Recent Research Notes: Grounded Attribution for AI Outputs
This note captures recent primary sources that directly shaped the current mechanism. The conclusion is clear: payout attribution and user-facing source attribution should share the same evidence trail.
2026-07-10 RDLLM 1.0 Evidence Boundary
Four recent papers sharpen the 1.0 design. Cited but Not Verified shows that
working and relevant links can coexist with substantially weaker factual support.
How Do LLMs Cite? shows why visible citation behavior is not proof of the
model's internal computational reliance. Data Attribution in Large Language
Models via Bidirectional Gradient Optimization treats training-data influence as
a distinct, model-access-dependent problem. What's a Credit Worth? shows that
the informativeness of an attribution signal can change whether royalty or
fixed-fee compensation is economically preferable.
RDLLM 1.0 therefore separates three evidence classes:
- answer-time source grounding bound before provider generation;
- observable support, text matching, and candidate allocation;
- separately scoped model or training-data influence evidence.
Unsupported, over-warranted, post-hoc, or disputed claims do not become direct settlement. The event records review or escrow, and direct execution remains disabled. Production deployment claims now require Ed25519-signed evidence from trusted external issuers; direct payout additionally requires a trusted payment processor attestation. Bundled profiles deliberately contain no such external evidence and cannot self-certify production readiness.
References: https://arxiv.org/abs/2605.06635, https://arxiv.org/abs/2606.28358, https://arxiv.org/abs/2606.04928, and https://arxiv.org/abs/2607.00641.
2026-07 GitHub, SEO, and GEO Documentation Update
The onboarding surface now treats documentation as part of the product boundary. Google's Search Central guidance frames SEO as making content easier for search engines to crawl, index, and understand while still serving users first. The GEO paper frames generative engines as systems that synthesize answers from multiple sources and need citation precision and recall. Stack Overflow's 2025 Developer Survey also shows why examples need to meet builders in common implementation languages rather than only in the Python library path. 2026 language-speaker rankings provide the basis for concise global quickstarts.
The repository therefore adds a GitHub start guide, live runtime use cases,
five implementation-language API examples, a public explanation ladder, and
localized quickstarts plus explainers in 15 languages. The new
github_docs_readiness_audit gate checks that these surfaces keep source
footers, Claim Evidence, support metrics, payout fields, and
source-disagreement status visible. This is not a ranking guarantee; it is a
maintenance rule that keeps the public docs crawlable, citeable, and useful to a
developer who has never seen RDLLM before.
References: https://developers.google.com/search/docs/fundamentals/seo-starter-guide, https://arxiv.org/abs/2311.09735, https://survey.stackoverflow.co/2025/technology, and https://en.wikipedia.org/wiki/List_of_languages_by_total_number_of_speakers.
2026-07 Production-Readiness Update
The production target is an open-source attribution platform that can be operated by individuals, companies, institutions, governments, model providers, and public sector deployments. The mechanism therefore adds a separate production-readiness gate for live operation. This gate is distinct from the artifact ship gate: it requires an operator profile, tenancy controls, runtime admission controls, source-footer enforcement, rights governance, dispute escrow, payment or instruction-only settlement controls, incident response, backup testing, public-sector controls, and public proof surfaces.
Those profile controls establish configuration readiness only. RDLLM 1.0 also requires current external attestations verified against an independently managed trust store before it permits a production-grade claim.
The controls are anchored to NIST AI RMF GenAI Profile, NIST SSDF, OWASP Top 10 for LLM Applications 2025, and SLSA because RDLLM production readiness is not only a model-quality claim. It is an operational claim about secure software, AI-risk controls, source-grounding integrity, supply-chain provenance, privacy, and auditable settlement.
2026-07 Runtime Display-Surface Update
Recent citation-failure work pushes the runtime boundary beyond "there is a
footer object somewhere in the JSON." Cited but Not Verified separates link
validity, relevance, and factual support; Detecting and Correcting Reference
Hallucinations in Commercial LLMs and Deep Research Agents shows that
source-looking URLs can be fabricated or stale even in search/deep-research
systems; CiteCheck separates real citations from metadata drift and fabricated
references; CiteGuard treats citation validation as retrieval-grounded
attribution alignment; and Correctness is not Faithfulness in RAG
Attributions shows why post-rationalized citations can look correct while not
reflecting model reliance.
The production rule is therefore that a response is not ready for display unless
the client-renderable answer surface is bound to the event hash and source-footer
hash. The service response now carries display.rendered_text, a
display_hash, and an answer_plus_verified_source_footer render policy.
rdllm-service-response-verify recomputes the display surface from the raw
answer and verified footer rows, then fails if the rendered answer/footer surface
is missing, modified, or detached from the summary hash. This keeps user trust
anchored to the exact answer the client can render, not to a hidden payout ledger
or decorative citation list.
2026-07 Inline Citation Marker Gate Update
Recent source-attribution results make one additional production rule necessary:
the answer text cannot contain citation-shaped markers that fail to resolve to
the verified footer. Cited but Not Verified shows that link validity and
topical relevance can mask much lower factual support, FACTUM treats citation
hallucination as distinct from generic RAG hallucination, SourceCheckup reports
large unsupported-statement rates even when answers cite sources, and
sub-sentence citation work argues that users need concise, sufficient evidence
anchors rather than broad citation spans.
The service therefore now treats unresolved inline markers such as [2],
[Source: 3], or [S9] as display-blocking audit errors unless they match a
verified source-footer source label or supported claim label. Numeric markers
resolve as source aliases, so [1] and [Source: 1] are accepted only when
verified source row S1 exists; numeric groups and ranges must fully resolve,
while four-digit bracketed years are ignored as non-citation text. The response
verifier publishes
answer_citation_marker_count, resolved_answer_citation_marker_count,
citation_marker_status, and detected/resolved/unresolved marker lists inside
source_grounding_acceptance, so operators can distinguish a grounded answer
with no inline markers from one that contains misleading citation-shaped text.
References: https://arxiv.org/abs/2605.06635, https://arxiv.org/abs/2601.05866, https://www.nature.com/articles/s41467-025-58551-6, and https://arxiv.org/abs/2509.20859.
2026-07 Answer-Link Source Binding Update
New deep-research citation papers add a display gap that bracket-marker parsing
does not cover: generated answers can include Markdown links or bare URLs that
look like source citations while bypassing the verified footer. Cited but Not
Verified evaluates inline citations in Markdown reports by parsing the
generated report, retrieving cited content, and separating link accessibility,
topical relevance, and factual support. ProvenAI similarly separates answer
correctness, citation fidelity, and document influence, showing that clean
surface citation audits can coexist with weak influence from cited sources.
How Do LLMs Cite? reinforces the product risk: inline citations can create a
false sense of security when citation behavior is not proof of source use.
The service therefore now treats answer-link URIs as a production display gate.
Any scheme://... URI in the generated answer, including Markdown links, must
match a verified footer source_uri; otherwise attribution returns a blocked
response and rdllm-service-response-verify fails ready responses. The verifier
publishes answer_link_uris, resolved_answer_link_uris,
unresolved_answer_link_uris, URI counts, and answer_link_status inside
source_grounding_acceptance. This does not claim the linked page is factually
sufficient by itself; it prevents generated answers from introducing foreign
source-looking locators outside the auditable footer.
References: https://arxiv.org/abs/2605.06635, https://arxiv.org/abs/2606.26449, and https://arxiv.org/abs/2606.28358.
2026-07 Answer-Claim Coverage Update
Recent claim-level verification work shows why valid source rows are not enough: the displayed answer can contain extra factual units that never receive a claim-evidence row. PaperTrail decomposes LLM answers and source documents into discrete claims and evidence to reveal unsupported assertions and omissions. MedRAGChecker similarly treats generated biomedical answers as atomic claims and aggregates claim decisions into answer-level diagnostics. The Google AI Overviews measurement study decomposed generated responses into 98,020 atomic claims and found unsupported claims even when cited pages were high quality, showing that source quality and claim fidelity are separate dimensions.
The service verifier now exposes answer-claim coverage inside
source_grounding_acceptance. It recomputes answer claim units from
event.answer_text, compares their hashes with footer claim_rows, and reports
answer_claim_unit_count, answer_claim_row_coverage_count,
uncovered_answer_claim_hashes, extra_claim_row_hashes, and
answer_claim_coverage_status. A ready response fails verification if answer
text claims and footer claim rows do not match exactly. This is a display
integrity gate: it prevents rehashed responses from adding unsupported answer
text outside the footer; the existing claim-evidence and support-score gates
still decide whether each covered claim is strongly supported.
References: https://arxiv.org/abs/2602.21045, https://arxiv.org/abs/2601.06519, and https://arxiv.org/abs/2605.14021.
2026-07 Evidence-Force Calibration Update
Recent cited-RAG evaluation work shows that claim coverage is still too weak
when a source is merely related to a claim. Relevant Is Not Warranted frames
this as citation laundering: the evidence may be real and relevant while failing
to warrant the exact force of the attached wording. Its FORCEBENCH stress test
uses five localized force axes: relation, modality, scope, temporal validity,
and numeric specificity. Know Before You Fetch also argues for calibrated
interfaces that make retrieval and abstention decisions explicit rather than
treating a raw score as a universal confidence signal.
The service footer therefore now renders claim-level warrant metadata. Each
Claim Evidence row includes claim_preview, claim_warrant_profile,
claim_force_flags, evidence_force_flags, warrant_mismatch_flags, and
warrant_strength_status. The response verifier recomputes those fields from
the claim and evidence text, fails grounded display when a supported claim's
evidence does not carry the same relation/modality/scope/temporal/numeric force,
and exposes claim_warrant_strength_status inside
source_grounding_acceptance. The public source-footer verifier enforces the
same gate for copied/exported footer artifacts. This moves RDLLM from "there is
a related evidence span" to "the evidence visibly warrants the strength of the
displayed claim."
References: https://arxiv.org/abs/2605.28044, https://arxiv.org/abs/2606.29959, and https://arxiv.org/abs/2605.06635.
2026-07 Visible Source-Disagreement Update
Recent conflict-aware RAG work adds a transparency requirement beyond source
support: systems should not assume retrieved or visible sources agree with each
other. ConflictRAG explicitly detects, classifies, and resolves knowledge
conflicts before answer generation. ArbGraph separates evidence arbitration
from generation by building support/contradiction relations over atomic claims.
Contradiction to Consensus shows why single-source verification is
insufficient: claim verification should retrieve both original and negated
perspectives and expose source-level disagreement.
The service footer now renders conservative source-disagreement metadata in
Claim Evidence rows. Each supported claim carries
source_disagreement_profile, agreement_source_labels,
disagreement_source_labels, and source_disagreement_status. The verifier
recomputes those fields from the visible source rows and fails grounded display
when a high-overlap visible source has opposite negation polarity for the claim.
The public source-footer verifier enforces the same gate for copied/exported
footers. This does not claim complete semantic conflict resolution; it prevents
the product from presenting a claim as cleanly grounded while a visible source
plainly says the opposite.
References: https://arxiv.org/abs/2605.17301, https://arxiv.org/abs/2604.18362, and https://arxiv.org/abs/2602.18693.
2026-07 Model-Reliance Claim Discipline Update
Recent attribution and explanation papers add a separate trust risk: answer text
can imply that the model internally used, reasoned over, or was influenced by a
source even when RDLLM has only observable source-support and allocation
evidence. How Do LLMs Cite? shows that citation behavior can diverge from the
model's internal computational pathway. ProvenAI separates citation fidelity
from per-document influence and gives examples where cited and influential
evidence diverge. Evaluating the False Trust Engendered by LLM Explanations
shows that reasoning traces and post-hoc explanations can persuade users without
being faithful provenance.
The service now blocks model-internal reliance claims on the public answer
surface. Phrases such as I used, my reasoning, or sources shaped the
answer fail display unless they are removed from the answer and represented
instead as observable footer support/allocation fields or as a separate,
verifiable model-signal artifact. source_grounding_acceptance publishes
model_reliance_claim_markers, model_reliance_claim_marker_count, and
model_reliance_claim_status, so operators can distinguish "this claim is
supported by a footer row" from the stronger and usually unverifiable statement
"the model internally relied on this source."
References: https://arxiv.org/abs/2606.28358, https://arxiv.org/abs/2606.26449, and https://arxiv.org/abs/2605.10930.
2026-07 Source-Usage Transparency Update
The latest papers point to a second display problem: even when the footer source
is real and claim-support verified, users still need to know what role the source
played in the answer. Explicit Evidence Grounding via Structured Inline
Citation Generation reports that document-level citation is easier than precise
evidence-span grounding; How Do LLMs Cite? warns that inline citations can
create a false sense of security because citation behavior may not reflect the
internal computation path; VERICITE shows that sentence-level citation support
can remain low even with retrieved medical abstracts; and What's a Credit
Worth? argues that noisy attribution signals materially change whether royalty
or fixed-fee compensation is welfare-optimal. A Human-Centric Framework for Data
Attribution in Large Language Models also frames attribution as a negotiated
stakeholder contract rather than a single universal score.
The runtime now renders source-usage metrics directly in service footers. Each
visible source row includes support, text_match, weight, and payout,
alongside source identity, verification handle, claim count, confidence,
selection rationale, settlement state, and content hash. These fields are
recomputed by both the saved response verifier and the public source-footer
verifier, so a copied or exported answer cannot silently replace the user-facing
usage explanation. source_grounding_acceptance now also reports
source_usage_metric_names, source_usage_metric_row_count, and
source_usage_metric_status, failing grounded display when a visible source row
does not expose the required usage/allocation metrics. The same verifier now
publishes claim_evidence_row_count and claim_evidence_status; every
supported claim must expose a visible Claim Evidence row whose claim hash,
support score, evidence span hash, character offsets, and evidence preview
recompute from the footer payload. That closes the gap between source-level
allocation and sentence- or claim-level factual support. RDLLM still
distinguishes observable attribution and settlement allocation from hidden
model-internal reliance; internal usage should only be claimed when provider
model-signal telemetry is separately verified.
References: https://arxiv.org/abs/2606.07130, https://arxiv.org/abs/2606.28358, https://aclanthology.org/2026.bionlp-1.62.pdf, https://arxiv.org/abs/2607.00641, and https://arxiv.org/abs/2602.10995.
2026-07 Source-Usage Metric Provenance Update
The latest attribution and verification papers add a sharper constraint:
showing a source-usage number is not enough unless the response also says which
metric profile produced that number and what the number is allowed to mean.
Do LLM Attribution Metrics Transfer? shows that attribution metrics do not
reliably transfer across datasets or constructs, so a product should not present
support or text_match as universal truth without exposing the metric method
and validation scope. LLM-as-a-Verifier treats verifier design choices such as
criteria decomposition, score granularity, and repeated scoring as part of the
calibration surface. ProvenAI separates answer correctness, citation fidelity,
and document influence, reinforcing that a clean source row is not proof of
hidden model-internal reliance.
The service footer now renders metric provenance beside every source-usage
number. Each visible source row carries usage_metric_profile,
usage_metric_scope, support_metric_method, text_match_metric_method,
weight_metric_method, and payout_metric_method, all hash-bound into the row
and visible in source_footer.rendered_text. The response verifier publishes
source_usage_metric_provenance_count and
source_usage_metric_provenance_status, failing grounded display unless every
visible source row exposes the expected profile, scope, and methods. The public
source-footer verifier enforces the same gate for copied/exported footer
artifacts. This keeps attribution auditable without implying that observable
support/allocation metrics reveal the model's private computational reliance.
References: https://arxiv.org/abs/2606.23915, https://arxiv.org/abs/2607.05391, and https://arxiv.org/abs/2606.26449.
2026-07 Attribution-Gap Closure Update
Recent LLM-search attribution work separates source use from source credit. The
attribution crisis in LLM search results defines the attribution gap as relevant
content consumed by a search-enabled LLM minus sources credited in the output,
and reports that deployed systems can cite fewer sources than they consume or
emit citations that do not appear in the disclosed search trace. Cited but Not
Verified adds the companion citation-quality problem: source links may resolve
and look relevant while factual support remains much lower. Citation Grounding
and CiteCheck show the same pattern in domain-specific settings by separating
existence, relevance, metadata fidelity, and temporal validity from generic
fluency.
The verifier now exposes attribution-gap closure as its own production display
dimension. source_grounding_acceptance publishes attribution_gap_status,
attribution_gap_verdict, accessed and credited source counts, and failure
counters for consumed-without-credit, cited-without-access, and paid-hidden
sources. Grounded display fails unless the attribution-gap verdict is closed
and those failure counters are zero. This keeps payout attribution, source
footer attribution, and source-access telemetry aligned before a ready response
can imply that its sourcing is complete.
References: https://www.cambridge.org/core/journals/data-and-policy/article/attribution-crisis-in-llm-search-results-estimating-ecosystem-exploitation/170DD0B88E5F5AEA8F69F2E9AF1328E3, https://arxiv.org/abs/2605.06635, https://arxiv.org/abs/2606.00898, and https://arxiv.org/abs/2605.27700.
2026-07 Source-Identity Binding Update
Recent citation-hallucination papers make source identity a separate runtime
failure mode. CiteCheck distinguishes exact citation matches from minor
metadata corruption and major fabricated references, showing that a plausible
reference must be grounded against external source metadata rather than trusted
from its surface form. Phantom References uses a stricter identity-level
definition of hallucinated citations: the referenced work must exist with
compatible authorship. Detecting and Correcting Reference Hallucinations
similarly treats URL validity as measurable and correctable, while Cited but Not
Verified separates link accessibility, topical relevance, and factual support.
The service response verifier now binds every visible source row back to the
event's source-reference metadata. source_grounding_acceptance publishes
source_identity_count and source_identity_status; grounded display fails if a
footer row's title, creator, work ID, chunk ID, URI, license, content hash,
support/allocation metrics, evidence preview, or evidence span hashes diverge
from the event source reference it claims to expose. This blocks a
self-consistently rehashed footer from swapping a displayed URI or creator while
leaving the private event trace unchanged.
References: https://arxiv.org/abs/2605.27700, https://arxiv.org/abs/2607.00738, https://arxiv.org/abs/2604.03173, and https://arxiv.org/abs/2605.06635.
2026-07 Temporal Grounding Disclosure Update
Recent temporal-RAG and public-information work adds another display failure
class: a source can be real, correctly identified, and relevant while still being
too stale for a current, latest, recent, or as-of claim. Freshness and the
Limits of Heuristic Trend Detection in Temporal RAG shows that semantic
similarity alone ignores time and can surface stale evidence for recency-sensitive
queries. Citation Grounding includes citation temporality as a separate
dimension from existence and relevance. Curated retrieval versus open web
search in public AI information services frames currency and source
trustworthiness as information-quality dimensions that remain invisible unless
measured. When Benchmarks Age shows the same temporal-misalignment problem in
factuality evaluation: static evidence can become wrong as the world changes.
The service verifier now exposes temporal_claim_markers,
temporal_claim_marker_count, source_temporal_metadata_count, and
temporal_grounding_status. Static answers pass without requiring temporal
metadata. If the rendered answer uses temporal wording such as current,
latest, recent, today, now, or as of, grounded display fails unless
every visible source row carries temporal freshness metadata. This does not
replace the full rdllm-source-freshness-audit/v1 artifact; it prevents the
runtime response from implying freshness when the footer has no timestamped
source basis to verify.
References: https://arxiv.org/abs/2509.19376, https://arxiv.org/abs/2606.00898, https://arxiv.org/abs/2607.05217, and https://arxiv.org/abs/2510.07238.
2026-07 Claim-Source Closure Update
Recent source-aware verification work adds a stricter failure class than
unsupported citation: a claim can be supported somewhere while being attributed
to the wrong visible source. ProvenanceGuard calls this cross-source
conflation and evaluates source-aware factuality with per-claim source IDs,
source-specific evidence routing, and attribution-swap probes. Citation-Closure
Retrieval and Per-Rule Attribution makes a related point for regulated QA:
systems need explicit claim-to-source closure rather than flattened, post-hoc
source lists. Cited but Not Verified reinforces the same runtime rule by
separating source accessibility, relevance, and factual support.
The service verifier therefore now treats claim-source closure as a production
display gate. source_grounding_acceptance publishes
claim_source_closure_count and claim_source_closure_status; every supported
claim must bind to the visible source row whose label, work ID, and chunk ID
match the claim row. A self-consistently rehashed footer that points a supported
claim at a different work or chunk now fails before display, even if the evidence
span itself remains syntactically valid.
References: https://arxiv.org/abs/2606.18037, https://arxiv.org/abs/2605.29742, and https://arxiv.org/abs/2605.06635.
2026-07 Source-Locator Integrity Update
Recent citation-reliability work separates source existence and accessibility
from claim support. Cited but Not Verified evaluates source attribution across
Link Works, Relevant Content, and Fact Check because factual support can remain
weak even when links resolve. Detecting and Correcting Reference Hallucinations
in Commercial LLMs and Deep Research Agents finds that retrieval-enabled and
deep-research systems still emit hallucinated or non-resolving URLs, and shows
that URL health is measurable and correctable. CiteCheck makes a similar point
for scholarly citations: verifying that the referenced work exists and is
correctly specified complements claim-level verification.
The service verifier now treats visible source locators as their own production
display dimension. source_grounding_acceptance publishes
source_locator_count and source_locator_status; every visible source row must
render its uri, deterministic verify handle, and content-hash prefix in the
footer. This keeps a client from showing fact-support and payout metrics while
hiding the locator information a user or auditor needs to check the source.
References: https://arxiv.org/abs/2605.06635, https://arxiv.org/abs/2604.03173, and https://arxiv.org/abs/2605.27700.
2026-07 Runtime Audit-Commit Update
The same recent work also changes the service audit boundary. A response is not
production-ready merely because it can render claim-level source rows; the
operator must later prove that the exact event, footer hash, display hash, source
count, and audit-error state were durably committed before release. PaperTrail
argues for user-facing claim-evidence provenance, CiteGuard and CiteCheck
show that citation attribution needs retrieval-grounded and metadata-aware
verification, and LLM-search attribution-gap measurements show that consumed
sources and displayed citations can diverge in deployed systems. The Nature
hallucination-incentives result also reinforces that systems need explicit
abstention/fail-closed incentives rather than rewarding plausible answers that
lack verifiable support.
The runtime therefore treats the hash-chained service audit log as part of the answer release gate. Readiness rejects corrupt existing audit chains, and attribution routes return a blocked response without answer, display, or source footer payload if the new audit entry cannot be committed. This keeps public source footers grounded in an operator-verifiable event history instead of becoming unprovable presentation text.
References: https://arxiv.org/abs/2602.21045, https://aclanthology.org/2026.acl-long.282/, https://arxiv.org/abs/2605.27700, https://www.cambridge.org/core/journals/data-and-policy/article/attribution-crisis-in-llm-search-results-estimating-ecosystem-exploitation/170DD0B88E5F5AEA8F69F2E9AF1328E3, and https://www.nature.com/articles/s41586-026-10549-w.
2026-07 Verification-Contract Update
The client verifier now publishes a standalone schema for
rdllm-service-response-verification/v1, not only for the response payload. This
is a product requirement, not a packaging detail: clients, launch gates, support
bundles, and auditors need to validate the verifier result that says whether an
answer/footer surface is displayable.
Recent work supports treating verifier outputs as first-class artifacts.
Cited but Not Verified evaluates citation quality by separating URL
accessibility, topical relevance, and factual support. CiteCheck separates
retrieval from verification and emits structured comparison results for exact,
minor, and major citation hallucinations. PaperTrail shows the value of
claim-evidence provenance, while also showing that provenance detail must be
usable enough to affect operator and user behavior. A Nature Communications
medical-citation study found high rates of unsupported cited statements, which
reinforces that a source-looking footer is not enough for display approval.
References: https://arxiv.org/abs/2605.06635, https://arxiv.org/abs/2605.27700, https://arxiv.org/abs/2602.21045, and https://www.nature.com/articles/s41467-025-58551-6.
The runtime now also fails closed on non-display-safe grounding. If the embedded
grounding verdict is warning, failed, blocked_by_policy, or
blocked_by_registry, the service response is blocked, includes a
grounding_quality: audit error, and renders unsupported claim hashes in the
footer. This follows the same recent evidence: citation systems can produce
working links and plausible metadata while the cited source does not support the
claim, so users need an explicit blocked status and unsupported-claim disclosure
instead of a source-looking footer alone.
2026-07 Source-Grounded Acceptance Update
The verifier now exposes a nested
rdllm-service-source-grounding-acceptance/v1 profile. This turns the research
distinction between link presence, source relevance, factual support, and
faithful citation into a machine-checkable product gate. A response can be
internally consistent for audit while still failing grounded display acceptance;
production display now requires visible verified source rows, supported claim
evidence, minimum support score, non-escrow royalty-share coverage, and a
rendered answer that includes the bound source footer.
This update follows three recent findings. Cited but Not Verified shows that
high link validity and relevance do not imply factual citation correctness.
How Do LLMs Cite? warns that inline citations can be produced by a distributed
attribution heuristic rather than faithful source use. Citation-Closure
Retrieval and Per-Rule Attribution shows that regulated workflows need
claim-to-source closure, not post-hoc source lists. Indirect-prompt-injection
benchmarks also reinforce that attribution must be evaluated end-to-end from
ingestion through answer release.
References: https://arxiv.org/abs/2605.06635, https://arxiv.org/abs/2606.28358, https://arxiv.org/abs/2605.29742, and https://arxiv.org/abs/2601.10923.
2026-07 Launch-Gate Update
The operator surface now has a fail-closed launch gate because the newest
citation and attribution papers keep separating visible citation shape from
actual source support. Cited but Not Verified shows that working, relevant
links can still fail factual support; Explicit Evidence Grounding via
Structured Inline Citation Generation shows that document-level citation is
easier than precise evidence-span grounding; How Do LLMs Cite? argues that
inline citations may follow shallow attribution heuristics rather than faithful
source use; Citation-Closure Retrieval and Per-Rule Attribution shows that
regulated workflows need closure over all required rules, not just a plausible
single source; and legal-citation hallucination benchmarks show that subtle
misquotes and mismatches remain hard even for agentic verifiers.
The practical product rule is that an operator cannot rely on a clean README, a valid service config, or a source-looking footer in isolation. Before production display traffic, the deployment must jointly verify the profile controls, runtime service readiness, bootstrap manifest when present, and at least one saved attribution response footer. The launch gate is therefore a product decision surface, while the support bundle is the redacted artifact to share when the gate blocks.
References: https://arxiv.org/abs/2605.06635, https://arxiv.org/abs/2606.07130, https://arxiv.org/abs/2606.28358, https://arxiv.org/abs/2605.29742, and https://arxiv.org/abs/2606.21155.
2026-07 Source-Rationale Update
Recent citation-verification papers changed the mechanism target from "show a
source and pay it" to "explain why this source is in the answer footer." The
implementation now treats every visible footer row as a compact attribution
claim: the row must name the source identity, claim-support path, source
availability, reliance/licensing state, settlement state, and a stable
source_rationale_hash. The rendered Markdown footer also carries a why=...
reason code, and the rendered-attribution audit rejects source rows that do not
explain source selection.
This update follows the recent evidence that citation-looking strings are not enough: source links can resolve without fact support, scholarly citations can have subtle metadata drift, and generative feature attribution needs an explicit contract for what is being explained. The practical rule is therefore fail closed: cite only rows with verified identity, support, availability, reliance, rights, and royalty/settlement coverage; otherwise downgrade the footer row to a review/escrow state.
Design Takeaways
- Cite every supported claim, not just the answer as a whole.
- Treat citation markers as constrained source pointers, not free-form text.
- Verify generated claims against cited evidence after generation.
- Keep retrieval, text matching, citation scoring, and training-data attribution as separate channels in the ledger.
- Show users a footer with source identity, evidence quote, source URI, content hash, evidence span hash, and grounding coverage.
- Hold value in escrow when no registered source can be traced.
- Enforce rights before attribution: a source can be retrievable but not licensed for generation, display, quote, external attribution, or training.
- Log span hashes and policy decisions because reproducible evidence identity can be weaker than document-level overlap.
- Treat conformance as a first-class product artifact. A provider should be able to prove that links, relevance, fact support, rights, receipts, and payouts agree.
- Separate source availability, source quality, citation integrity, and factual support. A working URL or registered source is not the same as evidence that supports a claim.
- Separate ownership proof from content matching. A hash can prove that two registrations point to the same text, but not which claimant is legally entitled to royalties; open conflicts should therefore escrow value until resolved.
- Make attribution portable. Receipts should not be trapped in one vendor's JSON API; they need credential-style proofs and provenance graphs that can move into wallets, registries, enterprise audit tools, and transparency services.
- Close the attribution gap explicitly. A source that was accessed, retrieved, or text-matched must be visible to the user, paid, or withheld with a reproducible rights/registry escrow reason.
- Separate public confidence from private audit. Users should see source and claim-support facts, while auditors can verify salted commitments for private prompts, evidence text, access traces, and economics.
- Bind provider telemetry to attribution receipts. A source footer is not enough if the provider can omit the retrieval/text-match trace; source-access spans, citation spans, and claim-support spans need shared hashes and portable verification.
- Treat probabilistic or model-internal attribution as evidence, not final truth. Token-level sensitivity, hidden-state probes, activation traces, and influence scores should be committed as auditable signals that can support, but not silently replace, runtime source and claim evidence.
- Make attribution configurable by stakeholder objective. Creator compensation, user confidence, publisher compliance, and platform auditability need different disclosure surfaces over the same signed event.
