Skip to content

Universal Reduction Plane

Universal Reduction Plane (URP) is an open-source, compatibility-first control and execution plane for reducing unnecessary storage, data transfer, and AI compute. It preserves application interfaces and requires verification before a reduced or cached result is accepted.

WorkUnit + Contract + Policy -> Plan -> Execute -> Verify -> Manifest + Ledger

Start with a working proof

git clone https://github.com/thewisecrab/urp.git
cd urp
python3 -m venv .venv
.venv/bin/python -m pip install -e ".[dev]"
.venv/bin/python examples/live/run_live_examples.py --reset

The proof runs without cloud credentials. It demonstrates exact object restore, range reads, legal-hold enforcement, an exact AI cache hit, a lakehouse adapter, redacted manifests, savings metrics, and ledger-chain validation.

The lifecycle

flowchart LR
    A[Input] --> B[WorkUnit]
    B --> C[Contract]
    C --> D[Policy]
    D --> E[Plan]
    E --> F[Execution]
    F --> G{Verification}
    G -->|Accepted| H[Manifest]
    G -->|Rejected| I[Baseline fallback]
    H --> J[Ledger and metrics]
    I --> J

What URP guarantees by default

  • Unknown data stays exact.
  • Tenants do not share cache or dedupe state.
  • A client cannot declare its own role or verifier success.
  • Exact outputs are checked before acceptance.
  • Semantic, approximate, lossy, and deletion behavior is disabled until policy explicitly permits it.
  • Failed candidate outputs fall back to the baseline path.
  • Every accepted action produces lineage and audit evidence.

Where to go next

Goal Read or run
Understand the product Product explainer
Evaluate the thesis and impact Technical paper
Deploy locally or on Kubernetes Deployment quickstart
Integrate an API or SDK API and protocols
Review security defaults Policy, security, and compliance
Check platform boundaries Platform readiness
Find a direct answer FAQ

Evidence, not universal claims

The included repetitive CSV fixture shrinks from 11,556 bytes to 249 bytes and restores exactly. The repeated mock AI request avoids one of two provider calls. Those results prove the paths, not production ratios.

Use the checked-in impact model with your own measured rates:

urp report impact --scenario examples/impact/illustrative-portfolio.json

The model reports gross and net impact, implementation payback, avoided bytes, tokens, and calls, plus optional energy and carbon only when measured telemetry is provided.