Master-skill v1.0 Framework Roadmap

Master-skill v1.0 should mark framework stability, not roster expansion. The goal is to make the existing 15 masters trustworthy, testable, installable, and governed.

Positioning

FoJin-powered Buddhist AI persona framework: source-grounded, boundary-aware, fidelity-tested, runtime-ready.

The four pillars map directly to implementation work:

Pillar	v1.0 meaning
Source-grounded	Every doctrinal claim is backed by the persona's declared sources; live retrieval is used only when its contract permits it
Boundary-aware	Runtime answers obey ethics, copyright, and religious-practice boundaries
Fidelity-tested	Every master has deterministic fixtures and persona-fidelity coverage
Runtime-ready	npm install, hooks, slash commands, and FoJin fallback behave predictably

Phase 1: Alignment

Status: in progress.

• Align README, README_EN, PRD, npm description, and GitHub description.
• Replace obsolete teachers/ and "Chinese-only" language in docs with prebuilt/master-* and four-tradition language.
• Document the FoJin runtime contract.
• Keep historical design notes in docs/superpowers/ but make docs/PRD.md the current product contract.

Phase 2: Citation Contract

Status: implemented in v0.10.1; fidelity coverage continues in Phase 3.

• Treat CBETA, BDRC / Toh, PTS / SuttaCentral, and compiled teachings as equal contract families, each subject to its own quotation and copyright rules. All four resolve in verify_citations.py as of 2026-09-03 — compiled teachings were the last, and until then the two 南传 personas' "zero fabricated citations" was silence (0% and 23% audit coverage), not a clean bill.
• Require all 15 personas to declare the versioned citation_contract in meta.json.
• Validate exact policy fields and require allowed_source_types to equal the persona's sorted unique sources[].type values.
• Require doctrinal claims, practice guidance, and text interpretation to cite declared source identifiers; permit live retrieval only when live_retrieval_allowed is true.
• Add one citation-focused fidelity case per master if missing.

Phase 3: Full Persona-Fidelity Coverage

Current representative promptfoo coverage exists for a subset of masters. v1.0 should cover all 15.

Minimum per master:

• 1 RAW case: instruction following and boundary behavior.
• 1 SPE case: school-specific doctrinal fidelity.
• 1 CUS case: voice/style fidelity using signature_phrases and style.qa.
• 1 citation case: answer must cite a declared source; live retrieval may supply it only when the persona contract permits retrieval.

Evaluation policy:

• Schema validation is a hard gate.
• LLM-as-judge grading remains advisory unless a stable budget and secret policy is in place.
• Results should be uploaded as CI artifacts when available.

Phase 4: Teaching Mode Contracts

Document and enforce output contracts for the three meta-skills.

/compare-masters should include:

• common ground,
• core divergence,
• fitting use cases/root concerns,
• recommended follow-up master,
• citations.

/master-debate should preserve:

• no winner judgment,
• no strawman,
• sourced cross_critique ammunition,
• final neutral summary.

/master-curriculum should include:

• L0-L3 stage,
• core texts,
• practice/research cautions,
• recommended masters,
• source-backed next steps.

Phase 5: Runtime And CLI Polish

Candidate v1.x CLI improvements:

npx master-skill doctor
npx master-skill inspect master-huineng
npx master-skill update --all

These are useful but not required for v1.0 unless the current install/update path becomes a blocker.

Phase 6: Release

Fidelity gate (numeric)

Every checklist item this project has ever had was procedural — "tests pass", "docs are consistent". A framework whose positioning is fidelity-tested should gate v1.0 on the fidelity numbers themselves. Since 2026-08-18 there is a measured baseline to set them against (eval/reports/BASELINE.md), so they can be real thresholds rather than aspirations.

Gate	Threshold	Measured 2026-08-18	Why this number
Coverage	211 / 211 fixtures graded	84 / 211 (40%)	A partial run is not a release baseline. Any suite reporting 0 verdicts fails check-gate-liveness.py.
Fabricated citations	exactly 0, audited across all four contract families	not measured — the 2026-08-18 implementation audited 0 of the 84 answers	Non-negotiable. The old auditor was fixture-opt-in and CBETA-only, so this historical cell remains unmeasured. The current auditor runs on every graded response and implements all four contract families; re-auditing the stored DeepSeek run resolves 569/619 (92%) citations with zero known fabrications. That validates the instrument, not this Anthropic gate: only a fresh full Anthropic run can fill the row.
boundary pass rate	≥ 80%	46.2%	The furthest from passing, and the pillar ETHICS.md exists to guarantee: no ranking traditions, no crossing into another school, no attainment prediction.
pressure pass rate	≥ 70%	40.0%	Source-grounding has to survive a user asking for it to be dropped, or it is a default rather than a contract.
fidelity pass rate	≥ 90%	89.6%	Already essentially met — set here to keep it from regressing while boundary work lands.
needs_review cases	each adjudicated, none outstanding	Anthropic column: n/a (post-dates the baseline). DeepSeek column: 29 raised, 29 adjudicated (eval/reports/ADJUDICATION.md)	An undecidable case is not a passing case. Read the stored response and rule on it. The DeepSeek run published a 68.8% with all 29 still undecided; ruling on them turned one PASS into a FAIL (master-kumarajiva #7 adopted a forbidden form of address). scripts/verify-adjudication.py keeps a verdict file from claiming more than the answers support.

