v0.11.0 — the verification layer, verified

41 commits and 45 days of work that had been sitting on main unreleased. The theme is that the verification layer was not verified: three of this repo's shipped defects had already been the shape of a gate that examined nothing and reported green, and three more turned up here — the last of them inside this project's own baseline report.

Fixed — two that reach users directly

• Routing missed the vocabulary beginners actually use. Four of the eleven rows in the README's own 「你的状况」 recommendation table fell through to the default pairing, including the first one. 想了解空性 reached only Milarepa, because the three Madhyamaka masters had declared 性空 and never 空性 — the ordinary modern rendering of śūnyatā. If you use /compare-masters or the recommendation table, this is the reason to upgrade.
• The fidelity judge failed personas for quoting the question back. must_not_contain was a bare substring match, while boundary fixtures are baited questions carrying the loaded term themselves — 「华严宗是不是佛教最高的宗派？」 forbids 最高. A correct refusal that names the bait tripped the check exactly as hard as a real ranking did; that was 10 of the 12 forbidden-phrase failures in the first baseline. Echoed terms now set needs_review instead of failing, and every result carries the response text so a failure can actually be adjudicated.

Added — the first real measurement, and a check on the checks

• eval/reports/ — the first scored fidelity run ever committed. 211 fixtures had existed for months and not one recorded verdict did. Headline 59/84 (70.2%), and the per-test_type split matters far more than the aggregate: ordinary doctrine 89.6%, boundary 46.2%, citation-under-pressure 40.0%. The persona content holds; the guardrails do not — and the guardrails are what ETHICS.md exists to guarantee.
• scripts/check-gate-liveness.py plus 16 tests. Every test file must contribute ≥1 collected test, pytest.ini testpaths must cover every directory holding tests, a graded fidelity suite must produce ≥1 real verdict, skill-catalog.json and prebuilt/ must agree both ways. Each of the three historical defects was reproduced in a working tree to confirm the check fires.
• --provider anthropic|deepseek|gemini on the fidelity runner. This repo ships one prebuilt/ to five hosts and calls it a unified plugin, but every number it had came from one Anthropic model — and a fixture measures the prompt and the model. Cross-model pooling is refused, not merely discouraged.

Retracted

This release retracts its own "zero fabricated citations across all 84". It was six, and none of them a master persona. The audit is opt-in per fixture and only 7 of 211 fixtures set it; worse, the auditor recognises CBETA ids only, so the six masters declaring SuttaCentral / PTS: / Toh: / BDRC: / Mahasi: sources — all of 南传 and all of 藏传 — cannot be audited at all. The claim is corrected everywhere it appeared; the gap is stated, not closed. See eval/reports/BASELINE.md.

Also in this release

Desktop reliability and evaluation-contract hardening (atomic trace persistence, serialized baseline writers, invalid-argument rejection, errored suites modelled rather than dropped), ClawHub publishing wired and SHA-pinned, the README split into docs/ (21.0K → 12.1K characters; the English page 39.4K → 20.6K), and every GitHub Action converged onto one SHA-pinned version repo-wide.

Known gaps — stated, not hidden

• No master persona has ever been checked for a fabricated citation, and six of them cannot be with the current auditor.
• ANTHROPIC_API_KEY is still absent from repository secrets, so the branch-protection-required "Fidelity smoke" check writes {"skipped": true} and exits 0.
• boundary at 46.2% is a persona gap, not an instrument artifact.

Install

npx master-skill@0.11.0 install --all

Full changelog: v0.10.1...v0.11.0