- Publish a provider-level attribution card. Per-answer receipts prove individual events; model buyers, regulators, creators, and platforms also need a signed, comparable deployment summary covering certification, source coverage, evidence channels, challenge handling, and known limitations.
- Publish the proof pack itself. Public confidence should not depend on a provider claiming that receipts, traces, statements, challenges, cards, summaries, and certification reports exist; the artifact set should have hash roots and inclusion proofs that independent verifiers can check.
- Make the user-facing answer verifiable. A source footer should have a companion answer provenance card that binds footer labels and claim span hashes back to the receipt and trace without exposing private prompt or evidence text.
- Materialize cited sources, not just citation strings. A grounded answer should carry a verifier report proving source hashes, quote hashes, claim evidence spans, source URIs, and answer-card bindings resolve to registered content without publishing private source text.
- Package the answer as a verifiable API object. Foundation-model adoption needs a response envelope that returns the rendered answer and embedded public proof artifacts together, so customers can verify the response boundary without private ledger access.
- Publish a provider integration profile. A foundation-model API needs a machine-readable contract for endpoints, response-envelope format, schemas, verifier commands, public surfaces, readiness checks, and bound proof hashes; otherwise attribution remains a custom integration instead of a deployable standard.
- Publish a well-known discovery manifest. Customers, creators, auditors, and downstream platforms need one stable entry point that tells them where the provider card, certification report, integration profile, response envelope, assurance bundle, schemas, verifier commands, and proof hashes live.
- Preserve derivative lineage. Foundation-model and agent pipelines increasingly use summaries, synthetic data, tool outputs, and transformed corpora; attribution must carry upstream work edges so derivative artifacts do not erase original owners from settlement.
- Benchmark provenance, not only payout. A provider should publish replayable attribution-quality evidence across clean, paraphrased, hard-decoy, escrow, and derivative-lineage cases so source finding is independently testable.
- Prove source influence, not just source presence. RAG attribution research now separates exposure, utility, and attribution, while counterfactual RAG papers test whether evidence changes decisions. Provider proof should therefore ablate credited sources and publish replayable influence margins.
- Treat consent as a lifecycle, not a one-time ingest flag. Rights reservations, opt-outs, and revocations need changed-policy hashes, preserved historical event hashes, and future-use denial probes so providers can prove they honor updated owner intent without rewriting past royalty records.
- Propagate consent changes across every attribution surface. A source can be correctly blocked at retrieval while still leaking through persistent memory, private reasoning, post-training signals, stale source footers, downstream exchange manifests, or settlement queues. The rights event therefore needs an SLA-bound propagation report over all those surfaces before future use or direct payout continues.
- Negotiate attribution before cross-provider use. Static proof publication is not enough when one AI system consumes another provider's answer or generated work; the caller needs a signed handshake that binds minimum level, required artifacts, source-footer relay, escrow relay, runtime headers, and downgrade rejection before the content enters a downstream workflow.
- Make attribution survive copying. Recent LLM-search work shows that relevant pages can be consumed without equivalent visible credit; therefore the answer itself needs a portable capsule binding the output hash, source-footer proof, federation handshake, C2PA-compatible assertion, and SCITT-like statement subject after it leaves the original product surface.
- Make footer confidence explicit. Recent citation-hallucination work shows that source-looking strings can be fabricated, stale, or weakly supportive, so the response should expose verified/warning/failed footer rows and a hallucination taxonomy grounded in source materialization and license proof.
- Bind the rendered footer itself. Recent verification work increasingly separates source identity, source support, confidence calibration, and human inspection, so the response should ship a signed client-rendering contract for the exact source rows and claim anchors users will see.
- Make private evidence challengeable. The public footer should not expose private prompts, evidence text, access traces, or payout accounts, but auditors still need a nonce-bound way to test whether redacted receipt paths open to the committed source, claim, rights, and royalty facts.
- Prevent attribution laundering after copy/reuse. Data-lineage and citation hallucination research both point to a second-hop problem: once an AI answer is copied into another workflow, downstream systems can cite only the wrapper or invent fresh citations. A portable report should bind the downstream event to the upstream capsule, verify the copied body hash, carry original source rows forward, and settle a pass-through pool to upstream owners.
- Clear obligations across providers. Data-valuation papers can estimate source influence, and citation-evaluation papers can test grounding, but a royalty market also needs a neutral clearing artifact that ingests provider statements and transitive reuse reports, deduplicates repeated submissions, holds overlapping evidence for review, and publishes payable/escrow totals.
- Bind attribution to payment instructions without overclaiming settlement. Once obligations are cleared, creators need remittance rows that can be reconciled by payment systems, but the attribution layer should not pretend a bank transfer has executed. Therefore the mechanism emits instruction-only rows with payout-account hashes, end-to-end IDs, remittance references, and hold preservation.
- Name attribution suppression explicitly. Recent attribution-bias work shows
that a model can have source or authorship information and still omit visible
attribution. RDLLM therefore treats footer omissions as an explicit
attribution_suppression_countin answer cards and source-confidence reports, so a provider cannot hide known sources behind a generic citation failure. - Require independent replay, not provider self-attestation. Recent citation and provenance work separates correctness, faithfulness, and support; RDLLM-L45 therefore turns the whole proof pack into a third-party audit attestation that binds public artifacts by hash, replays available verifiers, and records negative controls without publishing private prompts, source text, or payout accounts.
- Prove event revenue before payout. Attribution and clearing are not enough if
the
gross_revenuefield is trusted input, so RDLLM-L46 adds revenue allocation reports that bind hashed subscription, advertising, API, enterprise, and marketplace revenue pools to event-level gross revenue and creator-pool totals before statements, clearing, and remittance. - Prove the revenue pools themselves. A provider could conserve a fabricated pool, so RDLLM-L47 adds finance ledger attestations that reconcile hash-only invoice, billing, ad-server, API-meter, enterprise-contract, and marketplace exports to each revenue source without disclosing customer records.
- Publish the proof pack's replay order. A rich artifact set is not enough if auditors must infer verifier order by hand, so RDLLM-L48 adds proof dependency graphs that bind artifact hashes, separate hard replay dependencies from publication commitments, emit a topological replay order, and reject cycles.
- Monitor published proof surfaces after release. Attribution is only trustworthy if the provider card, certification report, response envelope, assurance bundle, and replay graph remain reproducible over time, so RDLLM-L49 adds an append-only publication monitor with Merkle snapshots, checkpoint history, artifact diffs, and certification-regression blocking.
- Bind proof artifacts to public trust roots. A witnessed checkpoint is still weak if verifier software does not know which provider, auditor, or witness keys are active or revoked, so RDLLM-L51 adds a trust registry for signer key hashes, key IDs, rotations, revocations, artifact signatures, and witness signatures.
- Make certification itself independently attestable. A certification report hash is useful, but adoption requires a signed object that a customer can verify without trusting the provider's web page. RDLLM-L52 adds certification attestations that bind the certification report hash, levels root, case-status root, certifier identity, target provider, and signature, then lets the trust registry verify the certifier key.
- Prove evidence sufficiency, not only source existence. Recent citation and evidence-ranking work shows that a valid source can still be redundant, ambiguous, or not the best support for a claim. RDLLM-L56 therefore ranks candidate spans per claim, requires the displayed span to be the top-ranked minimal sufficient evidence, and records a decoy margin before treating the footer as grounded.
- Adjudicate counterevidence before release. Citation verification and evidence ranking still allow one-sided answers when a claim has support but registered material directly disputes it. RDLLM-L57 therefore scans registered source spans for contradiction candidates and fails answers with unaddressed counterevidence before display or settlement.
- Close the release/application gap. A provider can publish correct late-stage grounding artifacts and still serve an answer envelope that does not depend on them. RDLLM-L58 therefore requires L55-L57 source availability, evidence sufficiency, and counterevidence reports to be embedded and bound in the response envelope before a high-certification release gate can emit.
- Close the answer-surface coverage gap. Sentence-level citation systems and claim-evidence interfaces still leave a serving risk: a provider can append an unsupported factual sentence to an otherwise grounded response. RDLLM-L59 therefore hashes every support-bearing public answer sentence and requires it to replay to a verified, sufficient, counterevidence-free claim row before release.
- Close the post-hoc citation gap. Retrieval-aware citation systems can still attach plausible sources after generation. RDLLM-L60 therefore binds every verified displayed claim to source-access spans and redacted generation-context block commitments captured before response delivery.
- Treat retrieved source text as evidence, not instructions. Recent prompt injection research shows that webpages, documents, and reusable prompts can carry machine-targeted instructions that alter model behavior. RDLLM-L61 therefore adds source boundary reports: each source packet must be hash-bound, role-labeled as evidence, excluded from control/instruction channels, and unable to modify attribution or payout policy.
- Prove which proof channels influenced the final answer. Recent attribution, citation, and prompt-injection work converges on the same operational gap: users need not only source footers, but a replayable reason why a source was cited, paid, or ignored. RDLLM-L62 therefore adds a decision provenance graph over claim grounding, visible footer rows, payout participation, and release decisions, with source text admitted as evidence only and never as payout or policy control.
- Calibrate attribution confidence instead of implying certainty. Recent RAG and faithfulness work shows that source correctness, actual reliance, and metric faithfulness can diverge. RDLLM-L63 therefore adds a calibrated attribution confidence report: every public claim, footer row, and payout-participation row carries a benchmark-backed lower confidence bound, and low-confidence rows must be disclosed or escrowed.
- Verify source authenticity before direct credit or payment. Recent poisoning work shows that retrieved knowledge bases can be adversarially injected, and source-attribution work shows that influence must be traced to the actual text responsible for a generation. RDLLM-L64 therefore binds public source rows to trusted origin evidence, archive consensus, active license terms, source-farm risk, poisoning risk, and synthetic-source disclosure; failed rows remain visible only with warnings and route payment to escrow until review.
- Make external verification economically accountable. Recent citation and provenance work motivates independent replay, but settlement systems also need consequences for incorrect or conflicted verifier attestations. RDLLM-L72 binds accepted verifier rows to trust-registry identity, non-revoked key hashes, slashable bond coverage, conflict disclosures, challenge rows, and slashing evidence before verifier-approved settlement can leave escrow.
- Make the economic usage log globally consistent, not merely signed. Certificate Transparency, SCITT, and witness systems show that inclusion proofs are not enough if different parties can be shown different roots. RDLLM-L73 therefore compares receipt-log snapshots across provider, creator, customer, and witness views, verifies append-only prefixes and required receipt inclusion, detects split-view roots, and escrows settlement when the usage log cannot be replayed as one consistent history.
- Make monitoring economically live. Transparency logs still fail in practice when nobody watches them, when monitors are incomplete, or when a provider can delay remediation after a public defect is discovered. RDLLM-L74 therefore requires registered independent watchtower attestations over the L73 subject, treats open or accepted public challenges as settlement blockers, and emits hash-only slashing and bounty rows for accepted verifier-observation failures.
- Prove that footer sources were actually relied on. Recent citation work shows that correct-looking references and source-support checks can still diverge from actual generation reliance. RDLLM-L100 therefore emits citation reliance receipts that bind visible footer rows to pre-generation evidence locks, rendered claim-evidence rows, claim-source replay, causal utility trials, current-turn trace membership, source-access leases, and content-protocol permission before a directly settled source can be treated as faithful.
- Make license authorization transactional, not just declared. Machine-readable rights protocols and provenance ledgers are necessary but insufficient if the provider cannot prove the publisher or license server actually authorized the exact source access. RDLLM-L101 therefore emits license transaction receipts: signed authorization tokens, license-ledger inclusion, source-access nonce and meter binding, protocol-term matching, transaction validity checks, and escrow routing for missing, expired, unsigned, non-ledgered, or mismatched tokens.
- Make the answer footer the verification surface. Recent work separates source identity, factual support, actual reliance, confidence calibration, and rights authorization, so RDLLM-L102 emits grounded source footer receipts that bind each visible footer row to confidence, availability, exact public evidence regions, citation reliance, license transactions, and public verifier handles without exposing prompt, answer, source, evidence, or payout data.
- Make delivery part of attribution, not a UI convention. Faithful citations and grounded footer receipts still fail users if the served answer drops footer rows, span handles, or verifier metadata. RDLLM-L103 therefore emits source footer delivery receipts that bind the grounded footer to the proof-carrying response, copied output, gateway egress hash, source labels, claim span handles, and public verifier metadata.
- Make attribution portable as a minimum model API contract. Recent citation, provenance, and content-credential work points to the same deployment problem: a citation row is not trustworthy unless generic clients know which headers, embedded proof objects, verifier paths, and fail-closed policies must be present. RDLLM-L104 therefore emits foundation API attribution profiles that bind the provider card, certification report, integration profile, discovery manifest, response envelope, and L103 source-footer-delivery receipt into a vendor-neutral response contract.
- Make the relying client prove enforcement. API-boundary attestation work and content-provenance deployments show that provider metadata is insufficient if downstream clients can strip, ignore, or rewrite it. RDLLM-L105 therefore emits client attribution enforcement receipts that bind observed response headers, embedded proof objects, verifier commands, source labels, and the client's render/block decision before a chat, search, agent, or relay surface displays an attributed answer.
- Make persistent memory a provenance-bearing derived work. Long-running agents increasingly write summaries, profile facts, and cross-session memories; if those memory cells are later used in an answer, the user and creator need the same source carry-forward that a direct retrieval answer would have. RDLLM-L106 therefore emits persistent memory provenance receipts that bind memory writes, reads, source labels, upstream proof hashes, license and retention policy, royalty obligations, delete tombstones, and visible footer labels.
- Make private reasoning attribution auditable without exposing chain-of-thought. Foundation models, routers, and agent runtimes increasingly use hidden scratchpads, delegated reviewers, and intermediate summaries. RDLLM-L107 emits private reasoning attribution receipts that commit to those private traces by hash, bind them to L106 memory and L105 client proofs, require every hidden source label to appear in the verified footer, and preserve royalty rows without publishing raw reasoning text.
- Prevent synthetic-source attribution laundering. Recent source-attribution and cited-RAG work shows that a visible source can be relevant, reachable, and cited while still being an AI-generated wrapper or weakly supported proxy for the original human work. RDLLM-L121 therefore adds source-origin lineage: direct payout requires trusted human-origin attestation; synthetic sources require upstream creator lineage and share conservation; unknown or unattributed synthetic sources are shown to users but settled to origin-review escrow.
- Make the footer human-inspectable without exposing full source text. Recent citation-auditing work emphasizes that users over-trust bare citations and need visible evidence that the source actually supports the answer. RDLLM-L122 therefore adds evidence-preview footers: each verified visible claim must publish a short permissioned snippet, source URL, warrant label, origin label, and proof hash, while full source text, prompts, answers, private reasoning, and payment data remain undisclosed.
- Make the source preview click-through verifiable. Recent citation-verification and source-auditing work shows that link labels and snippets are still weak unless the reader can resolve the exact passage that grounded the claim. RDLLM-L123 therefore adds evidence-locator manifests: each preview snippet must bind to an exact public resolver URL, resolver status, and snapshot or text-fragment proof without redisclosing full source or excerpt text.
- Make foundation-model attribution provider-neutral. Current model APIs expose different native response envelopes, tool-call shapes, citation annotations, streaming events, and metadata surfaces. RDLLM-L125 therefore adds a composite foundation adapter: OpenAI Responses, Anthropic Messages, Google Gemini, Meta/Llama-style, Mistral, Cohere, xAI, Bedrock, Azure OpenAI, and OpenAI-compatible responses must bind their native output hash, headers, JSON proof fields, citation/tool paths, and final streaming metadata to the same RDLLM response envelope, source-footer delivery, and citation URL-health receipts before a client can call the answer grounded.
- Make provider compatibility fixture-backed. RDLLM-L126 adds a foundation provider conformance matrix so every supported provider family publishes hash-only positive and negative fixtures for attribution-critical response modes, claim-support footers, parametric-memory fallback, and fail-closed citation or proof failures.
- Make provider compatibility executable at response time. RDLLM-L127 adds a foundation runtime adapter receipt so a concrete native provider response either normalizes into the RDLLM response envelope, footer, URL-health proof, attribution headers, JSON proof fields, citation paths, and stream-final hashes, or fails closed before display.
- Make multi-provider routing non-bypassable. RDLLM-L128 adds a foundation runtime router receipt so broker APIs, fallback stacks, latency routers, and cost routers must prove every candidate provider route is adapter-backed, conformance-backed, hash-committed, and fail-closed before any selected provider response is displayed.
- Make backend model substitution non-bypassable. RDLLM-L129 adds a foundation model deployment attestation so the selected route must bind to an active provider deployment key, signed model/version commitments, and request/response boundary hashes before display.
- Make composite answers non-bypassable. RDLLM-L130 adds a universal composition receipt so an answer assembled from multiple foundation providers must bind every segment to one released L129 deployment attestation, source-footer delivery hash, telemetry span commitment, and conserved provider weight.
- Make source-grounded composition settleable. RDLLM-L131 adds a universal composition settlement receipt so every L130 provider segment must bind its displayed source labels and claim IDs to source-entitlement weights, creator-obligation rows, preserved source-footer delivery, and conserved revenue allocation totals. Payable rows require licensed or verified sources; unresolved rows go to escrow, and disputed, blocked, or revoked rows are held.
- Make foundation-model adoption one contract, not a provider-by-provider assertion. RDLLM-L132 adds a universal foundation model contract so all supported provider families must share the same adapter, conformance, runtime, router, deployment, composition, settlement, discovery, and public surface proof chain before a response can be released as RDLLM verified.
- Make native provider calls pass through an attribution gate before invocation. RDLLM-L133 adds a universal invocation guard so a raw OpenAI-, Anthropic-, Gemini-, Bedrock-, Azure-, xAI-, Cohere-, Mistral-, Meta/Llama-, or OpenAI-compatible call cannot bypass the L132 contract, selected route, deployment attestation, request/response boundary, source-footer requirement, fail-closed headers, or GenAI telemetry.
- Make deployment-wide provider usage reconcile to the gate. RDLLM-L134 adds a universal invocation coverage report so every native provider meter event must match one L133 guard, gateway egress record, source-footer delivery hash, response-envelope hash, billed unit row, gross-revenue row, creator-pool row, and provider invoice row before complete attribution coverage can be certified.
- Make coverage non-repudiable. RDLLM-L135 adds a universal invocation witness report so every L134-covered native provider call must bind to a provider-signed usage receipt, independently observed egress event, and independent witness quorum before a deployment can claim omitted calls would be externally detectable.
- Make exported content carry attribution after it leaves the chat surface. RDLLM-L136 adds a universal content credential that binds output hashes, source-footer delivery, response envelopes, invocation witnesses, and C2PA/VC-shaped provenance assertions so copied text, media, files, and agent handoffs can still point back to their attribution proof chain.
- Make deployment adoption inspectable as one public passport. RDLLM-L137 adds a universal RDLLM passport that binds certification, provider card, integration profile, discovery manifest, assurance bundle, proof graph, foundation adapters, invocation controls, composition/settlement receipts, and content credentials into a single offline-verifiable provider object.
- Make adoption provider-neutral and procurement-ready. RDLLM-L138 adds a universal adoption standard package with required artifacts, implementer roles, SDK surfaces, public discovery surfaces, verifier commands, procurement gates, and standards mappings, so a foundation-model provider, gateway, enterprise client, auditor, registry, or clearinghouse can adopt the same attribution mechanism without private integration rules.
- Make adoption executable before compatibility claims. RDLLM-L139 adds a universal interop test kit with provider-family golden fixtures, SDK bindings, CI/offline runners, negative mutation cases, verifier commands, and discovery paths, so a provider, gateway, SDK, client, or auditor can replay compatibility instead of trusting an implementation declaration.
- Make runtime context access become attributable evidence before the answer is released. RDLLM-L140 adds a universal context provenance bridge for MCP tools, retrieval, vector stores, file search, browser search, enterprise connectors, creator license endpoints, memory, and media credentials. Every context event must be authorized, licensed, projected into a source claim, rendered in the footer when it supports the answer, bound to agent-step attribution, and projected into royalty settlement. Unauthorized context, stale leases, hallucinated citations, missing footers, unbound agent steps, rights-denied sources, token-audience mismatches, connector audit gaps, and private text leaks fail closed.
- Make displayed citations independently verifiable before users rely on them. RDLLM-L141 adds a universal citation verification contract that binds the L140 context bridge to source identity, resolver/URL health, exact evidence locators, metadata fidelity, claim support, evidence-force calibration, source confidence, warranted footer rows, rendered attribution audits, source authenticity, and royalty state. Nonexistent sources, metadata drift, inaccessible locators, unsupported cited claims, overclaimed evidence force, footer/context mismatches, hallucinated labels, missing royalty state, stale snapshots, and private-text leaks fail closed.
- Treat cached or reused answers as fresh attributable usage, not as free provider-side bypasses. RDLLM-L142 adds a universal grounded reuse contract for semantic answer caches, gateway response caches, native provider prompt/KV caches, reusable agent reports, and enterprise RAG caches. It requires query-equivalence and evidence-overlap checks, source freshness, consent/license continuity, L141 citation continuity, cache-collision resistance, provider-family coverage, and a new royalty-metered reuse event before a cache hit can be displayed as grounded.
- Cover every declared model capability, not just text responses. Recent citation-integrity work shows that a source-looking footer can be fabricated, unsupported, stale, or metadata-drifted, while current model APIs expose attribution-relevant behavior across tools, web search, files, vision, audio, realtime sessions, image/video generation, embeddings, reranking, fine-tuning, batch, safety, and local-runtime surfaces. RDLLM-L181 therefore makes every declared capability and modality pair prove either verified source-footer behavior or an explicit no-source abstention path, with settlement held and invocation blocked for uncovered capability rows.
2 June 2026 Verification Addendum
The L130 implementation pass rechecked the most relevant recent primary sources for the production bypass case: a model gateway, broker API, or fallback stack can strip attribution even if one upstream provider response was correctly grounded. The update maps those sources into a concrete invariant: attribution cannot be a single-provider response feature; it must be enforced by the router that chooses the model actually displayed to the user. The L129 pass extends the same evidence boundary from the router to the selected backend deployment: a route label is not enough unless the provider signs the model deployment identity and binds it to the concrete request/response boundary. The L131 pass extends the boundary from attribution proof to settlement proof: showing a source footer is not enough unless the same source rows, claim IDs, and provider segment weights reconcile to payable, escrowed, or held creator obligations under a conserved revenue allocation pool. The L132 pass turns L125-L131 into a single adoption contract: provider families with different native APIs can only claim RDLLM compatibility when they publish one reproducible proof chain that covers provider discovery, public surfaces, runtime route selection, deployment identity, source footer, and settlement readiness. The L133 pass moves that contract from publication to pre-call enforcement: the gateway must prove the exact native invocation carries the L132 hash, route hash, deployment hash, request-projection hash, response-binding hash, source-footer delivery hash, response-envelope hash, fail-closed header contract, and provider/model telemetry before it calls the model. The L134 pass moves enforcement from a single invocation to a deployment window: native provider meters, gateway egress logs, L133 guard receipts, source-footer delivery hashes, response-envelope hashes, invoice rows, gross revenue, and creator-pool rows must reconcile one-to-one, or the deployment cannot claim that all foundation-model usage is attributable and settlement-ready. The L135 pass adds non-repudiation to that window: provider-signed usage receipts, independent egress observations, and independent witness quorum evidence must bind to each covered invocation, or the deployment cannot claim omitted foundation-model calls are externally detectable. The L136 pass makes source attribution survive export and copying: the generated asset itself carries a credential-style proof pointer to the answer envelope, source-footer delivery, invocation witness, and upstream proof chain. The L137 pass makes provider adoption inspectable: every proof artifact needed to evaluate a deployment is bound into one passport with public surfaces and verifier commands. The L138 pass turns that passport into an implementer-facing standard: the package names the required artifacts, roles, SDK surfaces, procurement gates, discovery paths, and standards mappings needed for universal adoption across foundation model providers, gateways, enterprise clients, creator registries, auditors, and clearinghouses. The L139 pass turns the standard into an executable compatibility kit: provider families must ship golden request/response fixtures, SDKs must prove fail-closed footer/proof handling, CI and offline runners must replay the kit, and negative mutation cases must block before a deployment claims universal RDLLM adoption.
- OpenTelemetry's GenAI semantic conventions explicitly account for proxied or hosted provider paths: multiple providers can expose OpenAI-compatible APIs, and request model, response model, provider, and server attributes are needed to identify the actual system in use. RDLLM-L128 turns that observability warning into a release rule: each candidate route is hash-bound to provider identity, adapter coverage, conformance fixtures, and fail-closed fallback behavior before display: https://opentelemetry.io/docs/specs/semconv/gen-ai/gen-ai-spans/
- Sovereign Context Protocol argues that creator-owned data access should be logged, licensed, and attributable at runtime, not only summarized after training. RDLLM-L128 extends that runtime principle to multi-provider routing: a gateway cannot consume or relay model output unless the selected route carries the same attribution and license proof chain: https://arxiv.org/abs/2603.27094
- Cited but Not Verified and CiteCheck show why visible citations are not enough: a source row must resolve to accessible, relevant, fact-supporting, and metadata-faithful evidence. RDLLM-L128 therefore requires the selected route to bind to L123-L127 evidence-locator, URL-health, adapter, conformance, and runtime-normalization artifacts rather than accepting provider-written citation strings: https://arxiv.org/abs/2605.06635 and https://arxiv.org/abs/2605.27700
- AuthGraph shows the value of comparing actual execution provenance against an authorization graph at tool and parameter-source level. RDLLM-L128 applies the same idea to economic attribution: the actual router decision must match the authorized provider-route graph, and unproven route substitutions fail closed: https://arxiv.org/abs/2605.26497
- The Attribution Contract and probabilistic token-attribution work reinforce that attribution claims need an explicit contract and uncertainty-aware evidence. The router receipt therefore does not claim to prove hidden model causality by itself; it proves the deployment boundary: which provider output was selected, which attribution proof artifacts accompanied it, and whether fallback choices preserved or broke that proof chain: https://arxiv.org/abs/2605.23080 and https://arxiv.org/abs/2605.21726
- AEX and provider-neutral GenAI telemetry make the API-boundary problem explicit: model calls can pass through intermediaries, compatibility layers, and hosted gateways. RDLLM-L129 therefore binds the public selected-route row to a signed deployment statement, active deployment-key hash, opaque model commitments, and native response/output hashes rather than trusting model-name strings alone: https://arxiv.org/abs/2603.14283 and https://opentelemetry.io/docs/specs/semconv/gen-ai/gen-ai-spans/
- AEX also separates source outputs from transformed, buffered, aggregated, or repackaged outputs; OpenTelemetry GenAI spans provide provider/model/response correlation fields. RDLLM-L130 uses those observations for composition: it treats a merged answer as a new proof object whose provider segments must carry their own deployment attestations, trace-span commitments, source-footer hashes, and weight conservation before release.
