RDLLM Explained

ELI5

Imagine an AI answer is a school report. RDLLM makes the AI show which books it used, which sentence came from which book, and whether the people who made those books should get credit or payment. If the proof is missing, the answer should not pretend to be fully sourced.

Simple

RDLLM is an open-source layer for AI products. It adds a visible source footer to an answer, checks that the sources really support the claims, and records how source usage maps to payout or escrow.

Use it when you are building a chatbot, RAG app, search assistant, agent, model API, marketplace, or creator platform and you want users to see where an answer came from.

Non-Technical

RDLLM helps three groups:

Users see sources, claim evidence, and confidence signals instead of a bare AI answer.
Creators and publishers get attribution evidence and settlement records.
Operators get verification gates before showing an answer as grounded or royalty-bearing.

The system does not claim that it can read a model's hidden thoughts. It reports observable evidence: visible source support, answer-to-source overlap, claim evidence, source-disagreement checks, and payout allocation.

Technical

At runtime, RDLLM takes an answer, source references, and revenue context. It builds a source footer with source rows and Claim Evidence rows. Each claim row includes a claim hash, support score, evidence span hash, character offsets, evidence preview, warrant status, source-disagreement status, and source label.

Before display, verifiers recompute the display hash, footer hash, row hashes, source usage metrics, claim evidence, citation markers, answer links, claim-source closure, model-reliance wording, attribution-gap closure, and source disagreement. The response is public-facing only when production_display_ready is true and source_grounding_acceptance is passed.

For implementation, start with: