entroly

Entroly public evidence policy

Entroly separates distribution, implementation, reproducible measurements, production outcomes, and external marketplace status. A lower evidence tier must never be presented as proof of a higher one.

Evidence tier What it establishes Required support
Distribution A package or license is publicly available Direct registry or license link and exact package name
Implementation A capability exists in the repository Source, tests, and accurate default/optional wording
Reproducible measurement A result occurred under a defined protocol Committed artifact, configuration, sample size, command, and caveats
Production outcome A user workload achieved an outcome Provider-observed usage, workload, baseline, and uncertainty
Marketplace status A third party indexed or validated a release The current third-party page after publication

A package badge does not prove benchmark quality. A local token estimate does not prove a billing reduction. Repository-local tests do not prove that an external marketplace validated a release.

The base Python package has a pure-Python path. Rust acceleration is optional. The npm entroly package is a separate Node/WASM runtime.

MCP launch contract

The canonical installed-Python MCP registration is the argument-free entroly stdio command. Under an MCP client’s stdin pipe, it launches the installed server. The uvx and entroly-mcp registrations likewise use no serve argument.

entroly serve is the explicit Docker-first deployment path. ENTROLY_NO_DOCKER=1 entroly serve selects the installed Python runtime.

Reproducible evidence

Context Commit integrity

The committed synthetic conformance artifact reports:

These results measure artifact integrity and recovery, not answer quality or identical cross-engine selection.

Recovery-resilience holdout

The committed v5 recovery-resilience revalidation records 66/66 exact entries for Entroly 1.0.66 source and 66/66 for the External Baseline A 0.31.0 comparison under the frozen holdout protocol. This result establishes parity, not leadership for that recovery-integrity workload. It is historical, release-scoped evidence; later Entroly implementations require a new frozen revalidation before this result can be applied to them.

The clean revalidation reproduced no External Baseline A worker errors. It therefore does not permit a public leadership claim and does not establish universal recovery superiority, production reliability, task-quality improvement, or provider-cost reduction.

WITNESS HaluEval-QA

The committed faithful protocol reports 0.7976 full-dataset AUROC and 84.92% accuracy on the 16,000-decision held-out split. On the shared 1,200-decision GPT sample, committed accuracy is 86.58% for WITNESS and 86.25% for gpt-4o-mini.

The uncertainty overlaps, so Entroly does not claim superiority, universal truth, or general hallucination prevention from this run.

Token reduction and task quality

Token reduction varies by corpus, query, budget, tokenizer, integration, provider, cache behavior, and recovery path. Use entroly simulate for a local estimate and provider-observed request usage before making a production billing claim.

The same-input compression gauntlet is not production-outcome evidence. It is a versioned synthetic protocol for named fixtures.

Model-triggered recovery

A frozen 24-case local Qwen2.5-1.5B holdout recorded 24/24 exact final answers for Entroly and 18/24 for the published External Baseline A 0.31.0 baseline. This is a synthetic, versioned workflow—not a universal product, provider-savings, or model-quality claim.

PRISM-R neural research pilot

PRISM-R is an opt-in research prototype, not the default compressor.

On a frozen 200-pair same-document query-shift pilot at a nominal 25% active budget, PRISM-R retained 87.0% of current-query exact evidence versus 60.5% for lexical selection. A different future question was revealed only after compression; exact local span recovery raised future evidence retention from 9.0% to 90.5%. Active plus recovered text was approximately 50.6% of the original.

These results measure exact answer-string retention on short SQuAD paragraphs. They do not measure generated answers, general neural superiority, long-agent memory, production latency, or billing savings.

Marketplace status

The LobeHub listing is an external discovery surface. Only the live LobeHub page can establish the current external result. Repository readiness, package publication, or local MCP tests cannot establish third-party validation.

Marketplace badges are deliberately excluded from the simplified README until external indexing reflects the current release and validation state.

Quarantined public surfaces

Legacy savings, prompt-compression, hallucination, projected-dashboard, and stale setup pages remain noindex redirects until their claims and setup instructions are rebuilt. An HTTP-successful page is not sufficient evidence that its copy or runtime is current.

Archived translated READMEs must be regenerated from the canonical README, including trust links and caveats, before they return to primary navigation.

Retired and republished public pages

A set of topic pages was reduced to noindex tombstones redirecting here, because their claims could not be sourced: a universal 70–95% range, a 0.844 AUROC that a later tie-correction retired, equivalence conclusions drawn against an API judge, and verifier coverage described as “every response”.

Retirement removed those claims by removing the pages. It also removed every entry point to the topics, and left the repository with no indexable answer to questions Entroly can answer honestly.

Those pages were republished on 2026-08-15 under a stricter condition than the one they failed: every figure on a republished page must resolve to a committed artifact under benchmarks/results/, and the page must state the workload that produced it. Where a benchmark set contains a loss case, the page states it next to the wins rather than omitting it — the SQuAD 2.0 row (43.8% savings, 90% retention on 233-token inputs) appears alongside the long-context results it is worse than.

Retirement is no longer what keeps these pages honest. STALE_PUBLIC_CLAIMS in scripts/verify_context_assurance_public.py is, and every republished page is listed in CLAIM_SENSITIVE_PUBLIC_FILES so that scan applies to it. A page must not be republished by deleting its retirement entry without adding it there.

docs/dashboard.html remains retired: it is an application view, not content.

Maintainer rules

Before adding or strengthening a public claim:

  1. Link the exact package, source, or result.
  2. State the workload, model, budget, sample size, baseline, and caveats.
  3. Keep different benchmark protocols separate.
  4. Label estimates as estimates and provider-observed usage as observed.
  5. Do not infer a marketplace score from repository-local evidence.
  6. Remove or soften claims that cannot be reproduced.

Run:

python scripts/verify_public_trust.py
python scripts/verify_readme.py

Use --online only for bounded destination checks. After publication, --require-published-version can require PyPI and npm latest versions to match server.json.