- The Attribution Contract requires attribution claims to specify the output, eligible features, generation process, held-fixed values, and attributed score. RDLLM-L131 maps that contract shape into settlement: source-entitlement rows must name the L130 segment, source label, claim ID, entitlement weight, rights status, and claimant status before any creator obligation can be recomputed: https://arxiv.org/abs/2605.23080
- A Human-Centric Framework for Data Attribution in LLMs emphasizes that attribution systems must serve different stakeholder objectives. RDLLM-L131 therefore keeps public source footers, private audit hashes, revenue allocation, and creator payment/escrow/hold status as separate but linked artifacts: https://arxiv.org/abs/2602.10995
- SoK: Blockchain Agent-to-Agent Payments identifies a four-stage lifecycle for trustworthy agent payments: discovery, authorization, execution, and accounting. RDLLM-L131 applies the same separation to creator royalties: source attribution and rights authorization are not treated as final settlement unless the accounting artifact conserves creator-pool value and exposes payable, escrow, and held totals: https://arxiv.org/abs/2604.03733
- Official provider APIs expose materially different response, streaming, tool, citation, model, and deployment surfaces. RDLLM-L132 therefore treats provider compatibility as a contract over artifacts rather than as one JSON shape: OpenAI Responses, Anthropic Messages, Google Gemini, Amazon Bedrock Converse, Azure OpenAI Responses, Mistral, Cohere, xAI, Meta/Llama-style, and OpenAI-compatible rows all have to map into the same public RDLLM proof chain: https://platform.openai.com/docs/api-reference/responses
- AEX's API-boundary attestation model, SEAR's multi-provider gateway schema, OpenTelemetry GenAI provider/model attributes, AuthGraph's provenance-vs- authorization comparison, and DataDignity/OLMoTrace output-to-source tracing all point to the same engineering rule: attribution cannot be added only after a model answers. RDLLM-L133 therefore makes the gateway admission decision itself a signed artifact, with missing headers, route drift, request/response drift, source-footer omission, telemetry drift, or private-text leakage blocking the native provider call. RDLLM-L134 adds the complementary accounting control: provider-side meter and invoice rows must reconcile against those admission artifacts so compliant sample calls cannot hide unguarded production calls. RDLLM-L135 adds the non-repudiation control: coverage rows must be backed by provider receipts, independent egress observations, and independent witnesses: https://arxiv.org/abs/2603.14283 https://arxiv.org/abs/2603.26728 https://opentelemetry.io/docs/specs/semconv/gen-ai/gen-ai-spans/ https://arxiv.org/abs/2605.26497 https://arxiv.org/abs/2605.05687 https://arxiv.org/abs/2504.07096 https://docs.anthropic.com/en/api/messages https://ai.google.dev/gemini-api/docs https://docs.aws.amazon.com/bedrock/latest/userguide/conversation-inference.html https://learn.microsoft.com/en-us/azure/ai-foundry/openai/how-to/responses https://docs.mistral.ai/api/ https://docs.cohere.com/reference/chat https://docs.x.ai/docs/api-reference
- C2PA 2.2, W3C Verifiable Credentials/Data Integrity, SCITT/RFC 9943, and OpenTelemetry GenAI conventions provide the closest existing interoperability anchors for RDLLM-L136 through L139: signed provenance assertions, portable proof-carrying claims, transparency statements, and provider-neutral telemetry. RDLLM adds the missing AI-royalty-specific contract: visible source footers, exact claim/source bindings, rights gates, payout/escrow rows, provider passports, and adoption gates must all hash back to the same proof chain: https://spec.c2pa.org/specifications/specifications/2.2/specs/C2PA_Specification.html https://www.w3.org/TR/vc-data-integrity/ https://www.w3.org/TR/vc-data-model/ https://www.rfc-editor.org/rfc/rfc9943.html https://opentelemetry.io/docs/specs/semconv/gen-ai/
- The newest citation-integrity papers make L138's visible-footer requirement non-optional. Cited but Not Verified separates link validity, relevance, and factual support; CiteCheck separates existence from metadata faithfulness; and LLM Hallucinations in the Wild shows fabricated references entering scholarly corpora at scale. L138 therefore requires source-footer rendering and verifier commands as procurement gates, not as UI decoration: https://arxiv.org/abs/2605.06635 https://arxiv.org/abs/2605.27700 https://arxiv.org/abs/2605.07723
- Recent data-provenance and attribution-protocol work pushes the mechanism beyond one payout table. Sovereign Context Protocol argues for logged, licensed, runtime-attributable source access; Tracing the Data Trail surveys provenance, transparency, and traceability across LLM lifecycles; DataDignity frames output auditing as pinpoint provenance; and ProToken shows token-level attribution can be represented without exposing private training records. RDLLM-L137/L138/L139 therefore treats runtime footers, training summaries, invocation telemetry, private audit commitments, and settlement rows as linked but distinct proof surfaces: https://arxiv.org/abs/2603.27094 https://arxiv.org/abs/2601.14311 https://arxiv.org/abs/2605.05687 https://arxiv.org/abs/2601.19672
- The L181 pass adds capability coverage because the recent citation literature and provider API surfaces now point to the same deployment risk: attribution cannot be proved only for a canonical text chat endpoint. Cited but Not Verified, CiteCheck, LLM Hallucinations in the Wild, and URL-health work show that source rows need materialization, metadata, link-health, relevance, and support checks; official OpenAI, Gemini, Claude, and Bedrock documentation show that source-affecting work happens through tools, web search, files, streaming, realtime, citations, multimodal inputs, batch, and async surfaces. RDLLM-L181 therefore requires capability fixtures, modality-pair fixtures, operation surface fixtures, route-capability rows, and negative fail-closed fixtures before a provider can claim model invocation, response release, footer reliance, or creator settlement: https://arxiv.org/abs/2605.06635 https://arxiv.org/abs/2605.27700 https://arxiv.org/abs/2605.07723 https://arxiv.org/abs/2604.03173 https://platform.openai.com/docs/api-reference/responses https://platform.openai.com/docs/guides/tools https://platform.claude.com/docs/en/build-with-claude/citations https://ai.google.dev/gemini-api/docs/function-calling https://ai.google.dev/gemini-api/docs/live https://docs.aws.amazon.com/bedrock/latest/userguide/conversation-inference.html
- The L182 pass adds live provider-capability discovery because current source footers can otherwise imply confidence from stale or undocumented model capability claims. Recent May 2026 citation papers show that link validity and citation shape are not enough; source rows need verification against actual materialized evidence and citation metadata. Provider docs also show that model capability support lives in mutable catalogs and endpoint-specific docs: OpenAI Responses/tools, Claude citations/tool use/model docs, Gemini models, Search grounding, function calling, Live API, Batch API, and Bedrock Converse supported-model surfaces. RDLLM-L182 therefore binds every capability claim to fresh official or attested provider evidence, endpoint compatibility, lifecycle status, region/tenant scope, L181 route coverage, and negative fixtures before invocation, response release, footer reliance, or creator settlement: https://arxiv.org/abs/2605.06635 https://arxiv.org/abs/2605.07723 https://arxiv.org/abs/2605.08583 https://arxiv.org/abs/2605.27700 https://www.microsoft.com/en-us/research/publication/datadignity-training-data-attribution-for-large-language-models/ https://platform.openai.com/docs/api-reference/responses https://platform.openai.com/docs/guides/tools https://platform.openai.com/docs/guides/tools-web-search https://platform.claude.com/docs/en/build-with-claude/citations https://platform.claude.com/docs/en/agents-and-tools/tool-use/overview https://ai.google.dev/gemini-api/docs/models https://ai.google.dev/gemini-api/docs/google-search https://ai.google.dev/gemini-api/docs/function-calling https://ai.google.dev/gemini-api/docs/live https://ai.google.dev/gemini-api/docs/batch-api https://docs.aws.amazon.com/bedrock/latest/userguide/conversation-inference.html https://docs.aws.amazon.com/bedrock/latest/userguide/models-supported.html
- The L183 pass adds native source-annotation normalization because official
provider APIs expose source evidence in incompatible response shapes. OpenAI
Responses output text can carry URL, file, container-file, and file-path
annotations; Claude citations expose character, page, content-block, and
streaming citation-delta forms; Gemini Search grounding returns
groundingMetadatawith search queries, grounding chunks, and grounding supports; and Bedrock Converse document blocks can enable document-specific citations. RDLLM-L183 therefore treats provider-native annotations as source evidence that must be parsed, hash-bound, normalized into RDLLM footer fields, finalized for streams/batches, and rejected if dropped, unsupported, or mismatched against the rendered footer: https://platform.openai.com/docs/api-reference/responses/create https://platform.openai.com/docs/guides/tools-web-search https://platform.openai.com/docs/guides/tools-file-search https://platform.claude.com/docs/en/build-with-claude/citations https://platform.claude.com/docs/en/api/python/messages https://ai.google.dev/gemini-api/docs/google-search https://docs.aws.amazon.com/bedrock/latest/userguide/conversation-inference.html
May 2026 Literature and Standards Scan
The latest improvement pass focused on sources that directly affect attribution quality, training-data provenance, and privacy-preserving proof:
- Memori (arXiv, 23 March 2026) treats persistent memory as an explicit layer for context-aware LLM agents. RDLLM-L106 adds the missing attribution controls for that layer: memory writes and reads are not trusted unless they preserve source labels, upstream proof hashes, license/retention terms, royalty obligations, and delete tombstones, and unless downstream answers carry the memory's source labels into the verified footer: https://arxiv.org/abs/2603.19935
- MemIR (arXiv, 25 May 2026) identifies "provenance-role collapse" in long-term agents when historical interactions are stored as flat text. Its typed memory representation separates raw evidence, retrieval cues, and truth-bearing claims. RDLLM-L106 adopts the same failure boundary at the receipt layer: memory cells must declare source roles, upstream proof hashes, and visible footer carry-forward before later answers can rely on them: https://arxiv.org/abs/2605.25869
- MemORAI (arXiv, 2 May 2026) reports that graph memory systems can lack provenance tracking and proposes turn-level factual-origin tracking in a multi-relational graph. RDLLM-L106 turns that architectural idea into a portable audit receipt: source labels, memory reads, rendered-output hashes, and royalty obligations are replayable outside the memory runtime: https://arxiv.org/abs/2605.01386
- MemTrace (arXiv, 27 May 2026) frames memory reliability as executable evolution graphs and root-cause attribution for memory failures. RDLLM-L106 therefore treats memory writes, reads, tombstones, and carried source labels as audit events rather than opaque assistant state: https://arxiv.org/abs/2605.28732
- MemLineage, MemQ, and MemMark (May 2026) sharpen the remaining design constraints: memory entries need cryptographic lineage and derivation DAGs; credit can propagate through memory provenance DAGs; and attribution should survive leaked or migrated snapshots with commitments or watermark-style evidence. RDLLM-L106 does not make watermarking mandatory, but its proof slots can carry such evidence alongside source-label and royalty carry-forward receipts: https://arxiv.org/abs/2605.14421, https://arxiv.org/abs/2605.08374, and https://arxiv.org/abs/2605.25002
- The Model Context Protocol specification standardizes agent access to external context and tools. RDLLM-L106 extends the same systems view to persistent memory: durable context is treated as a provenance-bearing input, not invisible model state: https://modelcontextprotocol.io/specification/2025-06-18
- TRACER (arXiv, 11 May 2026), AEX, and agent tracing practice all point at a multi-stage provenance problem: generated outputs can depend on tool calls, intermediate reasoning, model handoffs, and runtime routing, not only the final answer text. RDLLM-L107 adds a privacy-preserving receipt for that boundary: source-influenced hidden steps are represented by commitments, artifact hashes, source labels, memory/tool/delegation references, and royalty rows, while raw chain-of-thought remains private: https://arxiv.org/abs/2605.09934, https://arxiv.org/abs/2603.14283, and https://openai.github.io/openai-agents-js/guides/tracing
- DataDignity (arXiv, 7 May 2026) frames output auditing as pinpoint provenance:
an auditor may need to identify the source document most likely supporting a
response. RDLLM maps that idea into
claim_support, source commitments, and a public source footer: https://arxiv.org/abs/2605.05687 - RISE (arXiv, 17 April 2026) makes data attribution and valuation more scalable by sketching output-layer influence signals rather than indexing full gradients. RDLLM treats that as an adoption-grade model-native evidence channel that can inform training-value priors while remaining separate from runtime source proof: https://arxiv.org/abs/2604.16197
- ZK-Value (arXiv, 5 May 2026) shows a practical zero-knowledge path for verifiable data valuation, including an LSH-Shapley primitive and compact verification. RDLLM-L95 uses the same privacy-verifiability direction for residual corpus royalty evidence, and RDLLM-L96 adds a benchmarked valuation-method audit before those residual rows can be trusted: public rows carry evidence hashes, benchmark commitments, and conservation checks rather than raw validation data or source text: https://arxiv.org/abs/2605.03581
- Cited but Not Verified (arXiv, 7 May 2026) and CiteAudit (arXiv v3, 1 May 2026) show that link validity, topical relevance, and plausible metadata are not enough; the cited content must factually support the claim. RDLLM-L97 responds by binding rendered claim spans and source-footer labels to exact page, line, character, bounding-box, or timecode regions and by requiring wrong-region controls to fail: https://arxiv.org/abs/2605.06635 and https://arxiv.org/abs/2602.23452
- Sovereign Context Protocol (arXiv, 28 March 2026) argues that attribution should become a default property of source access, with logged, licensed, and attributable creator-owned data consumption. RDLLM-L98 implements creator-side access leases, and RDLLM-L99 extends the idea to external web-scale protocol ingestion so RSL, CoMP, SCP, ODRL, Croissant, robots.txt, and C2PA/TDM signals can be preserved into RDLLM contracts, leases, and escrow routing: https://arxiv.org/abs/2603.27094
- RSL and CoMP style publisher protocols turn content access into a machine-readable licensing and compensation negotiation surface. RDLLM-L99 does not replace those protocols; it verifies that their terms were discovered, hash-preserved, and enforced inside the RDLLM proof chain: https://rslstandard.org/
- The 2026 ACL survey on attribution, citation, and quotation highlights the fragmented evaluation landscape for evidence-based text generation. RDLLM maps that taxonomy into separate artifacts for citation identity, claim support, footer rendering, and exact source-region binding: https://arxiv.org/abs/2508.15396
- Concept Influence (arXiv, 16 February 2026) attributes model behavior to semantic directions rather than only individual examples. RDLLM treats that as a possible valuation-evidence method for residual training value, but still requires rights gating, direct-attribution exclusion, creator caps, receipts, and escrow: https://arxiv.org/abs/2602.14869
- Probabilistic Attribution For Large Language Models (arXiv, 20 May 2026) defines token attribution from next-token probabilities and entropy, which is useful for identifying unstable or influential parts of a generation. RDLLM treats that as a model-internal evidence channel that can be attached to telemetry without substituting for source proof: https://arxiv.org/abs/2605.21726
- Attribution Bias in Large Language Models (arXiv, 6 April 2026) introduces attribution suppression: models can omit attribution even when authorship information is available. RDLLM maps that failure mode into visible footer checks, source-confidence taxonomy, and release-gate blocking rather than treating it as a vague citation-quality issue: https://arxiv.org/abs/2604.05224
- Key-discovery and trust-root standards show how this should surface in a production provider contract: RFC 7517 defines JSON Web Keys and key IDs, W3C DID Core defines verification methods for decentralized identifiers, and Sigstore distributes root-of-trust material through TUF. RDLLM-L51 maps those patterns into a provider-neutral trust registry for attribution proofs: https://www.rfc-editor.org/rfc/rfc7517 https://www.w3.org/TR/did-core/ https://docs.sigstore.dev/about/security/ https://docs.sigstore.dev/cosign/signing/overview/
- The Attribution Contract (arXiv, 21 May 2026) argues that feature-attribution claims need an explicit contract naming the output, eligible features, held-fixed assumptions, and attributed score. RDLLM maps that into signed artifacts: every footer, source-confidence row, model-signal row, and certification attestation declares exactly what is being explained and what hash roots were bound: https://arxiv.org/abs/2605.23080
- CiteCheck (arXiv, 26 May 2026) reinforces that citation verification must check both existence and metadata faithfulness. RDLLM uses that distinction in source verification and citation-footer contracts, then adds L52 so users can verify that the whole citation-proof mechanism was certified by a signer rather than only asserted by the response provider: https://arxiv.org/abs/2605.27700
- SURE-RAG (arXiv, 5 May 2026) treats evidence sufficiency as a set-level property: individual passage scores can miss missing hops, unresolved conflicts, and retrieval uncertainty. RDLLM-L56 and L57 now mirror that separation by requiring claim-level sufficient support and explicit counterevidence adjudication before the L58 release gate can emit: https://arxiv.org/abs/2605.03534
- Towards Dependable Retrieval-Augmented Generation Using Factual Confidence Prediction (arXiv, 4 May 2026) uses conformal prediction and diagnostic checks to attach confidence to whether retrieved chunks came from a correct source and whether generated answers are consistent with retrieved context. RDLLM-L63 maps that design pressure into public lower-bound confidence fields rather than a single uncalibrated source score: https://arxiv.org/abs/2605.05244
- Faithfulness Metrics Don't Measure Faithfulness (arXiv, 24 May 2026) warns that many faithfulness metrics perform near chance under ground-truth evaluation. RDLLM-L63 therefore does not let a provider publish a bare faithfulness number: it binds attribution claims to benchmark case counts, observed rates, Wilson lower bounds, and disclosure duties: https://arxiv.org/abs/2605.25052
- RAGVUE (EACL 2026) decomposes RAG evaluation into retrieval quality, answer relevance/completeness, strict claim-level faithfulness, and judge calibration. RDLLM-L63 uses the same decomposition principle for deployable source footers: claim confidence, footer confidence, and payout-participation confidence are separate public rows: https://aclanthology.org/2026.eacl-demo.35/
- GhostCite (arXiv, 6 February 2026) studies citation validity at scale and reports model hallucination rates and human-review gaps around fabricated references. RDLLM maps this into verifier-readable citation fields, footer contracts, source availability snapshots, and release blocking instead of relying on human-visible bibliography text: https://arxiv.org/abs/2602.06718
- Source or It Didn't Happen / CiteTracer (arXiv, 9 May 2026) reframes citation hallucination detection as field-level adjudication rather than a binary found/not-found check. RDLLM maps that into typed verifier rows, citation-footer contracts, source-availability reports, and counterevidence adjudication so source rows can fail for specific, auditable reasons: https://arxiv.org/abs/2605.08583
- A Human-Centric Framework for Data Attribution in LLMs (arXiv v2, 8 May 2026) argues that attribution design depends on stakeholder objectives and domain criteria. RDLLM maps those negotiated objectives into policy, public footer, selective disclosure, and settlement artifacts: https://arxiv.org/abs/2602.10995
Late-2025 to May-2026 Attribution Scan
This pass focused on papers and standards that affect the exact question the mechanism must prove: not only who should be paid, but which source material the model relied on, whether the visible citation is faithful, and whether retrieved source text was safely treated as data.
- Correctness is not Faithfulness in Retrieval Augmented Generation Attributions (ICTIR 2025) distinguishes citation correctness from citation faithfulness and reports that plausible citations can be post-rationalized rather than genuinely used. RDLLM-L60, L61, and L62 respond by proving pre-generation context presence, source-boundary integrity, and decision provenance before or at release: https://doi.org/10.1145/3731120.3744592
- Source Attribution in Retrieval-Augmented Generation (arXiv, July 2025) adapts Shapley-style attribution to identify influential retrieved documents in RAG, but notes that repeated LLM utility evaluations are expensive. RDLLM uses this as support for a two-channel design: runtime source proof for user trust and sampled counterfactual/Shapley evidence for payout weighting: https://arxiv.org/abs/2507.04480
- CiteGuard (arXiv v4, April 2026; ACL 2026) reframes citation evaluation as citation-attribution alignment and uses retrieval-aware validation rather than trusting LLM-as-judge alone. RDLLM maps this into source verification, source-confidence reports, and citation-footer contracts: https://arxiv.org/abs/2510.17853
- RAGOrigin / Who Taught the Lie? (arXiv, September 2025; IEEE S&P 2026) treats poisoned RAG output as a responsibility-attribution problem and assigns scores to candidate texts using retrieval ranking, semantic relevance, and influence on the generated response. RDLLM-L64 maps this into source-authenticity rows that must bind a footer source to origin evidence and low poisoning risk before direct payout: https://arxiv.org/abs/2509.13772
- Poisoned-MRAG (arXiv, March 2025) shows that a small number of malicious image-text pairs can steer multimodal RAG outputs, and M3Att (arXiv, May 2026; ACL 2026) shows query-agnostic poisoning in medical multimodal RAG using covert misinformation and visual triggers. RDLLM-L64 therefore treats source authenticity as a first-class gate across text and media, not as a cosmetic URL check: https://arxiv.org/abs/2503.06254 https://arxiv.org/abs/2605.10253
- Attribution Techniques for Mitigating Hallucinated Information in RAG Systems (arXiv, January 2026) surveys attribution methods across RAG hallucination classes. RDLLM reflects the same separation by splitting source availability, evidence sufficiency, counterevidence, claim coverage, and context closure into independent artifacts: https://arxiv.org/abs/2601.19927
- MaxShapley (arXiv v2, May 2026) targets incentive-compatible generative search compensation by computing fair context attribution more efficiently than exact Shapley. RDLLM uses the same economic direction but requires public proof artifacts before settlement so compensation does not rely on an opaque score: https://arxiv.org/abs/2512.05958
- LoRIF (arXiv v2, May 2026) makes gradient-based training data attribution more practical at large scale by reducing storage and query-time overhead. RDLLM treats these model-internal signals as useful training-value evidence, but keeps them separate from response-time source proof and visible footer grounding: https://arxiv.org/abs/2601.21929
- Mechanistic Data Attribution (arXiv, January 2026) traces interpretable LLM units back to training samples with influence functions and validates causal effects through interventions. RDLLM maps this to provider-private model-signal reports and training-content summaries rather than claiming deterministic event-level royalties from base-model memory alone: https://arxiv.org/abs/2601.21996
- Indirect Prompt Injection in the Wild (arXiv, April 2026) analyzes 1.2B URLs and finds validated machine-targeted prompt injections across webpages and HTTP responses. RDLLM-L61 maps that risk into source boundary telemetry and verifier checks that prove retrieved content was evidence-only: https://arxiv.org/abs/2604.27202
- ARGUS (arXiv, May 2026) constructs an influence provenance graph for LLM-agent decisions and reports strong reductions in context-aware prompt-injection success. RDLLM-L62 applies the same provenance-auditing direction to source attribution: claim, footer, payout, and release decisions must expose authorized influence edges rather than merely displaying citations: https://arxiv.org/abs/2605.03378
- AuthGraph (arXiv, May 2026) compares execution provenance against a clean authorization graph to detect tool and parameter-source deviations. RDLLM-L62 uses the analogous separation for royalties: retrieved source text may justify evidence, but policy and payout participation must come from explicit license, accounting, boundary, and release-gate proof channels: https://arxiv.org/abs/2605.26497
- AIP: Subverting Retrieval-Augmented Generation via Adversarial Instructional Prompt (EMNLP 2025) shows that trusted instructional prompts can be weaponized to alter RAG behavior. RDLLM-L61 therefore binds not only source packets but also the boundary profile that separates source data from control, instruction, attribution, and payout channels: https://arxiv.org/abs/2509.15159
- OWASP LLM01:2025 identifies direct and indirect prompt injection as a primary LLM application risk and recommends segregating external content. RDLLM-L61 converts that recommendation into a signed, replayable source-boundary artifact: https://genai.owasp.org/llmrisk/llm01-prompt-injection/
- TREC 2025 RAG evaluates attribution-rich answers using response completeness, attribution verification, and agreement analysis. RDLLM treats this as evidence that attribution quality must be benchmarked and published, not just asserted: https://pages.nist.gov/trec-browser/trec34/rag/proceedings/
- C2PA Technical Specification 2.2 and W3C PROV-O remain relevant interoperability anchors for provenance metadata and graph-shaped lineage. RDLLM uses analogous signed claims, content hashes, and provenance edges for AI response proof packs, while the 2026 C2PA security analysis reinforces why high-stakes use needs independent verification beyond metadata presence alone: https://spec.c2pa.org/specifications/specifications/2.2/specs/C2PA_Specification.html https://www.w3.org/TR/prov-o/ https://arxiv.org/abs/2604.24890
- Probing for Knowledge Attribution in LLMs (arXiv v2, 27 May 2026) separates context-use errors from internal-knowledge errors and shows that hidden-state probes can identify dominant knowledge source. RDLLM does not require hidden states in the reference implementation, but the trace exchange gives providers a stable place to attach model-internal attribution signals: https://arxiv.org/abs/2602.22787
- DATE-LM (arXiv, July 2025) proposes data attribution evaluation for language models. RDLLM uses that lesson operationally: attribution evidence has to be scoreable, but it also has to flow into receipts, statements, transitive reports, and now clearinghouse settlement rows: https://arxiv.org/abs/2507.09424
- Large-Scale Training Data Attribution with Efficient Influence Functions (ICLR 2025 submission, modified February 2025) frames training data attribution as useful for data citation, curation, and debugging. RDLLM keeps that as a training-value signal while requiring runtime source proof and separate settlement proof before creators are paid: https://openreview.net/forum?id=jZw0CWXuDc
- C2PA's technical specification and conformance program make content credentials portable across media supply chains. RDLLM maps that portability idea to AI text by binding output hashes, response envelopes, copy markers, transitive reports, and clearinghouse settlement artifacts: https://c2pa.org/specifications/specifications/2.2/index.html
- Verifiable Provenance and Watermarking for Generative AI (arXiv, 20 May 2026) combines cryptographic provenance, watermarking, and zero-knowledge attestation under a laundering and forgery threat model. RDLLM applies the same evidentiary stance to generated answers: provenance signatures are necessary, but high-level answer release also needs source availability, sufficient evidence, counterevidence adjudication, and trust-root verification: https://arxiv.org/abs/2605.21002
- RLVR Datasets and Where to Find Them (arXiv, 26 May 2026) documents provenance collapse in reasoning-training datasets and attributes more than 99% of a large RLVR corpus back to atomic sources. RDLLM's registry, trace exchange, and training-value channels are designed so derived datasets do not erase upstream source identity: https://arxiv.org/abs/2605.26971
- Tracing the Roots (arXiv, April 2026) introduces data lineage for post-training LLM datasets and reconstructs dataset-evolution graphs. RDLLM-L24 turns that lineage idea into settlement logic: derivative works carry upstream edges and pass-through royalty obligations: https://arxiv.org/abs/2604.10480
- Provable Model Provenance Set for LLMs (arXiv, February 2026) argues that model provenance claims need coverage guarantees across multiple possible sources, not heuristic single-source labels. RDLLM-L91 maps that concern into declared training-lineage reports, while RDLLM-L92 maps it into black-box provenance-set challenge reports for undisclosed derivative models: https://arxiv.org/abs/2602.00772
- Model Provenance Testing for LLMs (arXiv, February 2025) shows that derived models preserve statistically detectable output similarities under black-box API access. RDLLM-L92 operationalizes that into baseline-calibrated candidate model provenance sets and settlement challenges: https://arxiv.org/abs/2502.00706
- TextSeal (arXiv, May 2026) introduces localized watermarking whose signal can transfer through distillation, while VOW (arXiv, April 2026) adds verifiable oblivious watermark detection. RDLLM-L92 treats watermark evidence as one auditable signal, never as the only basis for payout: https://arxiv.org/abs/2605.12456 and https://arxiv.org/abs/2604.27666
- Behavioral Fingerprint via Refusal Vectors (arXiv, February 2026) and text-preserving fine-tune provenance auditing (arXiv, October 2025) show that derivative models can carry behavioral or document-level traces even when weights and training data are hidden. RDLLM-L92 combines those signals with multiple-testing controls and source-row settlement: https://arxiv.org/abs/2602.09434 and https://arxiv.org/abs/2510.09655
- Where does output diversity collapse in post-training? (arXiv, April 2026) shows that data composition can embed downstream behavior changes in model weights. RDLLM-L91 treats fine-tuning and distillation lineage as an ongoing attribution and settlement surface rather than a one-time dataset note: https://arxiv.org/abs/2604.16027
- Tracing the Data Trail (arXiv, January 2026) surveys LLM data provenance, transparency, and traceability. RDLLM maps those axes into source-access traces, training summaries, lineage reports, and public proof surfaces: https://arxiv.org/abs/2601.14311
- The TREC 2025 RAG Track overview (arXiv, 10 March 2026) evaluates systems on relevance, response completeness, attribution verification, and agreement analysis. RDLLM maps those evaluation dimensions into source footers, answer provenance cards, source verification reports, and response envelopes: https://arxiv.org/abs/2603.09891
- Who Benefits from RAG? (arXiv, 25 March 2026) separates group exposure, utility, and attribution, showing that fair source treatment requires more than retrieval frequency. RDLLM maps that distinction into source-access traces, provenance benchmarks, and counterfactual influence reports: https://arxiv.org/abs/2603.24218
- Counterfactual Reasoning for Retrieval-Augmented Generation (ICLR 2026) argues that RAG systems must distinguish causally decisive evidence from correlated but misleading evidence. RDLLM-L26 operationalizes that idea by removing each credited work and replaying the attribution path: https://openreview.net/forum?id=9U51rOnGko
- Beyond Semantic Relevance (arXiv, 2 May 2026) proposes counterfactual risk minimization for RAG and argues that semantic relevance can fail as a proxy for utility. RDLLM uses source-ablation reports as an audit artifact for the same distinction: https://arxiv.org/abs/2605.01302
- Source Attribution in Retrieval-Augmented Generation (arXiv, July 2025) adapts Shapley-style document attribution for RAG and studies lower-cost approximations. RDLLM uses a deployment-friendly replay layer: benchmark reports test source ranking, while counterfactual reports test source influence for individual paid events: https://arxiv.org/abs/2507.04480
- VISA (ACL 2025) introduces visual source attribution for RAG and localizes supporting evidence in the original document view with bounding boxes. RDLLM maps that requirement into media attribution reports and footer rows that can carry text spans, perceptual-hash commitments, source regions, and rendered evidence pointers instead of treating attribution as prose-only: https://aclanthology.org/2025.acl-long.1456/
- Ask in Any Modality (arXiv, February 2025), MAVIS (AAAI 2026), and MiRAGE (arXiv, October 2025) show that source attribution cannot remain text-only when models retrieve and generate from images, audio, video, and mixed documents. RDLLM-L27 adds a media attribution report with content hashes, perceptual-hash commitments, descriptor-hash commitments, ranked match scores, payouts, and unattributed-media escrow: https://arxiv.org/abs/2502.08826 https://ojs.aaai.org/index.php/AAAI/article/view/40585 https://arxiv.org/abs/2510.24870
- Verifying Provenance of Digital Media (arXiv, April 2026) warns that C2PA-style content provenance should not be treated as sufficient for high-stakes media trust by itself. RDLLM uses provenance metadata as an input but requires private replay, hash roots, payout conservation, escrow, and signed verifier artifacts: https://arxiv.org/abs/2604.24890
- Toward Accountable AI-Generated Content on Social Platforms (arXiv, April 2026) studies signed identifiers and multimodal attribution verification for generated imagery. RDLLM uses the same accountability direction but makes the payout and escrow ledger explicit: https://arxiv.org/abs/2604.10460
- The Attribution Contract (arXiv, 21 May 2026) argues that generative-model attribution must name the output being explained, eligible features, assumed generation process, held-fixed values, and attributed score. RDLLM-L28 now embeds that contract directly in model-signal reports: https://arxiv.org/abs/2605.23080
- Probabilistic Attribution for Large Language Models (arXiv, 20 May 2026) uses
next-token log-probabilities to infer token influence in a model-agnostic way.
RDLLM maps that idea into a
logprob_deltasignal channel rather than treating citation labels as the only evidence of source use: https://arxiv.org/abs/2605.21726 - Explaining the Reasoning of Large Language Models Using Attribution Graphs (arXiv, December 2025) argues that generated tokens influence later generated tokens through a causal graph, improving context attribution faithfulness. RDLLM uses this as a reason to keep model-signal attribution separate from public footer citations: https://arxiv.org/abs/2512.15663
- Cited but Not Verified (arXiv, May 2026) shows that link validity and topical relevance can coexist with much lower factual support in deep-research reports. RDLLM therefore keeps source materialization and answer provenance verification as public artifacts: https://arxiv.org/abs/2605.06635
- DAVinCI (arXiv, 23 April 2026) introduces a dual attribution and verification framework that attributes generated claims to internal model components and external sources, then verifies them with entailment reasoning and calibrated confidence. RDLLM-L40 maps that idea into a product surface: the exact footer rows and claim anchors shown to users must be verified before display, and RDLLM-L41 adds a private audit challenge for the redacted evidence behind those rows: https://arxiv.org/abs/2604.21193
- CiteVQA (arXiv, May 2026) introduces Strict Attributed Accuracy, crediting an answer only when both the answer and evidence region are correct, and reports attribution hallucination in current document models. RDLLM's source verification, media attribution, and model-signal layers target the same answer-plus-evidence requirement: https://arxiv.org/abs/2605.12882
- TRACER (arXiv, 11 May 2026) frames multimodal tool-using agents as having a provenance gap: tool trajectories and final answers do not say which observation supports which generated claim. RDLLM-L33 maps that concern to cross-provider runtime use: a downstream agent must verify the upstream provider's source-footer relay, escrow relay, artifact hashes, and downgrade policy before consuming the generated result: https://arxiv.org/abs/2605.09934
- The Attribution Crisis in LLM Search Results (Data & Policy, published 28 April 2026) measures a gap between relevant URLs consumed and URLs cited in real-world search-enabled LLM conversations. RDLLM-L34 directly targets the portability side of that gap by making the output hash and proof-chain pointer travel with copied answers: https://www.cambridge.org/core/journals/data-and-policy/article/attribution-crisis-in-llm-search-results-estimating-ecosystem-exploitation/170DD0B88E5F5AEA8F69F2E9AF1328E3
- VeriCite (arXiv, October 2025) verifies generated claims, selects supporting evidence, and refines answers around evidence. RDLLM hardens the same pattern by hashing claim evidence spans and binding them into receipts and answer cards: https://arxiv.org/abs/2510.11394
- SAFE (arXiv v2, September 2025) emphasizes sentence-level in-generation attribution so users can verify sentences while reading. RDLLM's answer card and source verification report provide the public proof surface for that level of granularity: https://arxiv.org/abs/2505.12621
- FRAMES (NAACL 2025) shows that end-to-end RAG must evaluate factuality, retrieval, and reasoning together rather than as isolated modules. RDLLM keeps retrieval traces, claim support, grounding quality, and payout evidence as cross-checking artifacts: https://aclanthology.org/2025.naacl-long.243/
- Mechanistic Data Attribution (arXiv, 29 January 2026) uses influence functions to trace interpretable model units back to training samples. RDLLM treats such mechanistic evidence as a future provider signal that can be committed in receipt telemetry without replacing runtime source evidence: https://arxiv.org/abs/2601.21996
- The European Commission's GPAI guidance requires copyright policies and public training-content summaries for general-purpose AI models. It says providers must identify and respect rights reservations, and that summaries help rightsholders assess whether lawful text and data mining conditions were respected. RDLLM-L29 turns that lifecycle requirement into a replayable changed-rights report: https://digital-strategy.ec.europa.eu/en/faqs/guidelines-obligations-general-purpose-ai-providers
- IPTC's Generative AI Opt-Out Best Practices recommend visible reservations, metadata declarations, robots/TDMRep/trust.txt signals, and asset-level metadata while warning that declaration protocols do not guarantee crawler compliance. RDLLM-L29 adds the missing provider-side proof that those rights changes were actually enforced after ingestion: https://iptc.org/wp-content/uploads/2025/05/IPTC-Generative-AI-Opt-Out-Best-Practices.pdf
- C2PA 1.3 defines a Training and Data Mining assertion for AI training, generative training, inference, and data-mining use states. RDLLM treats these asset-level signals as inputs to a broader text-and-media rights manifest, then requires remediation reports when those policy states change: https://c2pa.org/specifications/specifications/1.3/specs/_attachments/C2PA_Specification.pdf
- The Attribution Crisis in LLM Search Results (Data & Policy, 2026) argues that commercial LLMs need transparent search telemetry covering query, retrieval, reranking, and citation, and notes that OpenTelemetry GenAI conventions can carry source IDs and scores. RDLLM-L13 implements that bridge as a verifiable trace exchange: https://doi.org/10.1017/dap.2026.10064
- OpenTelemetry GenAI semantic conventions define provider-neutral attributes for GenAI spans, including data-source IDs and structured retrieval documents. RDLLM aligns source-access spans with those conventions while adding royalty-specific commitments: https://opentelemetry.io/docs/specs/semconv/gen-ai/gen-ai-spans/
- C2PA's January 2026 clarification states that Content Credentials record provenance and authenticity, not text-and-data-mining permissions. RDLLM keeps provenance, rights policy, and settlement as separate but linked artifacts: https://c2pa.ai/news/tdm-clarification
- W3C ODRL, C2PA provenance, SCITT transparency receipts, and EU GPAI copyright obligations point to a common deployment boundary: rights state cannot be a static registry field. RDLLM-L118 therefore adds an SLA-bound propagation audit proving that opt-outs, revocations, lease expiry, and license changes reached retrieval indexes, access leases, license transactions, source footers, persistent memory, private reasoning, post-training signals, attribution exchanges, downstream notices, and settlement before future direct use: https://www.w3.org/TR/odrl-model/ https://c2pa.org/specifications/specifications/2.2/index.html https://www.rfc-editor.org/rfc/rfc9943.html https://digital-strategy.ec.europa.eu/en/policies/contents-code-gpai
- OpenAI's May 2026 provenance update shows frontier providers moving toward multi-layer provenance, C2PA conformance, durable watermarking, and public verification tooling. RDLLM extends that ecosystem pattern from AI-origin detection to source attribution, royalty settlement, provider attribution cards, and assurance bundles: https://openai.com/index/advancing-content-provenance/
- Nature Machine Intelligence's March 2025 editorial on training-data transparency argues that resolving generative-AI copyright disputes starts with knowing which copyrighted data were used and where they entered the training pipeline. RDLLM converts that transparency goal into signed source registries, training-data summaries, source-footed responses, and provider-level discovery manifests: https://www.nature.com/articles/s42256-025-01023-9
- Samuelson's forthcoming PNAS paper on collective licensing for generative-AI training data emphasizes the normative, economic, and operational difficulty of compensating rightsholders at scale. RDLLM-L46 responds by separating policy choices from proof mechanics: providers can adopt licensing, statutory, pool, or marketplace regimes while still publishing conserved revenue-allocation evidence: https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6051014
- Peukert's September 2025 economics paper argues that royalties can help sustain fresh creative supply only if coverage is broad and administrative cost stays low. RDLLM-L46 therefore keeps allocation modes explicit, machine-verifiable, and hash-only for private billing fields, so creators can audit totals without forcing every provider into the same business model: https://papers.ssrn.com/sol3/papers.cfm?abstract_id=5532959
- PrivaRisk (Journal of Information Security and Applications, July 2025) and PrivChain-AI (Scientific Reports, December 2025) both point at the same finance-audit pattern: sensitive financial data should remain private while commitments, hashes, and authorized audit paths prove consistency. RDLLM-L47 applies that pattern to AI royalty finance exports: creators can verify that revenue pools reconcile without seeing customer records: https://www.sciencedirect.com/science/article/pii/S2214212625001310 https://www.nature.com/articles/s41598-025-32606-6
- CiteCheck (arXiv, 26 May 2026) verifies both whether a citation maps to a real scholarly work and whether metadata is faithful to that work. RDLLM uses the same philosophy at response time: source labels, claim-support spans, and footer entries must resolve to registered, hash-bound sources rather than model-written citation strings: https://arxiv.org/abs/2605.27700
- CiteAudit (arXiv v3, 1 May 2026) decomposes scientific citation verification into metadata extraction, retrieval, passage matching, reasoning, and calibrated judgment. RDLLM mirrors that decomposition by separating source materialization, claim-support spans, source quality, and certification failure modes: https://arxiv.org/abs/2602.23452
- Correctness is not Faithfulness in RAG Attributions (ICTIR 2025) shows why a citation can support a statement yet still be post-rationalized rather than the source the model genuinely relied on. RDLLM keeps corroborative source footers, contributive influence evidence, and independent audit attestation as separate proof channels: https://repository.tudelft.nl/record/uuid:bab4a12e-455b-49ea-bdf8-7f6a88478f60
- User-Centric Evidence Ranking (EACL 2026) argues that attribution systems should rank sufficient evidence early enough for humans to verify efficiently, instead of dumping redundant snippets. RDLLM maps that into visible source footers, support scores, claim-span hashes, and answer provenance cards: https://aclanthology.org/2026.eacl-long.340/
- SciTrue (EACL 2026) provides source-level accountability and evidence traceability for scientific claim verification, linking claim components to explicit sources that users can inspect and challenge. RDLLM generalizes that requirement beyond science through source verification reports and challenge-correction reports: https://aclanthology.org/2026.eacl-demo.27/
- SCITT/RFC 9943, RFC 9162 Certificate Transparency, Rekor, transparency-dev witness, in-toto, and the C2PA conformance program show that public verification ecosystems converge on signed statements, append-only Merkle logs, inclusion proofs, attestation subjects, and conformance evidence. Witness cosigning adds the missing anti-equivocation control: a log checkpoint should be rejected if an independent quorum has not seen the same checkpoint view. RDLLM-L18 turns the attribution proof pack into that kind of assurance bundle, RDLLM-L45 adds an independent audit attestation over the full public proof pack, RDLLM-L49 adds ongoing publication checkpoints, and RDLLM-L50 adds witness quorum and split-view detection over those checkpoints: https://www.rfc-editor.org/rfc/rfc9943.html https://www.rfc-editor.org/rfc/rfc9162.html https://docs.sigstore.dev/logging/overview/ https://pkg.go.dev/github.com/transparency-dev/witness https://github.com/in-toto/attestation https://opensource.contentauthenticity.org/docs/conformance/
- Sovereign Context Protocol (arXiv, March 2026) proposes an attribution-aware runtime access layer where every creator-content access is logged, licensed, and attributable. RDLLM's source-access trace and provider card make the same adoption pressure auditable across receipts, ledgers, and certification reports: https://arxiv.org/abs/2603.27094
- Cited but Not Verified (arXiv, 7 May 2026) shows that strong models can maintain high link validity and relevance while factual accuracy against cited sources remains much lower. RDLLM therefore reports citation integrity, fact support, and provider-level coverage as separate fields: https://arxiv.org/abs/2605.06635
- CiteCheck (arXiv, 26 May 2026) verifies whether generated scientific citations correspond to real scholarly works and faithful metadata, using retrieval and calibrated exact/minor/major labels. RDLLM maps that into source-confidence rows, citation-footer contracts, and verifier failure modes for metadata drift: https://arxiv.org/abs/2605.27700
- LLM hallucinations in the wild (arXiv, 8 May 2026) audits 111 million references and estimates 146,932 hallucinated citations in 2025 alone. RDLLM treats source identity and claim support as machine-checkable obligations rather than trust in model-written references: https://arxiv.org/abs/2605.07723
- MAVIS (AAAI 2026) shows that source attribution is already moving beyond text: multimodal systems need fact-level citations across visual and text documents. RDLLM-L27 declares multimodal source attribution support explicitly through a media attribution report instead of implying unimplemented coverage: https://doi.org/10.1609/aaai.v40i39.40585
- PaperTrail (arXiv, 24 February 2026) argues that document-level citations are not granular enough for scholarly verification and studies claim-evidence mappings. RDLLM therefore exposes public claim-support metadata, evidence span hashes, and private audit challenges for the redacted evidence text behind those hashes: https://arxiv.org/abs/2602.21045
- Efficient and Scalable Provenance Tracking for LLM-Generated Code Snippets (arXiv, May 2026) combines vector search and fingerprinting to trace generated code snippets back to likely source snippets. RDLLM treats code as another content class that needs the same source footer, license, and audit-challenge machinery as prose: https://arxiv.org/abs/2605.28510
- Concept Influence (arXiv, 16 February 2026) attributes behavior to semantic directions rather than only individual examples. RDLLM keeps runtime source attribution, training-value priors, and future model-internal influence evidence as separate ledger channels: https://arxiv.org/abs/2602.14869
- CUE-R (arXiv, April 2026) emphasizes intervention-style utility for individual retrieved evidence items in multi-step RAG. RDLLM's attribution-gap report treats every accessed item as an accountable object: https://arxiv.org/abs/2604.05467
- Source Attribution in RAG (arXiv, July 2025) applies Shapley-style attribution to retrieved documents but highlights LLM-call cost. RDLLM uses deterministic local weights by default and leaves room for Shapley/SHAP evidence as a higher-cost production signal: https://arxiv.org/abs/2507.04480
- CiteGuard (arXiv, October 2025) treats citation evaluation as citation-attribution alignment and uses retrieval-aware validation. RDLLM makes that alignment machine-checkable through footer labels, claim support, and receipt verification: https://arxiv.org/abs/2510.17853
- CiteAudit (2026): builds a benchmark and multi-stage verification pipeline for hallucinated scientific references, reinforcing RDLLM's split between citation identity, metadata fidelity, and claim support.
- User-Centric Evidence Ranking (EACL 2026): evaluates evidence ordering for attribution and fact verification, supporting RDLLM's choice to expose ranked source/support facts rather than opaque citation markers.
- SciTrue (EACL 2026): demonstrates source-level accountability and traceability for scientific claim verification, supporting RDLLM's source-materialization and creator-challenge layers.
- Correctness is not Faithfulness in RAG Attributions (arXiv, December 2024) shows that citation correctness alone is not enough; citations can be correct-looking but unfaithful to actual model use. RDLLM therefore records source-access traces separately from visible citations: https://arxiv.org/abs/2412.18004
- OLMoTrace (arXiv, April 2025) demonstrates real-time tracing from model output segments to massive training corpora for verbatim matches. RDLLM turns that class of output-to-training evidence into a receipt field and settlement signal: https://arxiv.org/abs/2504.07096
- W3C Verifiable Credentials Data Model v2.0 defines selective disclosure as fine-grained holder control and defines tamper-evident credentials and presentations. RDLLM's VC-shaped interop layer follows that portability pattern: https://www.w3.org/TR/vc-data-model/
- IETF RFC 9901 SD-JWT specifies issuer-signed JSON claims plus selected disclosures, digest verification, and salt entropy requirements. RDLLM's disclosure package mirrors those properties with per-path salts and root commitments: https://www.ietf.org/ietf-ftp/rfc/rfc9901.pdf
- W3C Data Integrity BBS Cryptosuites v1.0 targets selective disclosure and unlinkable derived proofs. RDLLM's current hash-based package is deliberately migration-friendly toward BBS-style production credentials: https://www.w3.org/TR/vc-di-bbs/
- W3C Data Integrity 1.0 explains that selective disclosure requires capable cryptosuites and can reveal only parts of a signed message while preserving verifiability. RDLLM applies that principle to attribution receipts: https://www.w3.org/TR/vc-data-integrity/
Papers and Relevance
- AIS, "Measuring Attribution in Natural Language Generation Models" (2023): defines attribution to identified sources as an evaluation target for generated statements.
- ALCE, "Enabling Large Language Models to Generate Text with Citations" (2023): evaluates citation quality alongside correctness and fluency, and shows citation support remains incomplete in long-form generation.
- RAGChecker (2024): argues for fine-grained diagnostics across retrieval and generation modules.
- RECLAIM, "Ground Every Sentence" (NAACL 2025): generates sentence-level citations through interleaved reference-claim generation.
- Think&Cite (ACL 2025): treats attributed text generation as a search problem and scores both generation progress and attribution progress.
- FRAMES (NAACL 2025): evaluates factuality, retrieval, and reasoning together for end-to-end RAG, showing that attribution belongs in the full task pipeline.
- VeriCite (2025): validates claims and supporting evidence before answer refinement, reinforcing claim-level verification.
- Source Attribution in RAG (2025): explores Shapley-style attribution for retrieved documents and highlights cost constraints.
- VISA (ACL 2025): localizes visual source evidence in document views, reinforcing that attribution must cover images, scanned PDFs, figures, screenshots, and multimodal source regions as well as plain text spans.
- SAFE (2025): classifies generated sentences before attribution so systems can decide whether a sentence needs no citation, one quote, or multiple quotes.
- OLMoTrace (2025): traces language model outputs to verbatim matches across very large training corpora, showing why runtime output-to-training provenance tools are becoming practical for open-data models.
- What Is Your Data Worth to GPT? (NeurIPS 2025, revised 2026): scales influence-function data valuation to recent LLMs and large training sets.
- DATE-LM (NeurIPS 2025): benchmarks data-attribution methods across training data selection, toxicity/bias filtering, and factual attribution, and reports that attribution methods have task-specific trade-offs rather than one universal winner.
- RISE (2026): compresses LLM influence evidence for retrospective attribution and prospective data valuation, reinforcing RDLLM's choice to expose model-native evidence as an auditable signal rather than a hidden payout oracle.
- Cited but Not Verified (2026): evaluates deep-research citations by checking link validity, topical relevance, and factual support against the actually cited content; it shows that surface citation quality can mask weak factual grounding.
- SourceBench (2026): scores cited web sources across relevance, factual accuracy, objectivity, freshness, authority/accountability, and clarity, reinforcing the need to score source quality separately from answer quality.
- LegalCiteBench (2026): shows that closed-book legal citation generation remains brittle and that concrete but incorrect authorities are a high-risk failure mode.
- Dependable RAG with factual confidence prediction (2026): argues for confidence measures around whether retrieved chunks support generated answers.
- DataDignity (2026): frames output auditing as pinpoint provenance, ranking candidate documents that best support a response.
- Probabilistic Attribution for LLMs (2026): attributes tokens through next-token probability inversions and entropy, reinforcing the need to expose uncertainty and token sensitivity separately from royalty splits.
- A Human-Centric Framework for Data Attribution in LLMs (2026): connects attribution methods to stakeholder objectives, negotiation, and sustainable creator incentives.
- Mechanistic Data Attribution (2026): traces interpretable LLM units back to training samples with influence-style methods, reinforcing that training-origin evidence and runtime citation evidence should be represented separately.
- DebugLM (2026): proposes training data provenance tags that trace model behavior to responsible datasets and enable source-specific remediation.
- Generation-Time vs. Post-hoc Citation (NeurIPS 2025 workshop): finds a practical trade-off between citation coverage and correctness, and recommends retrieval-centric post-hoc citation for many high-stakes applications.
- The Attribution Crisis in LLM Search Results (2025): documents a gap between web pages consumed by LLM search systems and sources credited to users.
- Document Overlap Is Not Evidence Continuity (ACL EvalEval 2026 submission): shows that document-level citation overlap can hide span-level turnover, motivating exact span hashes in audits.
- CiteGuard (arXiv 2025): validates citation attribution through retrieval-aware checking rather than relying only on LLM-as-judge.
- C2-Cite (2026): makes citation markers active knowledge pointers aligned to retrieved content.
- MedRAGChecker (2026): decomposes biomedical answers into atomic claims and checks evidence support.
- LUMINA (ICLR 2026): shows RAG can still hallucinate when the model underuses external context or overuses internal knowledge.
- MAVIS (AAAI 2026): extends source attribution to multimodal long-form VQA with fact-level citations.
- Samuelson (PNAS forthcoming, 2026 posting) and Peukert (SSRN 2025): show why compensation mechanisms need collective, statutory, or market-compatible revenue proof rather than only one-off permission or model-level claims.
- LLM hallucinations in the wild (2026): demonstrates that fabricated citations are entering scholarly corpora at scale.
- CiteTracer / Source or It Didn't Happen (arXiv, May 2026): reframes citation hallucination detection as field-level adjudication with a taxonomy of real, potential, and hallucinated citation errors, reinforcing RDLLM's typed source footers and verifier-readable citation metadata.
Standards and Policy Signals
- W3C ODRL 2.2 defines a rights language for permissions, prohibitions, duties, and constraints over content and services: https://www.w3.org/TR/odrl-model/
- MLCommons Croissant 1.1 adds machine-actionable provenance and structured usage policies for automated consent and licensing enforcement: https://docs.mlcommons.org/croissant/docs/croissant-spec-1.1.html
- SPDX 3.0.1 includes Dataset and AI profiles with hashes, intended use, limitations, and bill-of-materials relationships: https://spdx.github.io/spdx-spec/v3.0.1/
- The EU AI Act Article 53 requires GPAI providers to maintain a copyright policy and publish a sufficiently detailed summary of training content: https://eur-lex.europa.eu/eli/reg/2024/1689/oj
- The European Commission's 2026 FAQ says the public training-content template is mandatory under Article 53(1)(d), with updates generally required at six-month intervals or sooner for material changes: https://digital-strategy.ec.europa.eu/en/faqs/template-general-purpose-ai-model-providers-summarise-their-training-content
- C2PA clarified in January 2026 that Content Credentials are for provenance and authenticity, not TDM opt-out or AI-training preferences; RDLLM therefore treats C2PA-style provenance as complementary to ODRL/Croissant-style rights metadata: https://c2pa.ai/news/tdm-clarification
- The C2PA conformance program shows how provenance systems can move from specification to vetted product behavior: https://opensource.contentauthenticity.org/docs/conformance/
- IETF SCITT/RFC 9943 defines signed statements, transparency services, and verification roles for transparent digital supply chains: https://www.rfc-editor.org/rfc/rfc9943.html
- W3C Verifiable Credentials Data Model v2.0 is a W3C Recommendation for portable, proof-carrying claims and presentations: https://www.w3.org/TR/vc-data-model/
- OpenID Federation 1.0 defines a trust-chain and entity-statement pattern for cross-organization federation, which RDLLM maps to provider-scoped creator audit federation participants: https://openid.net/specs/openid-federation-1_0.html
- IETF RFC 9901 standardizes SD-JWT selective disclosure for JSON claims, including selected disclosures, digest verification, and salt-entropy security requirements: https://www.ietf.org/ietf-ftp/rfc/rfc9901.pdf
- W3C Data Integrity BBS Cryptosuites v1.0 describes selective disclosure and unlinkable derived proofs for VC data integrity: https://www.w3.org/TR/vc-di-bbs/
- W3C PROV defines a web provenance model built around entities, activities, agents, and relations, which RDLLM uses as the shape of answer-level provenance graphs: https://www.w3.org/TR/prov-overview/
Implementation Changes Made
- The engine now stores
answer_textseparately from renderedoutput. - Rendered output includes a
Sourcesfooter. - Source references expose source labels, owners, work IDs, chunk IDs, source URIs, evidence quotes, hashes, support scores, and payout shares.
- Claim support is computed per sentence with a lightweight local verifier and now records supporting work IDs, chunk IDs, evidence text, character offsets, and span hashes.
- The event hash includes the answer text, rendered output, payout shares, source references, claim evidence spans, and grounding report.
- Audits verify source-reference hashes in addition to royalty-share hashes.
- Conformance now fails if a supported claim's evidence span hash is absent from the rendered footer.
- Rights policies are now first-class: allowed uses, prohibited uses, jurisdiction, minimum royalty rate, and revocation are enforced before payout.
- Receipts now include policy decisions, and conformance fails if receipt rights decisions drift from the ledger.
- The engine exports a rights manifest aligned with ODRL, Croissant, SPDX, and PROV vocabulary concepts.
- The
certifycommand now emits an RDLLM-L0 through RDLLM-L186 report that tests grounded generation, external text attribution, escrow, rights blocking, receipt transparency, tamper detection, rights-manifest export, and grounding-quality scoring, registry dispute, escrow resolution, interoperability export, attribution-gap accountability, selective disclosure, trace exchange, and royalty-statement rollup, attribution-challenge correction, and provider-card verification, training-content-summary verification, assurance-bundle publication, answer-provenance-card verification, and source-materialization verification, response-envelope public verification, integration-profile contract verification, discovery-manifest publication, foundation-API attribution-profile verification, client-attribution-enforcement verification, persistent-memory-provenance verification, private-reasoning-attribution verification, post-training-signal-provenance verification, attribution-bom verification, creator-attribution-audit-index verification, creator-attribution-audit-federation verification, creator-audit-federation-transparency verification, derivative-lineage royalty verification, provenance-benchmark evaluation, counterfactual influence verification, multimodal media attribution verification, and model-internal signal attribution verification, rights-remediation verification, semantic text attribution verification, cross-provider attribution-exchange verification, portable conformance-vector verification, runtime federation-handshake verification, portable attribution-capsule verification, and response-release gate verification, proof-carrying-response verification, and serving-gateway verification, creator-license-contract verification, source-confidence verification, citation-footer-contract verification, private-audit-challenge verification, transitive-attribution-flow verification, and cross-provider-clearinghouse-settlement verification, verifiable-remittance verification, payment-execution attestation, payment-rail authenticity, creator-facing payout receipts, rendered-attribution-audit verification, training-memory provenance verification, and evidence-locked generation verification, emission evidence enforcement, live emission witness verification, live emission transparency verification, evidence-region binding verification, source-access lease verification, content-protocol ingestion verification, citation-reliance receipt verification, license-transaction receipt verification, grounded-source-footer verification, creator-audit federation verification, creator-audit federation transparency verification, source-freshness-audit verification, royalty-abuse-audit verification, and consent-revocation-propagation verification, citation URL-health verification, composite-foundation-adapter verification, foundation-provider-conformance verification, foundation-runtime-adapter verification, and foundation-runtime-router verification, and foundation-model-deployment-attestation verification, and universal-composition-receipt verification, and universal-composition-settlement verification, and universal-foundation-model-contract verification, and universal-invocation-guard verification, and universal-invocation-coverage verification, and universal-invocation-witness verification, universal-content-credential verification, universal-rdllm-passport verification, universal-adoption-standard verification, universal-interop-test-kit verification, and universal-context-provenance-bridge verification, and universal-citation-verification-contract verification, universal-grounded-reuse verification, universal-training-serving-contract verification, universal-confidential-attribution-audit verification, and universal-reliance-correction-ledger verification, universal-foundation-adoption-kernel verification, and universal-provider-adapter-harness verification, and universal-provider-drift-sentinel verification, and universal-attribution-negotiation-handshake verification, and universal-negotiated-invocation-enforcement verification, and universal-certification-trust-federation verification, and universal-foundation-provider-adoption-pack verification, and universal-industry-adoption-root verification, and universal-reference-implementation-distribution verification, universal-live-attribution-proof verification, universal-foundation-model-release-passport verification, and universal-composite-RDLLM-contract verification, and universal-foundation-provider-binding-matrix verification, and universal-provider-conformance-runner-receipt verification, and universal-production-invocation-admission verification, and universal-source-grounded-response-receipt verification. L158 turns static provider conformance into a continuously replayed drift control: provider API schemas, SDK response shapes, model aliases, streaming events, gateway transforms, citation locators, source-footer rendering, and settlement meters must keep matching the published RDLLM contract, or grounded display and creator payout release are revoked until remediation is published. L159 adds request-time negotiation: a client and provider route must agree on source footers, citation locators, claim provenance, model alias resolution, drift-sentinel status, telemetry, copy/export status links, privacy redaction, and settlement meters before generation, directly addressing recent findings that citation-looking output can remain unsupported unless source and metadata commitments are bound before answers are produced. L160 adds invocation enforcement: every actual SDK call, gateway proxy, stream, tool/MCP call, retrieval context, batch callback, fallback route, and cache reuse must bind an invocation receipt to the negotiated attribution contract before display or payout, aligning RDLLM with recent GenAI observability and citation-verification work that requires traceable calls and source-support metadata rather than self-asserted citations. L161 adds the portable trust-federation layer: attribution footers, invocation receipts, creator-settlement proofs, and anti-hallucination reliance signals must be certified through trust anchors, accredited conformance labs, trust marks, verifiable credentials, transparency inclusion, revocation status, and relying-party policy before a provider claim is treated as grounded rather than self-asserted. L162 turns that trust chain into a provider-neutral adoption pack: all hosted APIs, cloud gateways, OpenAI-compatible routers, local open-weight runtimes, and enterprise proxies must expose the same adapter evidence, standard exports, fail-closed invocation gates, source-footer delivery, telemetry mappings, copied-output status links, creator-audit paths, regulator exports, and settlement-release guards before attribution or anti-hallucination reliance can be claimed. The L117 layer adds platform-scale settlement integrity: source-farm, sybil, duplicate-source, reciprocal-boosting, undisclosed synthetic-source, and direct-payout concentration signals must route suspicious value to abuse-review escrow before direct payout. The L118 layer adds stale-rights prevention: revocation, opt-out, lease-expiry, and license-change events must propagate to source selection, footers, memory, private reasoning, post-training signals, attribution exchange, downstream notices, and settlement before future use or direct payout can continue. The L71 layer adds independent verifier quorum checks so direct settlement requires external signed replay agreement, not only provider-produced proof. The L72 layer adds bonded verifier accountability: accepted verifier attestations must bind to active trust-registry identities, non-revoked key hashes, slashable bond rows, conflict disclosures, and no open accountability challenge before settlement can leave escrow. The L73 layer adds receipt transparency consistency: required usage receipts must appear in append-only transparency-log snapshots with valid inclusion proofs, matching receipt payload/envelope hashes, and no split-view roots before verifier-approved settlement can leave escrow. The L74 layer adds watchtower challenge settlement: registered independent watchtowers must attest to the L73 subject, and open or accepted public challenges force settlement into escrow with slashing and bounty commitments. The L75 layer adds output provenance binding: copied or exported output must bind to proof-carrying responses, serving-gateway hashes, attribution capsules, watchtower-cleared settlement, content credentials, watermark commitments, fingerprint commitments, and public verification paths without exposing raw generated text. The L40 level rejects source-row drift, claim-anchor drift, inactive license or royalty duties, failed confidence rows, signature drift, and private text leakage before a client renders the public source footer. The L41 level rejects nonce replay, hidden-path drift, opening mismatch, and private evidence or payout leakage in audit challenges. The L42 level rejects copied-output marker loss, copied-body hash drift, upstream source row loss, transitive payout non-conservation, and copied prompt/output leakage. The L43 level rejects clearinghouse hash drift, non-ready transitive inputs, duplicate rows paid instead of held, payable/escrow conservation drift, and private prompt/output/source/copied-output leakage. The L44 level rejects remittance drift, missing payout-account hashes, inactive license terms, missing holds, and settlement overclaims. The L45 level rejects provider-only audit claims, missing audit discovery surfaces, failed public verifier replay, proof-pack hash drift, remittance amount drift, footer suppression, and private text or payout-account leakage in the audit attestation. The L46 level rejects non-conserved source revenue, event gross-revenue drift, creator-pool drift, missing receipt bindings, raw billing fields, and revenue-allocation tampering. The L47 level rejects finance-record hash drift, missing external record hashes, source-pool mismatch, total gross-revenue mismatch, unmapped finance records, unbacked allocation sources, private customer fields, and finance-attestation tampering. The L48 level rejects proof-dependency cycles, unknown dependencies, stale replay orders, overclaimed publication edges, private-field leakage, and graph tampering. The L49 level rejects publication-monitor checkpoint tampering, required public proof-surface removal, certification downgrades, snapshot-proof drift, non-reproducible hashes, and private-field leakage. The L50 level rejects publication-witness signature drift, missing witness quorum, split-view equivocation, inconsistent monitor subjects, and monitor payload leakage. The L51 level rejects unknown artifact signers, invalid artifact signatures, wrong witness keys, revoked active keys, broken key rotations, registry hash drift, trust-registry signature drift, and raw key-material leakage. The L52 level rejects non-reproducible certification report hashes, failed case-status roots, below-L51 attested reports, certification-attestation signature drift, missing trust-registry binding, and private case-payload leakage. The L53 level rejects copied-code scoring drift, missing compatible-owner payouts, incompatible copied code released without escrow/review, creator-pool non-conservation, private code leakage, and code-attribution report tampering. The L54 level rejects direct settlement without trusted signed ownership evidence, duplicate or weak ownership claims that bypass escrow, claim-row hash drift, private work/payment/issuer leakage, and claim-verification report tampering. The L55 level rejects stale, unreachable, footer-unbound, source-mismatched, or claim-span-mismatched citation sources and source-availability report tampering. The L56 level rejects cited spans that are not top-ranked minimal sufficient evidence, ambiguous decoy margins, broken source-availability bindings, footer-binding drift, and evidence-sufficiency report tampering. The L57 level rejects unaddressed counterevidence, answer acknowledgement drift, broken sufficiency bindings, and counterevidence report tampering. The L58 level rejects detached late-grounding reports by requiring source availability, evidence sufficiency, and counterevidence reports to be embedded and release-gate verified in the response envelope. The L59 level rejects unsupported factual sentences anywhere in the public answer surface. The L60 level rejects post-hoc citation attachment by requiring verified claims to be present in pre-generation context commitments. The L61 level rejects source packets that can act as instruction/control channels or mutate attribution and payout policy, and requires source-boundary telemetry to replay against the trace and generation context closure report. The L62 level rejects decision provenance graphs with missing proof edges, unknown influence channels, private prompt/source text, payout rows influenced directly by retrieved source text, or release decisions that do not bind response envelope, release gate, trace exchange, attribution capsule, and source-boundary proofs. The L63 level rejects calibrated attribution reports with benchmark drift, missing decision provenance, private text leakage, undisclosed low-confidence attribution, or confidence bounds that cannot be reproduced from the provenance benchmark. The L64 level rejects source-authenticity reports with unsigned or untrusted origin evidence, synthetic-source nondisclosure, source-farm or poisoning risk above policy thresholds, missing archive consensus, active-license drift, private text leakage, unauthenticated direct payout, or source-authenticity report tampering. The L65 level rejects streamed responses whose chunk lengths, chunk hashes, previous-chain hashes, final chain hash, gateway delivered-output hash, proof timing, or final attribution footer cannot be replayed from the public proof-carrying response and serving gateway report. - The L66 level rejects conversation ledgers that skip parent turns, reorder the turn chain, drop inherited source-obligation hashes, fail proof/gateway/stream replay, omit visible inherited source rows, or disclose private prompt/raw-output fields in ledger rows.
- The L67 level rejects agent-tool attribution ledgers that fail to bind tool observations to trace spans, visible source rows, supported claims, conversation-level source obligations, and raw-tool-output privacy boundaries.
- The L68 level rejects pinpoint provenance reports that accept topical anti-documents, publish footer sources without answer-critical fact support, fail to escrow uncertain claims, drift from private replay inputs, or leak prompt, response, claim, source, or evidence-phrase text.
- The L69 level rejects citation identity reports that allow fabricated, unresolved, metadata-swapped, or claim-unsupported citations into canonical footers or citation-linked payout. This directly answers the recent citation-verification finding that link validity, topical relevance, and source labels are insufficient unless the citation identity and claim support can be replayed against authority records.
- The L70 level rejects direct settlement unless source confidence, source
authenticity, evidence sufficiency, counterevidence adjudication, pinpoint
provenance, and citation identity agree on the same event. This converts recent
multi-source attribution and citation-verification work into a settlement rule:
disagreement is not silently averaged away; it becomes public
attribution_consensus_escrow. - The L71 level rejects direct settlement unless multiple independent verifier
attestations sign the same artifact root for the attribution consensus report,
provider card, certification report, and integration profile. This responds to
the same failure mode highlighted by recent citation-audit work: a plausible
footer or internally consistent proof pack is not enough unless an external
verifier can replay the cited evidence and sign the result. Disagreement
produces
verifier_quorum_escrowrather than direct settlement: https://arxiv.org/abs/2605.06635 https://arxiv.org/abs/2605.27700 https://arxiv.org/abs/2604.21193 - The L77 level rejects settlement-finality claims unless the instruction-only remittance report is reconciled against hash-only external processor records. This closes the gap left by attribution and clearing papers: they can identify which work should be credited and paid, but they do not prove that a payment rail or escrow rail actually executed the instruction. RDLLM-L77 requires matching amounts, currencies, end-to-end IDs, payout/escrow account hashes, processor statuses, duplicate-record checks, unmatched-record checks, hold preservation, and raw-bank-field rejection before a provider can mark payout value as externally executed. The resulting report is still privacy-preserving: public verifiers see hashes, statuses, counts, and reconciliation IDs, not creator bank details.
- The L78 level rejects payout-finality claims unless registered external payment or escrow processors sign the settlement batches behind the L77 execution rows. This closes the next gap after reconciliation: a provider can no longer present internally consistent processor rows unless the trust registry recognizes the processor identity and the processor signature covers the batch hashes, row hashes, totals, currencies, counts, and statuses. Public verifiers still see only hashes, signatures, roles, registry entries, and aggregate totals.
- The L79 level makes payout proof creator-facing rather than only auditor-facing.
Each
rdllm-creator-payout-receipt-report/v1row binds a credited creator, work, chunk list, origin hashes, clearinghouse settlement row, remittance instruction, execution row, processor record hash, signed rail batch, and paid, escrowed, or held status. The report rejects totals drift, missing signed rail coverage, and private prompt, source, customer, tax, or raw account fields. - The L80 level makes attribution proof user-facing rather than only artifact-facing.
Each
rdllm-rendered-attribution-audit/v1report parses the exact Markdown displayed to the user, verifies inline[S#]markers against response-envelope source labels, verifies footer source rows against the citation footer contract, and verifies claim-evidence span rows against answer coverage, source availability, evidence sufficiency, and counterevidence reports. This directly responds to recent citation-verification work: a generated footer is not enough unless the visible marker, footer row, source URI, content hash, and claim span all replay from signed evidence. - Current streaming model APIs return incremental chunks over event streams, so a final proof object is not enough for production adoption: the proof must commit each partial emission and prove the footer arrived by stream completion. RDLLM-L65 turns this into a hash-chain manifest rather than treating streaming as a UI concern: https://developers.openai.com/api/docs/guides/streaming-responses
- Current stateful response APIs carry prior responses into later turns through
conversation objects or
previous_response_id. RDLLM-L66 adds a conversation ledger so this state does not become an attribution laundering channel: https://developers.openai.com/api/docs/guides/conversation-state - Current model APIs expose hosted and remote tools, including web search, file search, function calling, and MCP. RDLLM-L67 treats each tool observation as a provenance-bearing evidence event rather than trusting the final citation list: https://platform.openai.com/docs/guides/tools?api-mode=responses
- Current agent SDK tracing records LLM generations, tool calls, handoffs, guardrails, and custom events, which is the right substrate for a replayable source-obligation ledger but not by itself a royalty or citation proof: https://openai.github.io/openai-agents-js/guides/tracing
- H-RAG at SemEval-2026 Task 8 shows why multi-turn RAG needs separate treatment: its generation task requires accurate answers and faithful grounding in retrieved evidence across conversation turns. RDLLM-L66 turns that research pressure into a production invariant by requiring dependent turns to inherit source obligations: https://arxiv.org/abs/2605.00631
- AILS-NTUA at the same SemEval-2026 task decomposes reference-grounded generation into evidence span extraction, candidate drafting, and calibrated multi-judge selection, reinforcing that multi-turn grounding needs explicit evidence-state handling rather than a final-answer-only citation pass: https://arxiv.org/abs/2603.10524
- Recent citation-aware and source-attribution RAG papers reinforce that sources are operational evidence, not decorative links. Hybrid retrieval/reranking with a claim judge, Shapley-style source attribution, and intervention-based evidence utility all motivate hash-bound source rows that survive generation and session state: https://arxiv.org/abs/2605.01664 https://arxiv.org/abs/2507.04480 https://arxiv.org/abs/2604.05467
- Recent agent-provenance papers make the L67 boundary sharper. TRACER argues for sentence-level provenance records tied to supporting tool turns and evidence units; TRACE and EnvTrustBench evaluate trajectories instead of only final answers; ARGUS and AuthGraph show why provenance/authorization must be tracked through tool parameters and untrusted context. RDLLM-L67 adapts those ideas to user-visible source rows and creator royalty obligations: https://arxiv.org/abs/2605.09934 https://arxiv.org/abs/2510.02837 https://arxiv.org/abs/2605.08828 https://arxiv.org/abs/2605.03378 https://arxiv.org/abs/2605.26497
- Deep-research citation audits show that links and cited source lists are not sufficient unless source accessibility, relevance, and factual support are checked. RDLLM-L67 therefore requires the tool observation that produced a source row to be replay-bound to the claim and footer: https://arxiv.org/abs/2605.06635
- ISO 20022 payment initiation and remittance-advice materials give RDLLM a practical shape for payment-file rows: an instruction should carry amount, currency, creditor identity/account reference, end-to-end ID, and remittance information for reconciliation. RDLLM-L44 uses that idea while keeping only payout-account hashes in the public artifact: https://www.iso20022.org/sites/default/files/2020-12/ISO20022_MDRPart2_PaymentsInitiation_2020_2021_v1_ForSEGReview.pdf
- W3C Verifiable Credentials 2.0 and W3C PROV reinforce the portability target: remittance claims and attribution chains should be independently verifiable statements, not rows trapped inside one provider database: https://www.w3.org/TR/vc-data-model/ https://www.w3.org/TR/prov-dm/
- Recent 2026 data-attribution work continues to make the same point from the model side: attribution evidence is becoming more scalable, but market adoption requires standard artifacts that bind influence evidence, visible source footers, clearing, and remittance. Examples include RAG attribution surveys, zeroth-order influence approximation, uncertainty-based data attribution, and rescaled influence functions: https://arxiv.org/abs/2601.19927 https://openreview.net/forum?id=KYaNRqJ7ho https://openreview.net/forum?id=IKB9uhMVH9 https://openreview.net/forum?id=edD3qTmoea
- DataDignity (arXiv, May 2026) frames robust training-data attribution as pinpoint provenance under paraphrase, hard anti-documents, and transformed query conditions. RDLLM-L30 now adds a runtime semantic text attribution report with concept fingerprints, hard-decoy margins, source footer rows, accepted-owner payouts, and semantic-text escrow: https://arxiv.org/abs/2605.05687
- Cited but Not Verified (arXiv, May 2026), CiteAudit (arXiv, February/May 2026), and User-Centric Evidence Ranking (EACL 2026) reinforce that footer sources must resolve to evidence users can inspect rather than to plausible citation strings. RDLLM-L55 therefore binds footer rows to reachable or archived source snapshots, registered content hashes, and cited claim-span hashes, while RDLLM-L56 ranks candidate spans and requires the cited span to be the minimal sufficient support with a margin over decoys and RDLLM-L57 rejects unaddressed contradiction candidates rather than treating citations as presentation text. RDLLM-L58 then requires those three reports to be embedded and gate-bound before a provider that claims L57 can emit the response, RDLLM-L59 requires the final answer surface itself to be claim-covered before release, RDLLM-L60 requires those claims to be closed over pre-generation context commitments, RDLLM-L61 verifies that cited source packets were evidence-only data rather than instruction or control channels, and RDLLM-L62 exposes a hash-only influence graph for the final claim, footer, payout, and release decisions: https://arxiv.org/abs/2605.06635 https://arxiv.org/abs/2602.23452 https://aclanthology.org/2026.eacl-long.340/
- Each event now includes a
rdllm-grounding-quality/v1report, and the certification suite reaches RDLLM-L7 by proving source accessibility, citation integrity, evidence relevance, fact support, policy alignment, and payout alignment. - The engine now emits
rdllm-claim-registry/v1reports that detect duplicate and near-duplicate ownership claims across creators. - Registry decisions are bound into ledger events, receipts, public receipts, and conformance checks.
- If registry enforcement is enabled and a matched work has an open ownership
conflict, the creator pool routes to
registry_dispute_escrowand the output verdict becomesblocked_by_registry. - The certification suite now reaches RDLLM-L8 by proving duplicate-claim detection, registry-dispute escrow, and registry-report hash binding.
- The settlement layer now emits
rdllm-escrow-resolution/v1reports that release registry-dispute escrow after resolution while preserving original event hashes. - The certification suite now reaches RDLLM-L9 by proving settlement reports recompute from the ledger and resolution, drain registry-dispute escrow, and release the held value to the resolved owner.
- The interop layer now emits
rdllm-interop/v1bundles: a VC-shaped receipt credential, a PROV-shaped answer graph, and optional VC-shaped escrow-settlement credential. - The certification suite now reaches RDLLM-L10 by proving bundle hashes, credential proofs, graph hashes, receipt bindings, and settlement-report bindings are all reproducible and tamper-sensitive.
- Each event now includes a source-access trace for retrieval and text-match paths
plus a
rdllm-attribution-gap/v1report that reconciles accessed, cited, paid, policy-withheld, and registry-withheld sources. - The certification suite now reaches RDLLM-L11 by proving closed attribution gaps for normal answers, detecting paid-but-hidden footer tampering, and treating rights-blocked or registry-blocked sources as accountable escrow rather than silent disappearance.
- Receipts now include
rdllm-selective-disclosure/v1roots and private per-path salts. Public packages disclose model/event/source/claim-support facts while redacting source quotes, evidence text, source-access traces, economics, detailed rights decisions, and registry decision internals. - RDLLM-L12 proves selective-disclosure packages verify standalone, verify against the private signed receipt, expose public claim support, redact private fields, and reject public receipt or leaf tampering.
- Receipts and interoperability bundles now include
rdllm-trace-exchange/v1artifacts aligned with OpenTelemetry GenAI spans. Verification proves source access, footer citations, claim-support spans, and receipt telemetry hashes agree. - The certification suite now reaches RDLLM-L13 by rejecting omitted provider source-access spans and citation relabeling, while proving trace summaries match the signed receipt.
- The statement layer now emits
rdllm-royalty-statement/v1reports that aggregate many events into creator, escrow, work, source-usage, receipt, and trace commitments without disclosing private prompts, outputs, quotes, evidence text, or matched text. - The RDLLM-L14 certification case proves aggregate payout conservation, receipt-root binding, trace-root binding, source-usage commitments, private-text redaction, and tamper rejection for creator statement drift.
- The challenge layer now emits
rdllm-attribution-challenge/v1reports that let creators contest omitted or weak attribution while binding the original event, optional statement, challenged content hash, overlap evidence, and non-rewrite remedy. - The certification suite now reaches RDLLM-L15 by proving accepted omitted-source correction, already-credited no-action, private-text redaction, and remedy tamper rejection.
- The provider-card layer now emits
rdllm-provider-attribution-card/v1reports that disclose certification level, coverage, supported evidence channels, public surfaces, challenge policy, limitations, and evidence roots without private text. - The RDLLM-L16 certification case proves provider-card binding to certification evidence and ledger coverage, private-text redaction, stale-level rejection, and coverage tamper rejection.
- The training-summary layer now emits
rdllm-training-content-summary/v1reports aligned with EU GPAI training-content summaries, Croissant usage policies, SPDX AI/Dataset BOMs, and ODRL rights metadata. - The certification suite now reaches RDLLM-L17 by proving training-rights coverage, license/use counts, policy roots, content-hash roots, training-value commitments, stale-certification rejection, and private-work-text redaction.
- The assurance layer now emits
rdllm-assurance-bundle/v1reports that publish hash-only proof-pack entries, Merkle roots, inclusion proofs, and profile flags aligned with SCITT, Rekor, in-toto, and C2PA-style conformance assurance. - The certification suite now reaches RDLLM-L18 by proving assurance bundles cover required artifacts, disclose no private payload text, verify inclusion proofs, and reject artifact-hash or proof tampering.
- The answer-card layer now emits
rdllm-answer-provenance-card/v1reports that bind visible source footers, claim span hashes, receipt hashes, trace hashes, grounding verdicts, and attribution-gap verdicts without disclosing private text. - The certification suite now reaches RDLLM-L19 by proving answer provenance cards verify against the ledger event, receipt, and trace, redact private prompt/source text, and reject card or footer tampering.
- The source-verification layer now emits
rdllm-source-verification-report/v1reports that prove cited source hashes, quote hashes, claim evidence hashes, offsets, visible footer spans, and answer-card bindings materialize from the registered corpus without publishing private text. - The certification suite now reaches RDLLM-L20 by rejecting fabricated source hashes, stale answer-card bindings, and claim evidence that cannot be reproduced from registered content.
- The response-envelope layer now emits
rdllm-response-envelope/v1packages that carry the rendered answer, answer provenance card, source verification report, public receipt, provider card, and certification posture as one API-verifiable object. - The certification suite now reaches RDLLM-L21 by rejecting rendered-output tampering and embedded-artifact tampering inside the response envelope.
- The integration-profile layer now emits
rdllm-integration-profile/v1contracts that declare generation and verification endpoints, required headers, embedded artifact requirements, verifier commands, schemas, public surfaces, readiness checks, and proof hashes. - The certification suite now reaches RDLLM-L22 by rejecting integration-profile tampering, provider-surface drift, certification regression, and response-envelope drift.
- The discovery-manifest layer now emits
rdllm-discovery-manifest/v1reports for/.well-known/rdllm.json, cataloging public artifact paths, artifact hashes, API contract hashes, schemas, verifier commands, readiness checks, and bound proof hashes. - The certification suite now reaches RDLLM-L23 by rejecting discovery-path tampering, artifact-hash drift, provider-surface regression, certification regression, and assurance bundles that omit integration-profile or response-envelope evidence.
- The lineage layer now emits
rdllm-lineage-report/v1reports that bind source royalty shares to derivative-work edges, upstream content hashes, recursive pass-through obligations, and payout-conservation commitments. - The lineage certification layer rejects missing upstream works, lineage cycles, declared upstream hash drift, payout drift, private-text leakage, and report tampering.
- The provenance-evaluation layer now emits
rdllm-provenance-evaluation/v1reports that replay clean-source, paraphrase, hard-decoy, unattributed-escrow, and derivative-lineage cases with hashed benchmark inputs and ranked source evidence. - The certification suite now reaches RDLLM-L25 by rejecting provenance-evaluation tampering, omitted cases, source-ranking drift, signature drift, and private benchmark/source-text leakage.
- The counterfactual layer now emits
rdllm-counterfactual-influence/v1reports that remove each credited work, replay the same prompt and answer, and publish source-removal status, payout reallocation status, substitute-source pressure, and influence margins without private text. - The public artifact pack now includes
examples/counterfactual_corpus.jsonandartifacts/counterfactual_ledger.json, a decisive-source vector whose signed counterfactual report has a positive influence margin. - The certification suite now reaches RDLLM-L26 by rejecting counterfactual-report tampering, ablation replay drift, signature drift, and private prompt, answer, source, quote, or claim-evidence leakage.
- The media layer now emits
rdllm-media-attribution/v1reports for image, audio, video, 3D, and text-shaped signatures. Reports publish only hashes, score components, ranked candidates, payout shares, and escrow decisions, while private descriptors and perceptual hashes remain verifier-side inputs. - The certification suite now reaches RDLLM-L27 by rejecting media-attribution tampering, replay drift, payout drift, signature drift, and raw media or descriptor leakage.
- The model-signal layer now emits
rdllm-model-signal-attribution/v1reports that bind private provider telemetry to an explicitrdllm-attribution-contract/v1, scalar signal scores, accepted owner payouts, and model-signal escrow without disclosing raw hidden states, token logits, private prompts, private outputs, or chain-of-thought. - The certification suite now reaches RDLLM-L28 by rejecting model-signal tampering, replay drift, payout drift, signature drift, unsupported attribution contracts, weak-score promotion, and private telemetry leakage.
- The rights-remediation layer now emits
rdllm-rights-remediation/v1reports that compare previous and updated policy roots, preserve historical event hashes, prove denied future training, retrieval, generation, external-attribution, display, and quote use after revocation, and verifyrights_conflict_escrowwithout leaking work text, prompts, outputs, matched text, or private ledger payloads. - The certification suite now reaches RDLLM-L29 by rejecting remediation replay drift, historical-event rewrites, future-use enforcement drift, escrow drift, signature drift, and private text leakage.
- The semantic-text layer now emits
rdllm-semantic-text-attribution/v1reports that rank paraphrased or external outputs against registered text using lexical, concept, source-distinctiveness, sequence, n-gram, and decoy-margin evidence, then publish source-footer commitments, accepted-owner payouts, and semantic escrow without exposing prompt, output, matched, or source text. - The certification suite now reaches RDLLM-L30 by rejecting semantic attribution replay drift, hard-decoy payment, escrow drift, footer tampering, signature drift, and private text leakage.
- The cross-provider exchange layer now emits
rdllm-attribution-exchange/v1manifests. This turns attribution from a provider-local proof into a relayable contract: downstream model providers can import provider cards, certification, integration profiles, discovery manifests, response envelopes, assurance bundles, and semantic-text reports by hash, preserve public source-footer rows, and keep ambiguous or unlicensed value in escrow. - This addition responds to the recent citation-verification and provenance papers: DataDignity shows that robust attribution must survive paraphrase and hard decoys; Cited but Not Verified, CiteCheck, CiteAudit, PaperTrail, and SciTrue show that source claims need retrievable evidence and claim-level grounding; and LLM hallucination audits show that fabricated or misassigned citations are a credit-allocation problem as well as an accuracy problem. RDLLM-L31 therefore requires an importable exchange with explicit source-footers and hash-bound upstream proof, not merely inline citations in generated text.
- The certification suite now reaches RDLLM-L31 by rejecting exchange replay drift, missing imported artifacts, missing exchange schema or verifier, missing discovery path, weak upstream certification, source-footer relay failure, hash drift, and private text leakage.
- The conformance-vector layer now emits
rdllm-conformance-vector-pack/v1reports with fixture hashes, expected public outcomes, verifier commands, and named negative mutations. This makes RDLLM testable as a public standard rather than a provider-specific promise. - The certification suite now reaches RDLLM-L32 by rejecting vector-pack hash drift, missing L32 vectors, missing negative mutations, weak upstream certification, missing verifier commands, expected-outcome drift, and private text leakage.
- The federation-handshake layer now emits
rdllm-federation-handshake/v1reports that bind requester/provider identity, negotiated RDLLM level, required artifact hashes, runtime headers, source-footer relay, escrow relay, and downgrade policy. This turns public proof publication into a runtime interop gate. - The certification suite now reaches RDLLM-L33 by rejecting handshake hash drift, stale artifact bindings, missing federation schemas or verifier commands, missing discovery paths, missing conformance vectors, disabled relay obligations, disabled downgrade protection, and private prompt, output, matched, source, or hidden-state leakage.
- The portable-attribution-capsule layer now emits
rdllm-attribution-capsule/v1reports that bind copied output markers, rendered-output hashes, delivered-body hashes, federation-handshake hashes, exchange hashes, conformance-vector hashes, provider proof hashes, C2PA-compatible assertion pointers, and SCITT-like statement subjects without exposing prompt, answer, matched, source, hidden-state, or private ledger payloads. - The certification suite now reaches RDLLM-L34 by rejecting capsule hash drift, stale artifact bindings, missing capsule schema or verifier commands, missing discovery paths, missing copy markers, copied-body tampering, incomplete runtime headers, signature drift, and private text leakage.
- The response-release-gate layer now emits
rdllm-response-release-gate/v1reports that make source attribution a serving-boundary decision. The gate binds the response envelope, answer provenance card, source verification report, attribution capsule, provider attribution card, and certification report, then emits only if footer labels, claim span prefixes, source materialization, capsule delivery contracts, provider public surfaces, and upstreamRDLLM-L34certification all verify. - The certification suite now reaches RDLLM-L35 by rejecting release-gate hash drift, failed gate checks, unsupported grounded claims, stale capsule binding, missing release-gate public surfaces, certification regression below L34, and private prompt, answer, matched, source, hidden-state, or ledger leakage. This is the direct implementation response to recent 2026 citation-verification papers: link validity, source relevance, and citation labels are useful, but source materialization and claim-level support must gate the answer before display.
- The proof-carrying-response layer now emits
rdllm-proof-carrying-response/v1objects that make the L35 decision enforceable at the API boundary. A released object carries the answer, capsule marker, release gate, response envelope, provider card, and certification report together; a failed gate produces a held-response object that suppresses the original answer. - The certification suite now reaches RDLLM-L36 by rejecting displayed-output tampering, missing capsule markers on released copied output, stale proof-response hashes, missing provider proof-carrying surfaces, certification regression below L35, blocked responses that leak the original answer, and private source/matched/hidden-state leakage.
- The serving-gateway layer now emits
rdllm-serving-gateway-report/v1objects that bind a concrete API request route to the proof-carrying response delivered at egress. It hash-commits the private prompt and raw model output, verifies the embedded proof response, and rejects delivered-output hash drift. - The certification suite now reaches RDLLM-L37 by rejecting gateway egress drift, missing provider gateway surfaces, certification regression below L36, embedded proof-response failures, and disclosure of private prompt/raw-output fields in the request context.
- The creator-license-contract layer now emits
rdllm-creator-license-contract/v1objects that bind registered source works to allowed AI uses, attribution duties, compensation duties, minimum creator-pool terms, revocation state, content hashes, and hashed payout-account commitments before use. This is the rights-side companion to source footers: an answer footer can now be checked against materialized sources, claim span hashes, serving egress proof, and pre-use license terms. - The certification suite now reaches RDLLM-L38 by rejecting license-scope drift, missing attribution or compensation duties, term-hash/signature drift, mismatched content hashes, and disclosure of private work text or raw payout accounts.
- The source-confidence layer now emits
rdllm-source-confidence-report/v1objects that bind answer provenance cards, source verification reports, and creator license contracts into public footer rows, claim confidence rows, and a hallucination taxonomy for fabricated sources, metadata drift, hash mismatch, footer omission, attribution suppression, license gaps, unsupported claims, and evidence-span gaps. - The certification suite now reaches RDLLM-L39 by rejecting source-confidence hash drift, stale answer-card or source-verification bindings, license-scope gaps, unsupported claim rows, failed footer rows, hallucination issue counts, and disclosure of private prompt, answer, source, evidence, or payout text.
- The citation-footer-contract layer now emits
rdllm-citation-footer-contract/v1artifacts that bind exact source display rows, claim display anchors, confidence labels, license status, royalty status, display order, row hashes, and rendered footer hash to a response envelope. - The certification suite now reaches RDLLM-L40 by rejecting citation-footer hash drift, source-row tampering, claim-anchor tampering, inactive license or royalty status, failed source-confidence rows, signature drift, and private prompt, answer, source, evidence, or payout-account leakage.
- The certification suite now reaches RDLLM-L41 by adding private audit challenge verification over redacted source-access, claim-evidence, rights, and payout paths; verifier tests reject replayed nonces and tampered opening hashes.
- The transitive-attribution layer now emits
rdllm-transitive-attribution-report/v1artifacts that bind downstream reuse of copied RDLLM output to the upstream attribution capsule, response envelope, answer-card source rows, copied-input hash, and pass-through settlement obligations. - The certification suite now reaches RDLLM-L42 by rejecting copied-output marker loss, copied-body hash drift, dropped upstream source rows, non-conserved transitive payout, and private copied-output or prompt leakage.
- The clearinghouse layer now emits
rdllm-clearinghouse-report/v1artifacts that ingestrdllm-royalty-statement/v1andrdllm-transitive-attribution-report/v1, normalize each obligation into payable, escrow, or held rows, deduplicate repeated artifact submissions and repeated transitive obligations, and bind the result with input, obligation, payable, escrow, held, duplicate, and policy roots. - The certification suite now reaches RDLLM-L43 by rejecting clearinghouse row tampering, duplicate payment, non-conserved payable/escrow totals, and private settlement text leakage.
- The remittance layer now emits
rdllm-remittance-report/v1artifacts that convert clearinghouse payable and escrow rows into instruction-only payment rows with payout-account hashes, creator-license bindings, ISO 20022-compatible reconciliation fields, preserved held rows, and explicit no-bank-settlement claims. - The certification suite now reaches RDLLM-L44 by rejecting remittance row tampering, missing payout-account hashes, inactive license terms, clearinghouse value drift, missing held rows, raw payout-account disclosure, and payment reports that imply executed settlement.
- The audit-attestation layer now emits
rdllm-third-party-audit-attestation/v1artifacts over the provider card, certification report, certification attestation, integration profile, discovery manifest, assurance bundle, response envelope, source confidence report, citation footer contract, clearinghouse report, remittance report, revenue allocation report, and finance ledger attestation. Settlement artifacts remain outside the stable assurance bundle and are bound here to avoid cyclic proof dependencies. - The certification suite now reaches RDLLM-L45 by rejecting provider self-attestation, missing audit surfaces, failed public verifier replay, proof-pack hash drift, remittance amount drift, stale artifacts, and private prompt, answer, source, evidence, or payout-account leakage.
- The revenue-allocation layer now emits
rdllm-revenue-allocation-report/v1artifacts over hashed billing, advertising, subscription, API, enterprise, and marketplace revenue sources, allocation policy, event allocation rows, receipt rollups, and creator-pool commitments. - The certification suite now reaches RDLLM-L46 by rejecting unsupported allocation modes, non-conserved source revenue, ledger gross-revenue drift, creator-pool drift, missing receipt bindings, duplicate event rows, raw billing field leakage, and allocation report tampering.
- The finance-ledger layer now emits
rdllm-finance-ledger-attestation/v1artifacts over hash-only finance export rows, revenue-source rollups, allocation report bindings, and privacy checks. The certification suite now reaches RDLLM-L47 by rejecting missing finance hashes, source-total mismatches, orphaned finance records, unbacked allocation sources, raw customer field leakage, and attestation tampering. - The proof-dependency layer now emits
rdllm-proof-dependency-graph/v1artifacts over hash-only artifact nodes, replay dependency rows, publication commitment rows, and topological verifier steps. The certification suite now reaches RDLLM-L48 by rejecting replay cycles, unknown dependencies, stale replay orders, publication edges that are not backed by the assurance bundle's own artifact entries, private-field leakage in the graph, and graph tampering. - The publication-monitor layer now emits
rdllm-publication-monitor/v1artifacts over hash-only public proof-surface snapshots, Merkle inclusion proofs, append-only checkpoint history, artifact diffs, and certification-regression checks. The certification suite now reaches RDLLM-L49 by rejecting required-artifact removal, non-reproducible hashes, broken inclusion proofs, checkpoint-chain tampering, certification downgrade, private-field leakage, and monitor tampering. - The receipt-transparency-consistency layer now emits
rdllm-receipt-transparency-consistency-report/v1artifacts over observed usage receipt-log snapshots, append-only prefix rows, split-view conflict rows, required receipt inclusion proofs, and settlement escrow rows. The certification suite now reaches RDLLM-L73 by rejecting omitted usage receipts, inconsistent Merkle roots, forked same-size log views, receipt payload/envelope drift, private-field leakage, and consistency-report tampering. - The watchtower-challenge-settlement layer now emits
rdllm-watchtower-challenge-settlement-report/v1artifacts over the L73 receipt-transparency subject, active watchtower trust-registry rows, watchtower attestations, public challenge rows, slashing evidence, bounty rows, and settlement escrow rows. The certification suite now reaches RDLLM-L74 by rejecting missing watchtower quorum, invalid watchtower signatures, open or accepted blocking challenges, private-field leakage, creator-pool drift, and watchtower-report tampering. - The output-provenance-binding layer now emits
rdllm-output-provenance-binding-report/v1artifacts over proof-carrying responses, serving-gateway output hashes, attribution capsules, L74 watchtower-cleared settlement, C2PA-style content credential assertions, watermark commitments, fingerprint-registry commitments, and public verification rows. The certification suite now reaches RDLLM-L75 by rejecting missing copy-survival signals, content-credential drift, gateway-output drift, watchtower-settlement failure, private text leakage, and binding-report tampering. - The post-release-discovery layer now emits
rdllm-post-release-discovery-report/v1artifacts over the base discovery manifest, output provenance binding report, proof dependency graph, proof-carrying response, serving-gateway report, attribution capsule, watchtower settlement, provider card, integration profile, and certification report. The certification suite now reaches RDLLM-L76 by rejecting stale base manifests, missing late output artifacts, catalog tampering, output-binding replay failure, self-referential hash cycles, and private text leakage. - The payment-execution layer now emits
rdllm-payment-execution-report/v1artifacts over remittance reports and hash-only external payment or escrow processor records. The certification suite now reaches RDLLM-L77 by rejecting missing settlement records, duplicate or unmatched processor rows, amount or currency drift, wrong payout or escrow account hashes, non-settled processor statuses, remittance-hold leakage into paid rows, raw bank/customer/tax fields, and payment-execution tampering. This makes payout finality independently verifiable instead of treating a remittance file as proof that money moved. - The payment-rail layer now emits
rdllm-payment-rail-attestation/v1artifacts over payment-execution reports, trust registries, and signed processor batch attestations. The certification suite now reaches RDLLM-L78 by rejecting unregistered processors, unauthorized processor roles, invalid signatures, uncovered execution batches, amount/status drift, and raw private payment fields. - The creator-payout-receipt layer now emits
rdllm-creator-payout-receipt-report/v1artifacts over clearinghouse, remittance, payment-execution, and payment-rail reports. The certification suite now reaches RDLLM-L79 by rejecting creator-visible payout receipts that cannot be replayed from attribution and settlement evidence, that lack signed rail-batch coverage, that drift in paid/escrow/hold totals, or that leak private payment, prompt, output, source, customer, or tax fields. - The rendered-attribution-audit layer now emits
rdllm-rendered-attribution-audit/v1artifacts over response envelopes, citation footer contracts, source availability, evidence sufficiency, counterevidence, and answer claim coverage reports. The certification suite now reaches RDLLM-L80 by rejecting visible Markdown whose inline source markers, source footer rows, or claim-evidence span rows drift from signed grounding artifacts, or whose audit leaks raw prompt, answer, source, or claim text. - The training-memory provenance layer now emits
rdllm-training-memory-provenance/v1artifacts over response envelopes, rendered attribution audits, creator-license contracts, training-content summaries, and registered source snapshots. The certification suite now reaches RDLLM-L81 by rejecting hidden memorized source spans in the exact displayed answer, stale snapshot roots, unlicensed training/generation/display use, and public reports that leak raw answer, source, prompt, or matched text. This closes the OLMoTrace/DataDignity gap where a model can rely on training memory even when retrieval logs and visible footers appear clean. - The evidence-locked generation layer now emits
rdllm-evidence-locked-generation/v1artifacts over response envelopes, answer-claim coverage reports, generation-context closure reports, citation footer contracts, rendered attribution audits, and training-memory provenance reports. The L82 certification case rejects locks created after generation start, support-required answer units without satisfied source/context/footer/claim-evidence locks, footer rows without a matching lock, hidden training-memory spans, and public reports that leak raw answer, source, prompt, claim, or context text. - The emission evidence enforcement layer now emits
rdllm-emission-evidence-enforcement/v1artifacts over response envelopes, answer-claim coverage reports, evidence-locked generation reports, proof-carrying responses, serving-gateway reports, and streaming attribution manifests. The certification suite now reaches RDLLM-L83 by rejecting streaming that starts before the evidence-lock/generation window, proof/gateway/stream output-hash drift, support-required units without satisfied locks, non-replayable stream chunks, and public reports that leak raw answer, source, prompt, claim, context, or chunk text. - The live emission witness layer now emits
rdllm-live-emission-witness/v1artifacts over emission enforcement reports and streaming attribution manifests. The certification suite now reaches RDLLM-L84 by rejecting missing independent witness quorum, preflight attestations created after the first streamed chunk, final stream-chain or chunk-subject drift, witness disagreement, invalid witness signatures, and public reports that leak raw prompt, answer, source, claim, context, or chunk text. - The live emission transparency layer now emits
rdllm-live-emission-transparency/v1artifacts over live witness reports and witness attestation subjects. The certification suite now reaches RDLLM-L85 by rejecting missing witness-report or witness-attestation log inclusion, non-reproducible log roots, append-only prefix violations, same-size split-view roots, payload hash drift, entry-type drift, invalid inclusion proofs, and public reports that leak raw prompt, answer, source, claim, context, chunk, or credential fields. This turns SCITT/Sigstore-style transparency from publication metadata into an emission-time attribution control. - The attested attribution runtime layer now emits
rdllm-attested-attribution-runtime/v1artifacts over live transparency, proof-carrying response, serving-gateway, and evidence-lock reports. The certification gate is RDLLM-L86 and rejects runtime measurement drift, untrusted attestors, stale or invalid quotes, missing attribution-enforcement capabilities, subject-binding drift, and public reports that leak private prompt, answer, source, chunk, or secret fields. This turns transparent attribution proof from an after-the-fact log into a claim about the measured code path that served the answer. - The claim-source attribution layer now emits
rdllm-claim-source-attribution/v1artifacts that replay every visible answer claim against candidate sources, Q&A nugget commitments, topical anti-documents, optional visual-region commitments, and LOO-style source contribution before a footer row or payout row is accepted. This directly incorporates the recent DataDignity finding that robust provenance must separate true answer support from topical resemblance, Chain of Evidence's requirement for pixel/bounding-box anchors in visual document RAG, GaRAGe and Trust-Align's claim-level grounding, attribution, and refusal framing, GuarantRAG's separation of parametric answers from evidence-backed refer answers, and source-attribution work that treats RAG attribution as a per-document contribution problem. The certification suite now reaches RDLLM-L87 by rejecting anti-document wins, insufficient LOO margin, unsupported claims, missing visual anchors for credited visual evidence, non-conserved attribution credit, footer omissions, replay drift, and public reports that leak raw prompt, answer, source, claim, nugget, or visual text. - The causal evidence-utility layer now emits
rdllm-causal-evidence-utility/v1artifacts that bind credited footer rows to current-turn retrieval/tool traces and REMOVE, REPLACE, DUPLICATE, and multi-source removal trials. This incorporates CUE-R's intervention-based view of per-evidence-item utility, AgenticRAG and retrieval-as-reasoning work on multi-step retrieval traces, MuRGAt's finding that fact-level multimodal reasoning can be correct while citations are wrong, and the recent attribution, citation, and quotation survey's call for unified evidence granularity. The L88 certification gate rejects spurious cited sources, prior-context citation drift, duplicate credit inflation, missing intervention trials, unsupported causal-use claims, non-conserved payout credit, and public reports that leak raw prompt, answer, source, claim, query, or observation text. - The parametric memory layer now emits
rdllm-parametric-memory-attribution/v1artifacts for claims answered from learned weights rather than current retrieval. It combines DataDignity's training-data provenance framing, knowledge-attribution probes for distinguishing memory from context, probabilistic token attribution for model-agnostic token sensitivity, ProToken-style token/client attribution, and protocol-layer attribution proposals such as SCP. The certification suite now reaches RDLLM-L89 by rejecting current-context contamination, anti-document wins, training-summary drift, weak memory or influence probes, non-conserved payout credit, and public reports that leak raw prompt, answer, source, claim, or probe text. - The style influence layer now emits
rdllm-style-influence-attribution/v1artifacts for generated outputs that imitate registered creator styles, voices, or modality-specific signatures without verbatim copying. It treats style credit as a stricter rights-controlled channel: accepted rows need active style-generation permission, sufficient style similarity, declared style intent, anti-style decoy separation, low content-copy overlap, footer disclosure, and creator-pool conservation. The L90 certification gate rejects copy-like outputs from style payout, revoked licenses, ambiguous style-profile ties, anti-style decoy wins, tampering, and public reports that leak private style exemplars or output text. - The model-lineage layer now emits
rdllm-model-lineage-attribution/v1artifacts for downstream models trained, fine-tuned, or distilled from attributed outputs, synthetic datasets, or teacher traces. It incorporates recent post-training data-lineage work such as Tracing the Roots and distillation/provenance research by treating training inheritance as an auditable royalty obligation rather than a one-time dataset artifact. The RDLLM-L91 layer rejects hidden synthetic training, missing teacher-distribution or teacher-output evidence, duplicate artifact over-credit, non-conserved future usage pools, revoked training rights, tampering, and public reports that leak private training examples, prompts, source text, or student outputs. - The black-box model-provenance layer now emits
rdllm-black-box-model-provenance/v1artifacts for API-visible models that may have been trained, fine-tuned, distilled, merged, or watermarked from protected upstream models without disclosure. It builds candidate provenance sets from challenge hashes, candidate scores, baseline distributions, adjusted p-values, watermark/fingerprint support counts, footer rows, and payout or escrow obligations. The RDLLM-L92 gate rejects insufficient challenge sets, excluding unrelated candidates, enforcing multiple-testing controls, conserving the provenance challenge pool, rejecting tampering, and preventing private prompts, source text, and model outputs from leaking into public reports. - The attribution dispute adjudication layer now emits
rdllm-attribution-dispute-adjudication-report/v1artifacts for cases where attribution, model provenance, ownership, or payout is challenged after a proof report exists. It incorporates recent provenance-testing, watermarking, behavioral-fingerprint, and audit-trail work by treating attribution as a contestable public decision: disputed value is frozen, claimant and respondent positions are admitted as hash-committed evidence, conflicted votes are excluded from quorum, appeal windows block premature release, and the final report publishes release, freeze, slash, bounty, and adjudicated footer rows without exposing prompts, outputs, source text, or challenge transcripts. The RDLLM-L93 gate rejects missing notice, failed quorum, weak evidence commitments, appeal-bypassing release, unconserved escrow, settlement drift, tampering, and private-text leakage. - The post-adjudication settlement adjustment layer now emits
rdllm-post-adjudication-settlement-adjustment-report/v1artifacts for the correction step after a dispute changes payout. The design borrows the proven platform pattern of separately held disputed revenue from YouTube Content ID, but makes it provider-neutral and verifiable for AI: original payment execution and creator payout receipt hashes are preserved, corrected entitlement rows produce top-ups or capped future-netting recoupments, open appeals freeze all movements, and creators receive hash-bound adjustment receipts. The RDLLM-L94 gate rejects historical payment rewrites, missing processor or receipt hashes, uncapped recoupment, appeal-bypassing release, unconserved corrections, tampering, and private payment/source-evidence leakage. - The residual corpus royalty layer now emits
rdllm-residual-corpus-royalty-report/v1artifacts for diffuse value learned from licensed training corpora when no visible answer source can be singled out with sufficient confidence. It incorporates recent DataDignity, DATE-LM, LLM-scale influence valuation, Concept Influence, RISE, LoRIF, and ZK-Value work by treating model-native valuation as evidence, not as an unchecked payout oracle: public reports bind training-content summary cohorts, creator-license terms, model-usage revenue rows, valuation evidence hashes, direct-attribution exclusions, creator-level caps, payable rows, escrow rows, and creator residual receipts. The RDLLM-L95 gate rejects unlicensed or low-confidence rows paid directly, double-counting of visible answer attribution, unconserved residual pools, creator-cap violations, missing residual receipts, tampering, and public leakage of prompt, output, source, training, customer, or payment text. The visible footer remains the answer-grounding surface; L95 only settles residual training value that survives rights and valuation checks. - The valuation-method audit layer now emits
rdllm-valuation-method-audit/v1artifacts before residual training value can be trusted for payout. It converts recent DATE-LM, DataDignity, RISE, LoRIF, Concept Influence, and ZK-Value directions into an executable audit: public method cards must cover every residual valuation method used, benchmark rows must cover known contributors, hard anti-documents, duplicate guards, confidence calibration, rights-denied escrow, and score stability, and privacy or zero-knowledge commitments must be present. The RDLLM-L96 gate rejects unaudited residual methods, benchmark drift, duplicate over-credit, unstable scores, missing privacy commitments, tampering, and public leakage of private benchmark, training, prompt, output, customer, or payment text. - The evidence-region binding layer now emits
rdllm-evidence-region-binding/v1artifacts after rendered attribution and claim-source attribution. It closes the remaining "right source, wrong passage" failure mode by binding every rendered claim span and footer span prefix to an exact source snapshot region, preserving page/line/character/bounding-box/timecode locators, and rejecting hard negative neighboring regions without publishing raw prompt, answer, source, customer, payment, or private evidence text. The RDLLM-L97 gate is intentionally stricter than a citation footer: a footer row is accepted only when the cited region itself supports the rendered claim. - The source access lease layer now emits
rdllm-source-access-lease/v1artifacts after evidence-region binding. It closes the "right source, no licensed access" failure mode by requiring directly settled consumed source rows to prove creator-issued active leases, matching access logs, license-contract permission, minimum creator-pool-rate compliance, and region binding. Denied, revoked, expired, or unleased usage is allowed only as hash-bound escrow, and public reports suppress raw prompt, answer, source, customer, and payout data. - The content-protocol ingestion layer now emits
rdllm-content-protocol-ingestion/v1artifacts after source-access leases. It closes the "external rights signal lost in translation" failure mode by requiring RSL, CoMP, SCP, ODRL, Croissant, robots.txt, C2PA/TDM, or custom publisher terms to be hash-preserved and carried into creator contracts, source-access leases, and escrow routing. Direct settlement fails when a direct source lacks an external protocol record, when the protocol denies the use, when rate or attribution duties drift, or when raw notices or private text leak. - The citation-reliance layer now emits
rdllm-citation-reliance-receipt/v1artifacts after content-protocol ingestion. It closes the "correct citation, unfaithful reliance" failure mode identified by recent RAG attribution work: every visible footer source must have a satisfied pre-generation evidence lock, a rendered claim-evidence row, accepted claim-source replay, accepted causal utility evidence, current-turn trace membership, source-access coverage, and protocol permission. Missing locks, non-causal sources, hidden paid sources, or private-text leakage downgrade the receipt toneeds_review. - The license-transaction layer now emits
rdllm-license-transaction-receipt/v1artifacts after citation-reliance receipts. It closes the dynamic authorization gap left by static rights signals: a directly settled source must have a signed license-server token, license-ledger inclusion proof, matching source-access nonce and metering hash, matching content-protocol terms, and a valid issued/ accepted/expiry window before direct settlement can leave escrow. - The grounded-source-footer layer now emits
rdllm-grounded-source-footer/v1artifacts after license-transaction receipts. It closes the user-facing confidence gap by requiring visible source rows to replay to confidence, availability, exact evidence regions, citation reliance, license transactions, and public verifier handles while redacting prompt, answer, source, evidence, and payout text. - The source-footer-delivery layer now emits
rdllm-source-footer-delivery/v1artifacts after proof-carrying response and serving-gateway evidence. It closes the display gap by binding the grounded footer to the response envelope, copied output, gateway egress hash, source labels, claim span handles, and client-visible verifier metadata. - The foundation API attribution profile layer now emits
rdllm-foundation-attribution-profile/v1artifacts after discovery and L103 delivery evidence. It closes the adoption gap by making the minimum attribution API surface explicit: response headers, embedded proof objects, well-known verifier paths, verifier commands, and fail-closed policy must be present before generic clients render an answer as attributed. - The client attribution enforcement layer now emits
rdllm-client-attribution-enforcement/v1receipts from the relying party. It closes the downstream stripping gap by proving the client observed the required L104 headers, replayed the embedded response envelope and source-footer-delivery receipt, matched source labels to delivered source rows, and blocked rendering when a required header or source label is missing. - The persistent memory provenance layer now emits
rdllm-persistent-memory-provenance/v1receipts after client attribution enforcement. It closes the cross-session memory gap: a stored assistant or agent memory cell must carry source labels, upstream proof hashes, license and retention policy, royalty carry-forward rows, and delete tombstones, and any later answer using that memory must display the carried labels in the verified footer. This follows recent persistent-memory agent work such as Memori and the broader MCP/tool ecosystem, where durable context can become an uninspected source of generated answers. - The private reasoning attribution layer now emits
rdllm-private-reasoning-attribution/v1receipts after persistent memory provenance. It closes the hidden-scratchpad and delegation gap: private reasoning steps, router handoffs, and memory-influenced synthesis must publish commitments, bind upstream proof hashes, preserve source labels and royalty rows, and prove those labels remain visible in the verified footer without exposing chain-of-thought. - The post-training signal provenance layer now emits
rdllm-post-training-signal-provenance/v1receipts after private reasoning attribution. It closes the RLHF/RLAIF/RLVR laundering gap: preference labels, reward scores, verifier outcomes, critiques, and alignment traces must bind to upstream L107 private-reasoning and L91 model-lineage hashes, carry source labels and royalties forward, disclose synthetic signal inputs, include annotator or verifier attestations, and avoid publishing raw feedback, prompt, answer, source, or reward-rationale text. This directly implements the post-training lineage lessons from RLVR Datasets and Where to Find Them, Tracing the Roots, and Provable Model Provenance Set: model behavior shaped by post-training signals is still a provenance-bearing use, not an attribution-free side channel. - The attribution bill of materials layer now emits
rdllm-attribution-bom/v1artifacts after L108 signal provenance. It closes the model-release supply-chain gap: source components, notice hashes, license-term hashes, proof artifact hashes, provider-card evidence, proof-dependency graph evidence, and post-training signal provenance must travel with the model or API release in a CycloneDX-aligned AI/ML BOM. This follows current AI bill of materials and ML-BOM standardization work and addresses permissive-washing findings that license notices and terms are easily lost unless the release artifact makes notice carry-forward replayable: https://cyclonedx.org/capabilities/mlbom/ https://arxiv.org/abs/2605.19755 https://arxiv.org/abs/2602.08816 - The creator attribution audit index layer now emits
rdllm-creator-attribution-audit-index/v1artifacts after L109 ABOM publication. It closes the creator self-audit gap: a creator query must replay across model-release source components, grounded source footers, delivery receipts, post-training signals, source-access leases, license transactions, model lineage, training summaries, and payout receipts. This maps transparency policy pressure and content-provenance practice into a concrete verifier: a provider cannot merely publish aggregate transparency; it must answer a creator-specific query with namespaced source identities and private-text redaction. https://crfm.stanford.edu/fmti/ https://c2pa.org/specifications/specifications/2.1/index.html https://leginfo.legislature.ca.gov/faces/billNavClient.xhtml?bill_id=202320240AB2013 - The creator attribution audit federation layer now emits
rdllm-creator-attribution-audit-federation/v1artifacts after L110 provider indexes. It closes the cross-provider self-audit gap: the same creator query must replay across multiple provider-local indexes under one query hash, with provider-scoped source-label namespaces, bound exchange or federation artifacts, and explicit conflict rows for unresolved creator identity mismatches. This combines recent source-attribution RAG work with SCITT-style transparent statements, W3C proof-carrying credentials, and OpenID Federation-style trust chaining so source footers can become portable confidence evidence rather than plausible citation strings. https://arxiv.org/abs/2507.04480 https://www.rfc-editor.org/rfc/rfc9943.html https://www.w3.org/TR/vc-data-model/ https://openid.net/specs/openid-federation-1_0.html - The creator audit federation transparency layer now emits
rdllm-creator-attribution-audit-federation-transparency/v1artifacts after L111. It closes the equivocation gap: a provider cannot give one creator, customer, or regulator a clean federation answer and another party a different answer for the same query/provider set without creating a public split-view or same-query conflict row. This imports Certificate Transparency, SCITT, and transparency-witness practice into creator attribution rather than leaving federation answers as private API responses. https://www.rfc-editor.org/rfc/rfc9162.html https://www.rfc-editor.org/rfc/rfc9943.html https://github.com/transparency-dev/witness - The creator audit transparency monitor now emits
rdllm-creator-audit-transparency-monitor/v1artifacts after L112. This is the creator-side counterpart to transparency publication: a creator or auditor can scan append-only logs for a query commitment, prove inclusion for every matching federation or participant-index entry, report newly observed appearances since a prior monitor run, and fail closed on contradictory same-query answers. The design directly answers recent source-attribution findings that citations are confidence infrastructure only when they can be checked externally, not merely displayed in a footer. https://arxiv.org/abs/2507.04480 https://arxiv.org/abs/2508.00838 https://www.rfc-editor.org/rfc/rfc9162.html https://www.rfc-editor.org/rfc/rfc9943.html - The creator audit private watch layer now emits
rdllm-creator-audit-private-watch/v1artifacts after L113. It closes the privacy gap created by public monitor query hashes: a creator can prove that a monitor found appearances and new observations through keyed watch tokens without exposing stable query hashes, provider-set hashes, subject hashes, provider IDs, prompt text, answer text, source text, license text, or payment data. The reference prototype uses deterministic HMAC watch tokens for offline replay; the production upgrade path is an RFC 9497 VOPRF/POPRF service or a PSI/PIR monitor backend so providers do not learn creator watchlists. https://www.rfc-editor.org/rfc/rfc9497.html https://www.rfc-editor.org/rfc/rfc9162.html https://arxiv.org/abs/2507.04480 - The deep research citation audit layer now emits
rdllm-deep-research-citation-audit/v1artifacts after L114. It closes the long-form citation gap identified in recent deep-research-agent work: a source string or reachable link is not sufficient evidence that a cited source supports the attached claim. The artifact parses rendered text markers, resolves each marker to materialized source rows, binds cited claims to source and quote hashes, records relevance and factual-support scores, and fails closed on unresolved markers, unmaterialized links, weak support, or private replay text leakage. https://arxiv.org/abs/2605.06635 https://arxiv.org/abs/2507.04480 https://arxiv.org/abs/2604.05467 - The source freshness audit layer now emits
rdllm-source-freshness-audit/v1artifacts after L115. It maps recent temporal-RAG and freshness work into a deployable verifier: claims are typed as static/current/latest/recent/rapidly-changing/as-of; selected sources must have retrieved-at and effective-at metadata, source-version hashes, validity windows, policy-bound source-age limits, and retrieval-lag limits; and candidate rows prove that a fresher supported source was not ignored unless it is materially weaker. This addresses the gap between "the citation exists" and "the citation was temporally valid when the model answered." https://arxiv.org/abs/2603.16544 https://arxiv.org/abs/2310.03214 https://arxiv.org/abs/2603.18012 - The evidence-force calibration layer now emits
rdllm-evidence-force-calibration/v1artifacts after L118. It maps the evidence-force warning in recent cited-RAG work into a deployable gate: a citation is not enough unless the claim's relation, modality, scope, temporal, and numeric strength are no stronger than the source warrants. Verified source footers and direct payout fail closed for over-warranted claims, and every visible verified footer claim must have a matching calibrated force row, even when the source is real, reachable, licensed, fresh, and relevant. - The warranted source footer layer now emits
rdllm-warranted-source-footer/v1artifacts after L119. It turns the private force-calibration replay into a public footer supplement: each visible claim exposes relation, modality, scope, temporal, and numeric warrant labels, claim hashes, and proof handles while withholding raw prompt, claim, evidence, and source text. https://arxiv.org/abs/2605.28044 https://arxiv.org/abs/2605.06635 https://arxiv.org/abs/2507.04480 - The source-origin lineage layer now emits
rdllm-source-origin-lineage/v1artifacts after L120. It closes the synthetic source laundering gap: a footer source is not directly payable merely because it is visible, relevant, reachable, and warranted. The report classifies each source row as human-original, synthetic-with-lineage, synthetic-unattributed, or unknown; synthetic-with-lineage rows split value to active upstream creators; unknown or unattributed synthetic rows route to origin-review escrow; and the user-facing footer can expose origin labels without revealing raw source text. https://arxiv.org/abs/2605.23684 https://arxiv.org/abs/2605.28044 https://arxiv.org/abs/2605.06635 https://arxiv.org/abs/2507.04480 - The evidence-locator layer now emits
rdllm-evidence-locator-manifest/v1artifacts after L122. It closes the remaining reader-facing gap between "the footer showed a source preview" and "the user can inspect the exact passage": every preview row must carry a public locator URL, resolver status, source-URI match, preview-row hash, excerpt hash, and snapshot or text-fragment proof. Missing, unresolved, non-exact, source-URI mismatched, or private-text-leaking locators fail closed. https://arxiv.org/abs/2605.27700 https://arxiv.org/abs/2605.23684 https://arxiv.org/abs/2605.28044 - The citation URL-health layer now emits
rdllm-citation-url-health/v1artifacts after L123. It closes the provider-neutral footer gap between "the locator points somewhere exact" and "the URL is not hallucinated": every locator must be live, content-addressed, DOI-resolved, or backed by archival snapshot evidence, with fabricated, never-seen, mismatched, unresolved, or unverified URLs failing closed before footer display or direct attribution. This is designed for composite adoption across foundation-model providers because it depends on replayable resolver receipts rather than a provider's internal model architecture. https://arxiv.org/abs/2605.27700 https://arxiv.org/abs/2605.23684 - The composite foundation adapter layer now emits
rdllm-composite-foundation-adapter/v1artifacts after L124. It closes the adoption gap between "RDLLM has a provider-neutral proof object" and "a real foundation-model API can carry it": native OpenAI Responses, Anthropic Messages, Google Gemini, Meta/Llama-style, Mistral, Cohere, xAI, Amazon Bedrock Converse, Azure OpenAI Responses, and OpenAI-compatible response rows must map their response IDs, model IDs, output hashes, attribution headers, embedded JSON fields, citation paths, tool paths, and final streaming hash events to the same RDLLM envelope, footer-delivery, and URL-health hashes. Missing provider-family coverage, native-output hash mismatch, missing proof headers, missing JSON proof fields, missing citation/tool paths, streaming drift, non-fail-closed policy, and private text disclosure fail closed. https://platform.openai.com/docs/api-reference/responses https://docs.anthropic.com/en/api/messages https://ai.google.dev/gemini-api/docs https://docs.mistral.ai/api/ https://llama-stack.readthedocs.io/ - The foundation provider conformance layer now emits
rdllm-foundation-provider-conformance/v1artifacts after L125. It converts "the adapter can map one native response" into "the provider family has public fixtures for all attribution-critical modes": sync responses, streaming, tool calling, citation or grounding metadata, URL-health binding, structured proof fields, claim-support footers, parametric-memory fallback, and fail-closed negative cases. This directly addresses recent citation-audit findings that plausible references can still fail source support, freshness, or origin checks. https://docs.cohere.com/reference/chat https://docs.x.ai/docs/api-reference https://docs.aws.amazon.com/bedrock/latest/userguide/conversation-inference.html https://learn.microsoft.com/azure/ai-services/openai/how-to/responses -
The universal foundation model contract layer now emits
rdllm-universal-foundation-model-contract/v1artifacts after L131. It closes the remaining adoption gap between "each component has a proof" and "a foundation-model provider, broker, or enterprise gateway is required to use the same proof chain for every supported provider family": the contract binds provider card, integration profile, discovery manifest, proof dependency graph, composite adapter, provider conformance, runtime adapter, runtime router, deployment attestation, universal composition, universal settlement, selected route, source footer, creator-pool conservation, and published public surfaces. Missing provider-family coverage, route/deployment drift, missing public surfaces, unready settlement, footer-settlement mismatch, or private text leakage fail closed. -
The L133-L136 pass closes the last provider-neutral adoption gap. AEX frames the API boundary as a signed request-output relation, SEAR and GenAI telemetry show that gateway traces can be standardized, DataDignity and SourceTracker show that output-to-source provenance must survive paraphrase, hard decoys, and code reuse, and Cited but Not Verified shows that visible citations can still fail factual support. C2PA's Content Credentials and durable credential model provide an interoperable content-provenance surface, while recent C2PA security and provenance/watermarking papers show why metadata alone is insufficient for high-stakes trust. RDLLM-L136 therefore emits a
rdllm-universal-content-credential/v1artifact that binds visible source rows, evidence previews, exact locators, URL-health rows, payout eligibility rows, output content-credential rows, durable watermark/fingerprint signals, public verifier surfaces, and the L135 universal invocation witness. Missing footer rows, unverifiable locators, missing payout rows, missing durable provenance signals, weak provider-call non-repudiation, or private text disclosure fail closed. https://arxiv.org/abs/2603.14283 https://arxiv.org/abs/2603.26728 https://arxiv.org/abs/2605.05687 https://arxiv.org/abs/2605.28510 https://arxiv.org/abs/2605.06635 https://spec.c2pa.org/specifications/specifications/2.2/specs/C2PA_Specification https://arxiv.org/abs/2604.24890 https://arxiv.org/abs/2605.21002 -
The L137 pass turns the component stack into one provider-neutral deployment passport. Recent evidence-based generation surveys show that attribution, citation, quotation, and evaluation terminology remain fragmented; source attribution in RAG shows that retrieved-document contribution can be measured but is expensive and approximation-sensitive; CiteGuard shows that faithful citation attribution benefits from retrieval-augmented validation rather than LLM-as-judge alone; DataDignity shows that training-data provenance needs hard anti-document controls instead of lexical similarity; and large-scale 2026 citation audits show fabricated references are now a real production risk. RDLLM-L137 therefore emits a
rdllm-universal-rdllm-passport/v1artifact that binds the L136 content credential to certification, certification attestation, provider card, training summary, integration profile, discovery manifest, assurance bundle, replay DAG, composite foundation adapter, provider-family conformance, runtime adapter/router, deployment attestation, universal composition, universal settlement, foundation-model contract, invocation guard, invocation coverage, invocation witness, public well-known surfaces, verifier commands, and explicit research-control rows. Missing provider-family coverage, missing public passport discovery, missing verifier commands, unready source footers/content credentials, unverifiable artifact hashes, weak research controls, or private prompt/output/source/payment leakage fail closed. https://arxiv.org/abs/2508.15396 https://arxiv.org/abs/2507.04480 https://arxiv.org/abs/2510.17853 https://arxiv.org/abs/2605.05687 https://arxiv.org/abs/2605.07723 https://www.w3.org/TR/vc-data-integrity/ https://opentelemetry.io/docs/specs/semconv/gen-ai/ https://modelcontextprotocol.io/specification/2025-06-18 -
The L138 pass turns the deployment passport into a universal adoption standard that a foundation-model provider, gateway, SDK, auditor, enterprise buyer, creator registry, or clearinghouse can implement without bespoke negotiation. Recent citation-attribution audits show that source-looking footers are not enough: citations must resolve to actual cited content and be checked against claim support, while large-scale hallucinated-citation measurements show that fabricated references are now observable at ecosystem scale. Official C2PA, W3C Data Integrity, OpenTelemetry GenAI, MCP, and SCITT work provide reusable surfaces for content credentials, signed proof objects, provider-call telemetry, tool/resource discovery, and transparency publication, but none of them by itself defines the royalty and attribution procurement gate. RDLLM-L138 therefore emits
rdllm-universal-adoption-standard/v1: a signed implementer contract binding the L137 passport, conformance vectors, integration profile, well-known discovery, provider attribution card, certification report, attribution exchange, federation handshake, trust registry, SDK surfaces, role obligations, procurement gates, standards mappings, verifier commands, public paths, and private-text leakage checks. Missing provider families, missing public standard surfaces, missing verifier commands, weak conformance vectors, failed procurement gates, or private prompt/output/source/payment leakage fail closed. https://arxiv.org/abs/2605.06635 https://arxiv.org/abs/2605.07723 https://spec.c2pa.org/specifications/specifications/2.2/specs/C2PA_Specification https://www.w3.org/TR/vc-data-integrity/ https://opentelemetry.io/docs/specs/semconv/gen-ai/ https://modelcontextprotocol.io/specification/2025-06-18 https://datatracker.ietf.org/doc/draft-ietf-scitt-architecture -
The L139 pass turns the adoption standard into an executable interop test kit. Recent provider-neutral telemetry, source-attribution, provenance, and citation-verification work all point to the same operational risk: a provider can publish the right policy words while SDKs, gateways, renderers, CI jobs, or offline auditors accept responses with missing footers, missing proof objects, stale provider routes, invalid signatures, rights-denied sources, hallucinated citations, unguarded native calls, or private-text leaks. RDLLM-L139 therefore emits
rdllm-universal-interop-test-kit/v1: a signed compatibility package binding the L138 standard to provider-family golden fixtures, SDK bindings, execution targets, negative mutation cases, public verifier commands, source footer delivery, response envelopes, invocation guards, and trust roots. A deployment cannot claim practical universal compatibility unless those fixtures replay and those negative cases block. https://opentelemetry.io/docs/specs/semconv/gen-ai/ https://arxiv.org/abs/2605.06635 https://arxiv.org/abs/2605.27700 https://arxiv.org/abs/2605.05687 https://www.w3.org/TR/vc-data-integrity/ -
The L140 pass turns executable compatibility into runtime source provenance. Recent attributed-generation work shows that citations need generation-time planning and progress checks; source-attribution RAG shows that retrieved document contribution can be measured but must be made cheaper and replayable; multimodal attribution benchmarks show that evidence may be text, image, or region-level context; and large citation audits show that fabricated references are now an ecosystem-scale failure mode. At the same time, MCP makes external tools and data sources a first-class model context surface, while C2PA, W3C Data Integrity, PROV, and the EU GPAI transparency/copyright regime provide reusable credential and disclosure patterns. RDLLM-L140 therefore emits
rdllm-universal-context-provenance-bridge/v1: a signed runtime bridge that requires every MCP/tool/retrieval/file/browser/vector/enterprise/license/media context event to be authorized, licensed, converted into source claims, projected into visible source footers, bound to agent steps, and reconciled to royalty projection before a context-influenced response can be released. https://aclanthology.org/2025.acl-long.490/ https://arxiv.org/abs/2507.04480 https://ojs.aaai.org/index.php/AAAI/article/view/40585 https://arxiv.org/abs/2605.07723 https://modelcontextprotocol.io/specification/2025-06-18 https://spec.c2pa.org/specifications/specifications/2.4/specs/C2PA_Specification.html https://www.w3.org/TR/vc-data-integrity/ https://digital-strategy.ec.europa.eu/en/policies/contents-code-gpai -
The L141 pass turns runtime provenance into user-verifiable citation grounding. Recent citation audits and RAG verification work show that a visible source row is not enough: the source must exist, the locator must resolve or have durable fallback evidence, the cited passage must support the claim, and confidence must be calibrated rather than implied by formatting. RDLLM-L141 therefore emits
rdllm-universal-citation-verification-contract/v1, a signed contract that requires every displayed label to pass source identity, URL health, evidence locator, claim support, evidence-force, confidence, rendered-footer, L140 context, authenticity, and royalty checks before the UI can call a citation verified grounding. https://arxiv.org/abs/2605.07723 https://aclanthology.org/2025.acl-long.490/ https://arxiv.org/abs/2507.04480 https://arxiv.org/abs/2605.28044 https://arxiv.org/abs/2605.23684 -
The L142 pass closes the production cache/reuse bypass. Recent semantic-cache work shows that production LLM deployments use static/dynamic caches to reduce latency and cost, but similarity-threshold reuse creates correctness and collision risks. Recent RAG attribution work also shows that evidence utility and source attribution must be measured at the retrieved-item level rather than inferred from final-answer quality. RDLLM-L142 therefore requires every cached or replayed answer to revalidate query equivalence, evidence overlap, source-version freshness, consent/license continuity, L141 citation continuity, cache-collision resistance, and a new royalty-metered reuse event before the cached answer can be served as grounded. https://machinelearning.apple.com/research/semantic-caching https://www.nature.com/articles/s41598-026-36721-w https://arxiv.org/abs/2601.23088 https://arxiv.org/abs/2604.05467 https://arxiv.org/abs/2508.15396
-
The L143 pass closes the training-to-serving attribution bypass. Recent source attribution and citation-verification work shows that response grounding must expose which retrieved sources support the answer, while training-data, post-training, distillation, and synthetic-data lineage work shows that some obligations originate before retrieval-time context ever appears. RDLLM-L143 therefore emits
rdllm-universal-training-serving-contract/v1: a signed provider-neutral matrix binding pretraining, fine-tuning, adapters, RLHF/RLAIF/RLVR, distillation, synthetic-data ingestion, release snapshots, runtime routes, grounded reuse, citation footers, revocation propagation, residual royalties, valuation methods, and serving meters before a foundation-model answer can claim that source attribution and royalty obligations survived from training into serving. https://arxiv.org/abs/2507.04480 https://arxiv.org/abs/2508.15396 https://arxiv.org/abs/2605.05687 https://arxiv.org/abs/2604.05467 -
The L144 pass closes the private-evidence trust gap. Public source footers make answers inspectable, but foundation-model providers still need a way to prove training corpus membership, license/consent state, post-training signal lineage, distillation lineage, serving metering, citation replay, residual valuation, and revocation/unlearning without publishing proprietary corpora, reward data, prompts, customer logs, model weights, hidden states, or source text. RDLLM-L144 therefore emits
rdllm-universal-confidential-attribution-audit/v1: a signed provider-neutral matrix over 10 provider families and 10 audit domains, backed by confidential evidence-room commitments, ZK/TEE/selective-disclosure proof handles, encrypted evidence escrow, independent auditor quorum, creator challenge routes, regulator export support, differential privacy budget commitments, and fail-closed leakage fixtures. https://arxiv.org/abs/2507.04480 https://arxiv.org/abs/2508.15396 https://arxiv.org/abs/2605.05687 https://arxiv.org/abs/2604.05467 https://arxiv.org/abs/2605.14421 -
The L145 pass closes the runtime-authority gap. Attribution and payout proofs can still fail in production if an agent, tool, memory store, retrieval connector, browser action, model route, or settlement gateway acts without a signed actor/intent/context authority chain. Recent work makes this gap explicit: Sovereign Context Protocol proposes attribution-aware access to creator-owned data; HDP signs multi-hop human delegation; PROV-AGENT binds agent interactions into workflow provenance; MCP production analysis identifies missing identity propagation and observability; and explicit-provenance work argues that agentic responsibility must be computable and interventionable. RDLLM-L145 therefore requires a 10-provider by 10-runtime authority matrix, public verifier commands, intervention logging, revocation checks, and fail-closed negative fixtures before an attributed answer, source footer, tool action, memory-influenced response, or settlement can be released. https://arxiv.org/abs/2603.27094 https://arxiv.org/abs/2604.04522 https://arxiv.org/abs/2508.02866 https://arxiv.org/abs/2603.13417 https://arxiv.org/abs/2605.17169
-
The L146 pass closes the adoption-root gap. Payouts alone do not make users confident that an answer is grounded, and isolated source footers do not prove that the provider, runtime, training-to-serving continuity, confidential audit, public discovery, assurance bundle, and settlement authority all agree. RDLLM-L146 therefore emits
rdllm-universal-rdllm-root/v1: one signed, provider-neutral root that binds certification, attestation, provider card, integration profile, discovery manifest, assurance bundle, proof graph, L143 training-to-serving continuity, L144 confidential auditability, L145 runtime authority, source-footer delivery, provider-family coverage, negative fixtures, and settlement authorization. This maps recent explicit-provenance work, CUE-R evidence utility, CiteGuard citation-attribution alignment, AIBOM lifecycle assurance, SCITT/RFC 9943 transparency statements, W3C Data Integrity, C2PA content credentials, and recent creator data-sharing/compensation market design into one verifiable deployment object. https://arxiv.org/abs/2605.17169 https://arxiv.org/abs/2604.05467 https://arxiv.org/abs/2510.17853 https://www.rfc-editor.org/rfc/rfc9943.html https://www.w3.org/TR/vc-data-integrity/ https://spec.c2pa.org/specifications/specifications/2.2/specs/C2PA_Specification.html https://ssrn.com/abstract=6761318 -
The L147 pass closes the runtime-emission gap. A deployment root can tell a buyer which provider proof pack is trustworthy, but it does not by itself prove that a concrete answer actually passed through the root-bound release path. RDLLM-L147 therefore emits
rdllm-universal-emission-enforcement-gateway/v1: one signed per-response artifact that binds the L146 root to the release gate, proof-carrying response, serving gateway, response envelope, delivered source footer, emission evidence, live witness, transparency log, invocation guard, invocation coverage, invocation witness, foundation runtime route, deployment attestation, trust registry, and client display enforcement. This maps recent work on deep research citation parsing and evaluation, argument-level provenance for agent enforcement, runtime protection controllers, verifiable training/release attestations, Warrant Certificate Authorities, runtime governance receipts, and dual attribution/verification into a fail-closed answer-emission mechanism. https://arxiv.org/abs/2605.06635 https://arxiv.org/abs/2605.11039 https://arxiv.org/abs/2604.17562 https://arxiv.org/abs/2603.28988 https://www.ietf.org/archive/id/draft-bondar-wca-00.html https://attestedintelligence.com/documents/verifiable-runtime-governance.pdf https://arxiv.org/abs/2604.21193 https://opentelemetry.io/docs/specs/semconv/gen-ai/ -
The L148 pass closes the provider-adoption completeness gap. A root and per-response gateway are still not enough for a market standard unless buyers, creators, auditors, and SDK implementers can verify that every provider family, native API route, source footer, client display surface, telemetry mapping, settlement route, and public proof receipt follows the same contract. RDLLM-L148 therefore emits
rdllm-universal-composite-rdllm-profile/v1: one signed adopter-facing profile that binds the universal passport, adoption standard, interop test kit, L146 root, L147 gateway, source-footer delivery, client attribution enforcement, foundation API/runtime adapters, invocation guard, invocation witness, trust registry, provider-family rows, API binding rows, expected API-to-provider-family mappings, standards mappings, and fail-closed fixtures. It explicitly covers OpenAI, Anthropic, Google, Meta, Mistral, Cohere, xAI, DeepSeek, AWS Bedrock, Azure OpenAI, OpenRouter-style aggregators, local open-weight deployments, and enterprise gateways. It maps the 2026 citation reliability result that citations must be checked for link validity, relevance, and factual support; RAG source-attribution work on document utility; mechanistic training-origin attribution; SCP-style creator access logs; SourceTracker-style code provenance; LLM-Rosetta's hub-and-spoke cross-provider API translation; OpenTelemetry GenAI; C2PA 2.4; SCITT/RFC 9943; W3C Data Integrity; WCA; MCP; ISO 20022; and RSL/TDM rights signals into one deployable contract. https://arxiv.org/abs/2605.06635 https://arxiv.org/abs/2507.04480 https://arxiv.org/abs/2601.21996 https://arxiv.org/abs/2603.27094 https://arxiv.org/abs/2605.28510 https://arxiv.org/abs/2604.09360 https://opentelemetry.io/docs/specs/semconv/gen-ai/ https://spec.c2pa.org/specifications/specifications/2.4/specs/C2PA_Specification.html https://www.rfc-editor.org/rfc/rfc9943.html https://www.w3.org/TR/vc-data-integrity/ https://www.ietf.org/archive/id/draft-bondar-wca-00.html https://docs.aws.amazon.com/bedrock/latest/userguide/conversation-inference.html https://learn.microsoft.com/en-us/azure/foundry/openai/how-to/responses -
The L149 pass closes the deployment-runtime conformance gap. A universal provider profile is not sufficient unless live routes can prove that each answer was generated through guarded provider APIs, source-attribution modes, visible footer injection, client display enforcement, OpenTelemetry GenAI export, proof download, challenge routing, privacy filtering, and settlement metering. RDLLM-L149 therefore emits
rdllm-universal-runtime-conformance-receipt/v1: one signed receipt binding sync generation, streaming generation, tool calls, agent actions, RAG answers, memory-influenced answers, batch generation, enterprise proxying, and client rendering to the L148 profile, L147 gateway, L146 root, citation identity, claim-source attribution, evidence-region binding, deep-research citation audits, source leases, invocation witnesses, finance ledgers, and trust registry. This maps DataDignity-style pinpoint provenance, CUE-R evidence-item utility, PaperTrail claim-evidence interfaces, source-attribution RAG findings, hallucinated-citation audits, mechanistic/influence attribution, LLM-Rosetta cross-provider API translation, OpenTelemetry GenAI, C2PA, MCP resources/tools, OpenAI Responses, Azure OpenAI Responses, AWS Bedrock Converse, and OpenRouter-style API routes into one runtime rule: no L149 receipt, no verified footer display and no creator settlement claim. https://arxiv.org/abs/2605.05687 https://arxiv.org/abs/2604.05467 https://arxiv.org/abs/2602.21045 https://arxiv.org/abs/2507.04480 https://arxiv.org/abs/2605.07723 https://arxiv.org/abs/2601.21996 https://arxiv.org/abs/2601.21929 https://arxiv.org/abs/2604.09360 https://opentelemetry.io/docs/specs/semconv/gen-ai/ https://c2pa.wiki/specifications/ https://modelcontextprotocol.io/specification/2025-06-18/basic/index https://platform.openai.com/docs/api-reference/responses https://learn.microsoft.com/en-us/azure/foundry/openai/how-to/responses https://aws.amazon.com/about-aws/whats-new/2026/02/amazon-bedrock-batch-inference-supports-converse-api-format/ https://openrouter.ai/docs/api-reference/chat-completion https://openrouter.ai/docs/api-reference/chat-completion https://modelcontextprotocol.io/ https://rslstandard.org/ -
The L150 pass closes the post-hoc citation gap. A runtime receipt proves that a route was guarded, but it does not by itself prove that each generated claim was emitted with source support rather than decorated with plausible citations after the answer existed. RDLLM-L150 therefore emits
rdllm-universal-claim-provenance-envelope/v1: one signed envelope that binds every displayed claim or sentence to a claim hash, response segment hash, generation-time provenance record, support relation, source proof, exact evidence region, citation identity, locator health, visible footer row, tool or memory trace, payout basis, and settlement meter. It covers quotation, compression, inference, retrieval, tool observation, conversation memory, parametric memory, and residual corpus value, and it requires API, Markdown, HTML, streaming, mobile, and export surfaces to preserve the verified footer and downloadable proof. This maps TRACER-style verifiable generative provenance, AgentSim verifiable trace simulation, PaperTrail claim-evidence interfaces, DataDignity pinpoint provenance, RAG source-attribution findings, and hallucinated-citation audits into one user-facing rule: no L150 envelope, no grounded display and no direct creator settlement. https://arxiv.org/abs/2605.09934 https://arxiv.org/abs/2604.26653 https://arxiv.org/abs/2602.21045 https://arxiv.org/abs/2605.05687 https://arxiv.org/abs/2507.04480 https://arxiv.org/abs/2605.07723 -
The L151 pass closes the provider-wire gap. AEX shows that LLM API users need request-output attestation, transform receipts, and streaming checkpoints at the API boundary, while SCP argues that creator-owned content access must be logged, licensed, and attributable by default. RDLLM-L151 turns those findings into one universal provider rule: OpenAI-style Responses routes, Anthropic Messages routes, Gemini generate-content routes, Bedrock Converse routes, OpenRouter/aggregator routes, local OpenAI-compatible routes, and enterprise proxies must preserve L150 envelope hashes, L149 runtime receipt hashes, footer metadata, telemetry spans, transform receipts, and settlement meters across headers, JSON bodies, streams, SDK metadata, batch callbacks, webhooks, and exported copies. If a provider, proxy, SDK, or export path strips the proof, the route cannot claim grounded display or direct creator settlement. https://arxiv.org/abs/2603.14283 https://arxiv.org/abs/2603.27094 https://opentelemetry.io/docs/specs/semconv/gen-ai/ https://platform.openai.com/docs/api-reference/responses https://platform.claude.com/docs/en/api/messages https://ai.google.dev/gemini-api/docs/text-generation https://docs.aws.amazon.com/bedrock/latest/userguide/conversation-inference.html https://openrouter.ai/docs/api-reference/chat-completion
-
The L152 pass closes the accountability-chain gap. Recent audit-trail work defines LLM audit records as chronological, tamper-evident lifecycle ledgers that connect technical provenance with approvals, waivers, and attestations. AttriGuard shows why agent tool invocations need causal attribution rather than raw transcripts. PROV-AGENT and the IETF SPICE inference-chain draft show how agent workflows and inference events can be exported as verifiable provenance chains. LDP delegation contracts and HDP human delegation provenance show that multi-agent authority needs explicit scopes, attested identity, and append-only delegation history. RDLLM-L152 turns those findings into one settlement rule: no append-only audit trail binding provider-wire calls, policy versions, actor authority, tool/memory actions, exports, challenges, and settlement meters, no grounded display or direct creator settlement. https://arxiv.org/abs/2601.20727 https://arxiv.org/abs/2603.10749 https://arxiv.org/abs/2508.02866 https://datatracker.ietf.org/doc/html/draft-mw-spice-inference-chain-00 https://arxiv.org/abs/2603.18043 https://arxiv.org/abs/2604.04522
-
The L153 pass closes the split-view accountability gap. Certificate Transparency/RFC 9162 defines signed tree heads, inclusion proofs, and consistency proofs; SCITT/RFC 9943 applies transparency receipts to signed supply-chain statements; Sigstore/Rekor shows practical artifact-signature transparency; C2SP checkpoint and witness work gives portable checkpoint and cosignature formats; transparency-dev's verifiable data-structure work and CoSi-style witness cosigning show how independent observers can prevent equivocation. RDLLM-L153 turns those mechanisms into an attribution rule: the L152 audit-trail checkpoint must be public-log included, consistency-proven, continuously monitored, and independently witnessed before users, creators, customers, auditors, or regulators can rely on source attribution or direct settlement. https://www.rfc-editor.org/rfc/rfc9162.html https://www.rfc-editor.org/rfc/rfc9943.html https://docs.sigstore.dev/logging/overview/ https://github.com/C2SP/C2SP/ https://transparency.dev/verifiable-data-structures/ https://arxiv.org/abs/1503.08768
-
The L154 pass closes the grounded-reliance gap. Recent attribution and citation work shows that a model can display a plausible source while the cited page does not actually support the generated claim; provenance metadata alone is not the same as user-safe reliance. DataDignity frames output attribution as a ranked source-support problem, Cited-But-Not-Verified and RefLens emphasize evidence-grounded citation verification, and ecosystem analyses of LLM search attribution show the economic cost of uncredited or weakly credited sources. C2PA content credentials, SCITT transparency receipts, NIST AI RMF traceability, W3C Data Integrity, and EU AI Act Article 50-style transparency support the publication layer. RDLLM-L154 turns those pieces into a verifier rule: no verified source footer, support coverage, locator health, freshness, warrant calibration, client rendering, L153 witness root, and finance reconciliation, no user reliance or direct creator payout. https://arxiv.org/abs/2605.05687 https://arxiv.org/abs/2605.06635 https://doi.org/10.1609/aaai.v40i48.42361 https://doi.org/10.1017/dap.2026.10064 https://spec.c2pa.org/specifications/specifications/2.2/specs/C2PA_Specification.html https://www.rfc-editor.org/rfc/rfc9943.html https://www.nist.gov/itl/ai-risk-management-framework https://www.w3.org/TR/vc-data-integrity/ https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-50
-
The L155 pass closes the post-release correction gap. April and May 2026 citation studies show that citation URLs can be fabricated, stale, or non-resolving at scale, and that tool-assisted checking can reduce those failures only when correction is part of the serving workflow. FineRef shows that citation mismatch and irrelevance need fine-grained reflection and correction, while the W3C status-list, C2PA update-manifest/revocation, SCITT receipt, RFC 9162, and Rekor patterns show how public status changes can be published as verifiable records. RDLLM-L155 turns those findings into a post-release verifier rule: no live status row, correction broadcast, stale citation revalidation, copied-output status link, cache invalidation, creator notice, regulator export notice, and settlement hold or adjustment, no continuing reliance on the answer footer or payout claim. https://arxiv.org/abs/2604.03173 https://arxiv.org/abs/2605.07723 https://ojs.aaai.org/index.php/AAAI/article/view/40547 https://www.w3.org/TR/vc-bitstring-status-list/ https://spec.c2pa.org/specifications/specifications/2.4/specs/C2PA_Specification.html https://www.rfc-editor.org/rfc/rfc9943.html
-
The L156 pass closes the cross-provider adoption gap. Recent citation-failure and provider-API work shows that source attribution must survive not only generation, but also transport through provider-specific response shapes, streaming events, SDK metadata, tool calls, gateway transforms, copied outputs, and client rendering. RDLLM-L156 turns that into a verifier rule: no provider-family adapter, response binding, text-attribution guarantee, client gate, source-status resolver, standard mapping, and negative fixture, no trusted source footer or direct creator settlement. https://arxiv.org/abs/2605.06635 https://arxiv.org/abs/2604.03173 https://modelcontextprotocol.io/specification/2025-11-25 https://opentelemetry.io/docs/specs/semconv/gen-ai/ https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.600-1.pdf
-
The L157 pass closes the provider-native fixture gap. L156 says every provider route must expose the same attribution contract; L157 requires replay evidence that real provider-shaped outputs normalize into it across sync text, streaming text, tool calls, retrieval context, batch callbacks, webhooks, and copied-output exports. This directly reflects recent findings that citation labels can look plausible while unsupported, and that API translation layers, SDKs, gateways, streams, and copied outputs can drop provenance unless the normalized response fields are verified end to end. https://arxiv.org/abs/2605.05687 https://arxiv.org/abs/2605.07723 https://platform.openai.com/docs/api-reference/responses https://docs.anthropic.com/en/api/messages https://ai.google.dev/gemini-api/docs/text-generation https://docs.aws.amazon.com/bedrock/latest/userguide/conversation-inference.html https://openrouter.ai/docs/api-reference/chat-completion https://cyclonedx.org/capabilities/mlbom/ https://www.rfc-editor.org/rfc/rfc9162.html https://docs.sigstore.dev/logging/overview/
-
The L158 pass closes the provider-drift gap. Recent source-attribution failures are not static: citation validity, source freshness, API response schemas, SDK event fields, model aliases, tool metadata, and streaming footer behavior can change after certification. RDLLM-L158 therefore requires pre-release, hourly, daily, weekly negative-rotation, and incident-triggered canaries over provider schemas, response shapes, citation locators, source-footer rendering, telemetry semantics, and settlement meters. A stale route loses grounded display and direct creator settlement until remediation is published. https://arxiv.org/abs/2605.06635 https://arxiv.org/abs/2605.07723 https://opentelemetry.io/docs/specs/semconv/gen-ai/ https://modelcontextprotocol.io/specification/2025-06-18/basic/index
-
The L159 pass closes the request-time negotiation gap. Provider compatibility cannot be inferred from a public profile alone because each call may select a different model route, source locator, citation format, privacy policy, copy policy, or settlement meter. RDLLM-L159 requires the client and provider route to negotiate the attribution contract before invocation; unnegotiated or stale requests fail before generation instead of becoming post-hoc citation claims. https://modelcontextprotocol.io/specification/2025-06-18/basic/index https://opentelemetry.io/docs/specs/semconv/gen-ai/ https://arxiv.org/abs/2603.27094
-
The L160 pass closes the runtime-bypass gap. Negotiation is useful only if the selected contract is carried by the actual API call, stream, SDK retry, gateway proxy, agent step, MCP tool call, retrieval context, batch callback, fallback route, or semantic-cache reuse. RDLLM-L160 makes every invocation emit privacy-preserving receipts, telemetry span hashes, response-envelope hashes, source-footer hashes, claim-provenance hashes, bypass guards, and settlement meters before output display or payout release. https://arxiv.org/abs/2603.14283 https://opentelemetry.io/docs/specs/semconv/gen-ai/ https://modelcontextprotocol.io/specification/2025-06-18/basic/index
-
The L161 pass closes the untrusted-conformance gap. A provider-hosted certification page is not enough for source attribution because relying parties must know who certified the proof pack, whether the certifier was accredited, whether credentials or trust marks are revoked, and whether the statement was logged. RDLLM-L161 binds conformance claims to trust anchors, accredited certifiers, conformance labs, W3C Verifiable Credentials, C2PA conformance practice, SCITT-style transparency statements, revocation status, discovery, and relying-party policy. https://www.w3.org/standards/history/vc-data-model-2.0/ https://spec.c2pa.org/specifications/specifications/2.2/specs/C2PA_Specification.html https://queue.rfc-editor.org/final-review/rfc9943/
-
The L162 pass closes the provider-adoption gap. Trust federation still leaves implementers asking what to ship. RDLLM-L162 packages the federated trust chain with provider-family adapters, standard exports, SDK/gateway contracts, OpenTelemetry GenAI mappings, MCP tool/resource contracts, C2PA assertions, W3C Verifiable Credential trust marks, SCITT transparency statements, creator-audit APIs, regulator exports, and guarded settlement release. No adoption pack means no grounded answer claim or direct creator settlement. https://opentelemetry.io/docs/specs/semconv/gen-ai/ https://modelcontextprotocol.io/specification/2025-06-18/basic/index https://spec.c2pa.org/specifications/specifications/2.2/specs/C2PA_Specification.html https://www.w3.org/standards/history/vc-data-model-2.0/ https://queue.rfc-editor.org/final-review/rfc9943/
-
The L163 pass closes the industry-root gap. The latest citation-verification papers show that a footer with valid links can still fail factual support, and standards work shows that provenance claims must be portable, signed, logged, and independently verifiable. RDLLM-L163 therefore binds the L162 adoption pack, current proof dependency graph, public verifier endpoints, adoption-role obligations, creator audit routes, regulator exports, copied-output status links, and fail-closed negative root fixtures into one acyclic public root. No verified root means no source-footer confidence signal, copied-output reliance, or direct creator settlement. https://arxiv.org/abs/2605.06635 https://arxiv.org/abs/2605.07723 https://arxiv.org/abs/2507.04480 https://opentelemetry.io/docs/specs/semconv/gen-ai/ https://spec.c2pa.org/specifications/specifications/2.2/specs/C2PA_Specification.html https://www.w3.org/standards/history/vc-data-model-2.0/ https://queue.rfc-editor.org/final-review/rfc9943/
-
The L164 pass closes the installability gap. Recent citation-attribution work such as VeriCite, CiteGuard, Source Attribution in RAG, and CUE-R shows that source footers need claim verification, retrieval-aware validation, and evidence-utility tests rather than link-only presentation. Recent training-data attribution work such as DATE-LM, Distributional Training Data Attribution, Mechanistic Data Attribution, and the EMNLP 2025 influence-functions critique also shows why RDLLM must keep runtime source proof, training influence, and payout evidence as separate replay channels. RDLLM-L164 therefore packages the L163 root into signed reproducible SDKs, gateway middleware, MCP adapters, OpenTelemetry GenAI mappings, OpenAPI contracts, C2PA assertions, W3C VC trust marks, SCITT statements, offline verifier containers, CI workflows, procurement policies, and SBOM/SLSA provenance. No signed reproducible distribution means no provider installation claim, no source-footer reliance, and no direct creator settlement release. https://arxiv.org/abs/2510.11394 https://arxiv.org/abs/2510.17853 https://arxiv.org/abs/2507.04480 https://arxiv.org/abs/2604.05467 https://arxiv.org/abs/2507.09424 https://openreview.net/forum?id=UBRFn7YKMe https://arxiv.org/abs/2601.21996 https://aclanthology.org/2025.findings-emnlp.775/ https://slsa.dev/spec/latest/ https://in-toto.io/
-
The L165 pass closes the live response-attribution gap. Recent work makes clear that a correct final answer is not enough: attribution can be wrong, suppressed, or merely decorative. Probing for Knowledge Attribution separates context-driven answers from parametric memory, CUE-R measures intervention-style evidence utility, factual-confidence prediction gates unsupported RAG answers, CiteGuard frames citation validation as attribution alignment, FalseCite supplies deceptive-citation negative controls, and Attribution Bias names suppression as a distinct failure mode. RDLLM-L165 therefore requires every response to bind visible footer rows to source identity, claim support, current-turn or memory path, causal utility, factual confidence, settlement participation, and negative fixtures before release or direct creator settlement. https://arxiv.org/abs/2602.22787 https://arxiv.org/abs/2604.05467 https://arxiv.org/abs/2605.05244 https://arxiv.org/abs/2510.17853 https://arxiv.org/abs/2602.11167 https://arxiv.org/abs/2604.05224 https://docs.sigstore.dev/logging/overview/ https://spec.openapis.org/oas/latest.html https://opentelemetry.io/docs/specs/semconv/gen-ai/ https://modelcontextprotocol.io/specification/2025-06-18/basic/index https://spec.c2pa.org/specifications/specifications/2.2/specs/C2PA_Specification.html https://www.w3.org/TR/vc-data-model-2.0/ https://scitt.io/
-
The L166 pass closes the model-release claim gap. Current foundation-model governance and AI supply-chain practice is converging on model-level documentation, copyright/TDM transparency, downstream provider documentation, signed model provenance, software/model supply-chain attestations, content credentials, transparency statements, and runtime telemetry. RDLLM-L166 turns those signals into an executable passport for each named model version: no passport means no RDLLM model claim, provider invocation claim, source-footer reliance, or direct creator settlement for that model release. https://digital-strategy.ec.europa.eu/en/policies/contents-code-gpai https://www.nist.gov/itl/ai-risk-management-framework https://openssf.org/blog/2025/06/05/model-signing-is-here/ https://slsa.dev/spec/latest/ https://in-toto.io/ https://spec.c2pa.org/specifications/specifications/2.2/specs/C2PA_Specification.html https://scitt.io/ https://opentelemetry.io/docs/specs/semconv/gen-ai/
-
The L167 pass closes the fragmented-reliance gap. A provider can no longer claim RDLLM by showing a correct-looking footer, a model passport, or a settlement report in isolation. The composite contract binds OpenAPI-visible routes, MCP or agent-tool calls, OpenTelemetry GenAI spans, C2PA credentials, SCITT transparency statements, W3C credentials, SLSA/in-toto provenance, OpenSSF model-signing posture, EU GPAI transparency, and NIST AI RMF/GeneAI governance into one verifier decision for model claims, invocation, answer release, footer reliance, procurement reliance, and creator settlement. https://spec.openapis.org/oas/latest.html https://modelcontextprotocol.io/ https://opentelemetry.io/docs/specs/semconv/gen-ai/ https://spec.c2pa.org/specifications/specifications/2.2/specs/C2PA_Specification.html https://scitt.io/ https://www.w3.org/TR/vc-data-model-2.0/ https://slsa.dev/spec/latest/ https://in-toto.io/ https://openssf.org/blog/2025/06/05/model-signing-is-here/ https://digital-strategy.ec.europa.eu/en/policies/contents-code-gpai https://www.nist.gov/itl/ai-risk-management-framework
-
The L168 pass closes the named-provider binding gap. AI Transparency Atlas reports that model documentation is fragmented across platforms and provider docs, while Cited but Not Verified and PaperTrail show that visible citations still need source accessibility, relevance, factual support, and claim-evidence checks. RDLLM-L168 therefore requires each named provider family and route type to publish native API, streaming, tool/MCP, retrieval/citation, source-footer, live-proof, telemetry, revocation, copy/export, audit, and settlement mappings to the L167 contract before universal provider adoption can be claimed. https://arxiv.org/abs/2512.12443 https://arxiv.org/abs/2605.06635 https://arxiv.org/abs/2602.21045 https://academic.oup.com/grurint/article/75/3/224/8441974
-
The L169 pass closes the executable-provider-evidence gap. Static route mappings do not prove that OpenAI-style, Anthropic-style, Google, cloud, router, local-runtime, RAG, or MCP/agent paths actually executed the required attribution and citation checks. RDLLM-L169 therefore requires an official signed runner receipt with fixture-suite rows, runner-stage rows, per-provider native transcript hashes, public result hashes, and negative canary rejection before provider onboarding, source-footer reliance, procurement reliance, or creator settlement can be claimed. Recent source attribution work motivates the claim-evidence and citation verification suites; OpenTelemetry GenAI motivates portable trace semantics; OpenAPI, Arazzo, and MCP motivate machine-readable API/tool replay surfaces; and SLSA, in-toto, and SCITT motivate signed runner/test-result attestations and transparency publication. https://arxiv.org/abs/2605.06635 https://arxiv.org/abs/2602.21045 https://opentelemetry.io/docs/specs/semconv/gen-ai/ https://spec.openapis.org/oas/ https://spec.openapis.org/arazzo/latest.html https://modelcontextprotocol.io/ https://slsa.dev/spec/latest/ https://in-toto.io/ https://scitt.io/
-
The L171 pass closes the final-answer/source-footer gap. A live invocation can be admitted while the final answer still carries fabricated citations, unsupported claims, unavailable links, wrong-source support, copied output with stripped footer rows, or settlement rows for hidden payable sources. RDLLM-L171 therefore requires a source-grounded response receipt that binds L170 admission and L165 live attribution to source rows, claim rows, response surfaces, citation metadata checks, copy/export preservation, and settlement rows before user reliance or creator payout. Cited but Not Verified and CiteCheck motivate source retrieval plus citation metadata verification; hallucinated-citation studies motivate fail-closed negative fixtures; C2PA Content Credentials motivate portable provenance claims for exported answers; and DataDignity-style pinpoint provenance motivates ranking and exposing the most likely source documents supporting response knowledge. https://arxiv.org/abs/2605.06635 https://arxiv.org/abs/2605.27700 https://arxiv.org/abs/2603.07287
-
The L172 pass closes the distribution-survival gap. Even when a response is grounded at release time, the footer, source locator, content credential, status resolver, revocation snapshot, reuse meter, or settlement obligation can be dropped by copy/paste, PDF export, screenshot capture, web embeds, API relays, social shares, marketplace exports, or downstream RAG ingestion. The universal distribution reliance passport binds L171 and content-credential evidence to those surfaces and fails closed when the distributed copy cannot resolve current source status or carry settlement obligations forward. This extends the same motivation behind Content Credentials and citation-verifier work from original answer display to downstream reliance. https://spec.c2pa.org/specifications/specifications/2.4/specs/C2PA_Specification.html
-
The L173 pass closes the adversarial-provenance gap. Recent work argues that C2PA-style credentials and watermarking need security analysis, trust-list handling, anti-replay controls, and coordination across independent verification layers; other work shows provenance/watermarking contradictions and RAG poisoning risks when provenance is valid but retrieval context is compromised. RDLLM-L173 therefore treats a single footer, manifest, credential, watermark, or resolver as insufficient. It requires independent provenance signals plus adversarial negative fixtures for manifest substitution, signature replay, footer spoofing, locator phishing, resolver split views, watermark removal, proxy rewrites, downstream RAG poisoning, credential time-shift, creator impersonation, settlement-meter forks, and private payload leaks before high-stakes reliance or creator settlement. https://arxiv.org/abs/2604.24890 https://arxiv.org/abs/2603.02378 https://arxiv.org/abs/2604.00387 https://c2pa.org/specifications/specifications/2.0/security/Security_Considerations.html https://arxiv.org/abs/2605.07723 https://spec.c2pa.org/specifications/specifications/2.4/specs/C2PA_Specification.html https://www.microsoft.com/en-us/research/publication/datadignity-training-data-attribution-for-large-language-models/
-
The L174 pass closes the procurement and regulatory reliance gap. Recent provenance work pushes beyond decorative citations toward claim-evidence interfaces; recent training-data attribution work treats source support for generated output as a measurable ranking problem; recent legal/evidentiary provenance work maps content provenance, watermarking, and attestations to court-facing proof requirements; and current policy surfaces such as EU AI Act GPAI summaries and U.S. Copyright Office AI reports make training-content, copyright, disclosure, and responsibility obligations procurement-relevant. RDLLM-L174 therefore turns attribution from a provider self-claim into a signed reliance contract: marketplace listing, enterprise procurement, regulator export, creator challenge routing, footer survival, source-machine readability, correction SLAs, settlement holds, and jurisdiction mappings all have to bind to the L173 adversarial provenance quorum, with negative fixtures for provider terms overriding creator payment, router bypass, missing regulator export, marketplace listing without proof, and source footer removal. https://arxiv.org/abs/2602.21045 https://arxiv.org/abs/2605.05687 https://arxiv.org/abs/2605.21002 https://digital-strategy.ec.europa.eu/en/factpages/general-purpose-ai-obligations-under-ai-act https://www.copyright.gov/ai/ https://c2pa.org/specifications/specifications/2.0/security/Security_Considerations.html
-
The L175 pass closes the provider onboarding and migration gap. Recent fabricated-citation audits show that source-looking footers cannot be trusted unless the generated citation is verified against real external records; PaperTrail and Source Attribution in RAG show that useful attribution must bind claim/evidence rows and source utility rather than decorate whole answers; AEX shows that provider and gateway APIs need request-output and streaming provenance at the boundary; and DataDignity shows that true answer support must be separated from lexical or topical resemblance. RDLLM-L175 therefore requires each foundation provider family, native API surface, SDK shim, gateway proxy, marketplace listing, regulator export, customer migration guide, rollout gate, rollback plan, and negative fixture to bind to the L174 reliance contract. A provider cannot claim adoption if legacy endpoints omit footers, SDKs strip source rows, streaming finals miss receipts, tool outputs are unattributed, model aliases drift, proxies bypass enforcement, or private payloads leak in migration logs. https://arxiv.org/abs/2605.07723 https://arxiv.org/abs/2603.03299 https://arxiv.org/abs/2602.21045 https://arxiv.org/abs/2507.04480 https://arxiv.org/abs/2603.14283 https://arxiv.org/abs/2605.05687
-
The L170 pass closes the production-invocation-admission gap. A provider can have a current conformance receipt while an individual live call still bypasses the route, model alias, tenant scope, telemetry span, source-footer gate, revocation snapshot, tool authorization, or settlement meter. RDLLM-L170 therefore requires a signed admission decision for each live invocation before response release, footer reliance, tool/MCP execution, retrieval reliance, or creator settlement. Cited but Not Verified motivates blocking unsupported source-footer reliance at response time; OpenTelemetry GenAI motivates per-call trace binding; MCP motivates explicit tool/resource authorization; OpenAPI and Arazzo motivate machine-readable request/workflow admission; and SLSA, in-toto, and SCITT motivate signed verifier artifacts and transparency publication for the resulting admission decisions. https://arxiv.org/abs/2605.06635 https://opentelemetry.io/docs/specs/semconv/gen-ai/ https://modelcontextprotocol.io/ https://spec.openapis.org/oas/ https://spec.openapis.org/arazzo/latest.html https://slsa.dev/spec/latest/ https://in-toto.io/ https://scitt.io/
-
The L176 pass closes the dynamic model/provider registry gap. Provider onboarding does not cover the rate at which model catalogs, hosted open-weight services, marketplace aliases, local runtimes, regional catalogs, private deployments, and router/gateway routes change. Current catalog APIs expose model IDs, publishers, modalities, limits, and capability metadata, so RDLLM now treats each concrete route as a registered service object with lifecycle, adapter, source-footer, and settlement metadata. Unregistered routes, stale catalogs, silent router fallback, capability overclaims, and private payloads fail closed before RDLLM support or settlement is claimed. https://docs.github.com/en/rest/models/catalog
-
The L177 pass closes the visible source-footer enforcement gap. Recent 2026 work shows that citation-looking output is not enough: citation hallucinations are measurable at scale, deployment constraints can increase fabricated references, RAG source attribution needs document-level utility analysis, and nugget/evidence-pool RAG designs preserve provenance by carrying structured evidence through generation. RDLLM-L177 therefore binds every L176 route to source discovery, retrieval trace capture, claim decomposition, evidence-span binding, source identity resolution, citation metadata verification, counterevidence scanning, no-source abstention, visible and machine-readable footer rows, copy/export preservation, and settlement holds before answer release. This makes the footer a release gate, not decoration. https://arxiv.org/abs/2605.07723 https://arxiv.org/abs/2603.07287 https://arxiv.org/abs/2507.04480 https://arxiv.org/abs/2601.13222 https://arxiv.org/abs/2604.14170
-
The L178 pass closes the provider catalog coverage gap. Dynamic provider catalogs and model marketplaces can add, rename, deprecate, region-lock, or silently route models after a registry is published. RDLLM-L178 therefore requires provider endpoints, OpenAI-compatible model lists, cloud catalogs, marketplace listings, hosted open-weight registries, routers, private catalogs, local runtime manifests, regional catalogs, SDK metadata, billing meters, and lifecycle feeds to be exhaustively normalized. Every discovered model must either map to an L176 route with L177 footer enforcement or be explicitly blocked with settlement held. This turns "works with every model" into a replayable catalog-coverage claim rather than a stale hard-coded list. https://docs.github.com/en/rest/models/catalog
-
The L179 pass closes the live route-binding gap. A provider catalog can be complete and still fail at runtime if an SDK, gateway, router, cache layer, stream final event, tool callback, or fallback path returns a different model than the preflight route. RDLLM-L179 therefore requires request model IDs, resolved model IDs, provider response model echoes, route IDs, telemetry spans, source-footer route hashes, and settlement-meter route hashes to remain bound to the same L178 catalog-covered route. Runtime alias drift, silent fallback, stale cache reuse, missing streaming model echoes, and source-footer or meter mismatches fail closed before release.
-
The L180 pass closes the verified source-footer reliance gap. Recent 2026 citation audits show that working links and plausible citation strings are insufficient: systems need source materialization, metadata fidelity, link health, relevance, factual support, claim-evidence matching, and explicit accounting for sources used but omitted. RDLLM-L180 therefore binds every visible footer row to the L179 runtime route, L177 footer enforcement, L171 source-grounded response receipt, and L141 citation verification contract. A footer row cannot support user confidence, copy/export, or creator settlement unless the source resolves or is archived, matches its metadata, factually supports the claim, is relevant to the claim, has claim/evidence hashes, survives copied output, and passes negative fixtures for fabricated citations, wrong-source claims, post-hoc footer insertion, and omitted used sources. https://arxiv.org/abs/2605.06635 https://arxiv.org/abs/2605.07723 https://arxiv.org/abs/2605.27700 https://arxiv.org/abs/2604.03173 https://arxiv.org/abs/2602.21045 https://arxiv.org/abs/2604.05467 https://arxiv.org/abs/2507.04480
-
The L181 pass closes the model-capability coverage gap. A route can be catalog-covered, runtime-bound, and source-footer verified for text chat while other declared surfaces still bypass attribution. RDLLM-L181 therefore requires fixture-backed coverage for capabilities, modality pairs, operation surfaces, catalog-covered routes, source-footer-or-abstention behavior, settlement-meter binding, and negative fail-closed cases before provider capability claims are treated as attribution-safe.
-
The L182 pass closes the live provider-capability discovery gap. Capability support can drift through model renames, endpoint changes, lifecycle retirement, region restrictions, pricing/rate-limit changes, and tool or multimodal feature changes after L181 fixtures were written. RDLLM-L182 therefore requires fresh official or attested provider-source rows, discovery-channel rows, capability discovery rows, route discovery rows, and negative live-discovery fixtures before source footers, answer release, or creator settlement can rely on a provider capability claim. https://arxiv.org/abs/2605.06635 https://arxiv.org/abs/2605.07723 https://arxiv.org/abs/2605.08583 https://arxiv.org/abs/2605.27700 https://www.microsoft.com/en-us/research/publication/datadignity-training-data-attribution-for-large-language-models/ https://platform.openai.com/docs/api-reference/responses https://platform.claude.com/docs/en/build-with-claude/citations https://ai.google.dev/gemini-api/docs/models https://docs.aws.amazon.com/bedrock/latest/userguide/conversation-inference.html
-
The L183 pass closes the native source-annotation normalization gap. Even when a provider has current capability evidence, its response may encode sources as URL annotations, file annotations, document citations, grounding chunks/supports, streaming citation deltas, router annotations, RAG contexts, local manifest source maps, or media source metadata. RDLLM-L183 therefore requires parser fixtures for every native annotation format, normalized footer-field rows, route annotation rows, verified footer bindings, and fail-closed negative fixtures before the answer can present source footers or release settlement. https://platform.openai.com/docs/api-reference/responses/create https://platform.openai.com/docs/guides/tools-web-search https://platform.openai.com/docs/guides/tools-file-search https://platform.claude.com/docs/en/build-with-claude/citations https://ai.google.dev/gemini-api/docs/google-search https://docs.aws.amazon.com/bedrock/latest/userguide/conversation-inference.html
-
The L184 pass closes the claim-evidence footer verification gap identified by the newest citation-failure literature. Cited but Not Verified shows that working links and relevance can coexist with poor factual support; Verified Misguidance / CITETRACE reports structurally misleading citations where real accessible sources still fail intent-purpose, source-suitability, or answer-source-fidelity checks; PaperTrail argues for claim/evidence mappings rather than coarse citations; CiteAudit and the large-scale hallucinated citation audit show why fabricated or unsupported references need automated verification before they enter scholarly or user-facing outputs. RDLLM-L184 therefore requires structured citation parsing, materialized cited content, claim hashes, evidence spans, source suitability, intent-purpose alignment, answer-source fidelity, factual-support verdicts, and fail-closed negative fixtures for correct-answer/wrong-source, inaccessible-source, post-hoc citation, and private-source-text leakage cases before source-footers can be relied on or settlement released. https://arxiv.org/abs/2605.06635 https://arxiv.org/abs/2605.28565 https://arxiv.org/abs/2602.21045 https://arxiv.org/abs/2602.23452 https://arxiv.org/abs/2605.07723
-
The L185 pass closes the provider-meter normalization gap that appears once source-grounded answers have to run across multiple model providers, routers, tools, media endpoints, and cache/batch surfaces. Recent attribution work reinforces that systems must separate true answer support from topical resemblance and citation appearance: CiteCheck verifies citation existence and metadata fidelity, DataDignity evaluates whether a source document actually supports expressed response knowledge, Source Attribution in RAG studies per-source contribution accounting, and CUE-R measures per-evidence utility beyond final-answer quality. L185 therefore normalizes provider-native usage fields into RDLLM settlement meters for input/output tokens, cached input, reasoning or thinking tokens, tool/search units, media units, batch/asynchronous units, embeddings/rerank/fine-tune units, hosted-runtime units, pricing snapshots, quota/rate-limit surfaces, and provider invoice rows. The official provider docs currently expose these meter shapes through OpenAI Responses
usagefields, Anthropic Messagesusageand prompt-cache counters, GeminiusageMetadata, Bedrock Converseusage, Mistral cached prompt usage, Coherebilled_units, xAI OpenAI-compatible usage and cached-token details, and OpenRouter pass-through usage/cost accounting. https://arxiv.org/abs/2605.27700 https://www.microsoft.com/en-us/research/publication/datadignity-training-data-attribution-for-large-language-models/ https://arxiv.org/abs/2507.04480 https://arxiv.org/abs/2604.05467 https://platform.openai.com/docs/api-reference/responses https://docs.anthropic.com/en/api/messages https://docs.anthropic.com/en/docs/build-with-claude/prompt-caching https://ai.google.dev/api/generate-content https://docs.aws.amazon.com/bedrock/latest/userguide/conversation-inference.html https://docs.mistral.ai/studio-api/conversations/advanced/prompt-caching https://docs.cohere.com/docs/how-does-cohere-pricing-work https://docs.x.ai/developers/advanced-api-usage/prompt-caching/usage-and-pricing https://openrouter.ai/docs/cookbook/administration/usage-accounting -
The L186 pass closes the provider response-state gap between a cited answer and a reliable cited answer. Recent work supports treating citations as claim/evidence objects, not decorative footnotes: Citation-Closure Retrieval and Per-Rule Attribution requires source-bound rule attribution, PaperTrail decomposes answers and evidence into discrete claims to expose unsupported assertions, Cited but Not Verified reports that link validity can remain high while factual support is much lower, and SourceCheckup finds many cited medical answers are not fully source-supported. L186 therefore requires provider-native terminal states to be normalized before the answer footer can imply grounded reliance: OpenAI
status,incomplete_details, refusals, andfinish_reason; Anthropicstop_reasonand streaming refusals; GeminipromptFeedback,blockReason, andsafetyRatings; BedrockstopReasonand guardrail traces; Azure OpenAI content-filter results; Mistral guardrails; Cohere finish reasons and safety errors; xAI/OpenAI-compatible refusal and finish states; OpenRouterfinish_reason/native_finish_reason; router errors; local runtime exits; and streaming final events. Blocked, refused, filtered, truncated, tool-only, errored, or unknown terminal states must hold answer release, footer reliance, and creator settlement until the response is either regenerated with support or publicly labeled as an abstention/refusal. https://arxiv.org/abs/2605.29742 https://arxiv.org/abs/2602.21045 https://arxiv.org/abs/2605.06635 https://www.nature.com/articles/s41467-025-58551-6 https://arxiv.org/abs/2504.01032 https://platform.openai.com/docs/api-reference/responses https://platform.openai.com/docs/api-reference/chat/object https://docs.anthropic.com/en/api/messages https://docs.anthropic.com/en/docs/test-and-evaluate/strengthen-guardrails/handle-refusals https://ai.google.dev/api/generate-content https://docs.aws.amazon.com/bedrock/latest/userguide/conversation-inference.html https://docs.aws.amazon.com/bedrock/latest/userguide/guardrails-use-converse-api.html https://learn.microsoft.com/azure/ai-foundry/openai/how-to/content-filters https://docs.mistral.ai/api/ https://docs.cohere.com/reference/chat https://docs.x.ai/docs/api-reference https://openrouter.ai/docs/api-reference/chat-completion -
The July 2026 runtime pass applies the same citation-faithfulness lesson to the default engine path. External attribution now discovers sources from the generated output body rather than from prompt intent alone, so a topical prompt cannot cause unrelated retrieved works to appear in the user footer. Footer rows are emitted only for output-supported sources, carry visible claim counts, support scores, payout amounts, content hashes, and confidence labels, and each claim-evidence row exposes a claim hash plus exact evidence-span hash. This implements the practical lesson from recent source-attribution work: users need a footer that says which work supports which answer claim, not a decorative list of plausible citations or a hidden payout ledger. https://arxiv.org/abs/2605.06635 https://arxiv.org/abs/2601.05866 https://arxiv.org/abs/2510.17853 https://arxiv.org/abs/2605.27700 https://www.cambridge.org/core/journals/data-and-policy/article/attribution-crisis-in-llm-search-results-estimating-ecosystem-exploitation/170DD0B88E5F5AEA8F69F2E9AF1328E3 https://aclanthology.org/2025.findings-acl.1087/
-
The July 2026 public-surface privacy pass separates answer grounding from private payload publication. Cited but Not Verified shows that end-user trust depends on retrievable, fact-checked citations; CiteCheck and CiteAudit show that citation identity and metadata fidelity must be externally verified; and A Human-Centric Framework for Data Attribution in LLMs argues that creator, user, platform, and intermediary goals must be negotiated rather than collapsed into a single payout metric. RDLLM therefore treats public source footers as selective-disclosure artifacts: they show source labels, titles, URIs, support scores, claim counts, hashes, verifier handles, and settlement commitments, while hosted public artifacts are audited to exclude prompts, raw source text, private reasoning, secrets, customer records, payment details, and payout account data. https://arxiv.org/abs/2605.06635 https://arxiv.org/abs/2605.27700 https://arxiv.org/abs/2602.23452 https://arxiv.org/abs/2602.10995
-
The July 2026 attribution-metric pass adds a release rule for evaluation claims: no single automatic attribution scorer should be treated as a universal grounding proxy. Do LLM Attribution Metrics Transfer? reports that scorer rankings invert across generated-answer attribution datasets and that metric choice must be validated against the target construct and domain. RDLLM therefore treats claim support, source materialization, footer rendering, confidence calibration, and payout attribution as separately verified channels rather than collapsing them into one score. https://arxiv.org/abs/2606.23915
-
The July 2026 influence-attribution pass keeps user-facing source footers separate from model/training influence evidence. TokenTrace shows that generated media can require multi-concept influence attribution through watermark recovery, while RDLLM's answer footer is a runtime claim/evidence and settlement surface. Implementations should not use footer rows as a substitute for training-data influence receipts, and should not use influence signals alone to imply that a displayed answer claim is source-grounded. https://arxiv.org/abs/2602.19019