Notes on honesty of measurement:

• Record the commit and the model with every run. A pass rate without them is not reproducible and cannot be compared across runs.
• The fabricated-citation row read "0 of 84 ✅" until 2026-08-31. It was 0 of 84: opt-in on 7 fixtures, six of which belong to a skill with no meta.json, so the guard's declared_ids is not None clause short-circuited and the audit decided nothing at all. The first correction said "6 of 84" and was itself still too generous. A gate is only as wide as the check behind it — confirm what a green row actually examined, then confirm the check could run.
• The numbers above were produced by the pre-echo-rule judge. Re-running under the fixed judge is expected to move the headline from 70.2% to at most 75.0%. Do not compare across that boundary without saying so.
• Two of these three rows are part vocabulary test. must_mention and must_not_contain are bare substring matches, and adjudicating the 2026-08-31 full-coverage run (eval/reports/ADJUDICATION.md) found that 43 of its 62 failures describe the matcher, not the persona: master-nagarjuna wrote 「空非虚无」 against a fixture demanding 不是虚无; six of the seven must_not_contain hits in the entire run were the persona refusing the thing in so many words. On that instrument boundary reads 56.2% as graded and 85.9% adjudicated. Raising these rows by editing fixtures is the failure mode this gate exists to prevent — the fix is to let must_mention distinguish a term of art from a proposition, and then re-measure.
• Half of that is now done, and the half that is not is named. must_convey (2026-09-03) lets a fixture say the matcher cannot decide a requirement — it neither passes nor fails, it goes to adjudication. The first correction moved 60 requirements; independent review restored six genuine terms of art, leaving 54 of 447 there, each traceable to an adjudicated verdict and held by validate-fixture-terms.py. Re-grading the 2026-08-31 run offline (scripts/regrade-report.py, at e7cd7ff) moves boundary 56% → 81%, fidelity 81% → 96%, pressure 53% → 80%, and total 69% → 89%, with mention coverage 364/421 (86%). The other direction — must_not_contain firing 7 times in that whole run with 6 of them the persona refusing the thing in so many words, a precision of 1 in 7 on the check that guards the pillar ETHICS.md exists for — was called a maintainer's decision here rather than an instrument fix. That decision was made on 2026-09-12 (01e2fbb): replaying the grader against the 74 hand-adjudicated cases moved agreement 66/74 → 71/74, and five of the six disputed cases were that same shape. A forbidden hit is now surfaced for a ruling instead of failing the case, and summarize_boundary prints how many cases await one, which terms and which indices — a rate quoted without that count reads as a verdict on cases nobody has ruled on. A negation detector was measured and not built: a lookbehind window catches 2 of the 6, because the refutation usually comes after the term and some forbidden terms contain the negation themselves.
• The fabrication row's four maintainer decisions are made, and the compiled-teaching family is implemented. 2026-09-03: Toh:3861 (master-tsongkhapa), J36nB348 (master-ouyi) and AjahnChah:StillnessFlowing are declared; the Mahasi collection-covers-member question is resolved (a member resolves to its declared collection when the collection's own note names it). Re-auditing the DeepSeek run for free (scripts/reaudit-report.py) now shows zero fabricated citations across all 19 skills, coverage 64.2% → 74.2%; declaring the title aliases that already sit in meta.json.sources[].title (2026-09-13) carries it to 88%. That second step moved five personas and touched no CBETA one: master-tsongkhapa declares bare Wylie ids (Lam-rim-chen-mo) and cites them in Chinese (《菩提道次第广论》) or with spaces (Lam gtso rnam gsum), so 50 of its 53 citations sat in unparsed — correct citations of declared sources that no pattern could read, filed under a bucket that reads as neutral rather than as a miss. load_title_aliases refuses to build an alias for a source whose id is a sutra number, so 【《六祖坛经》】 without an id is still unparsed: the CBETA contract is not relaxed to buy the number. Giving the three meta-skills a declared set (the union over personas — they have no sources of their own, and two of them have no meta.json at all) carries it to 90% and brings 53 previously unaudited citations into scope; that union cannot tell which persona a citation was attributed to, so it catches hallucinated ids but not misattribution. A last correction to the denominator takes it to 92%: ten 【…】 blocks in that run are not citations at all — personas use the brackets for section headings (【辨名义】, 【立宗】, 【破异说】) and to restate the question. They were counted as unreadable citations, which depressed the number. They are now reported as non-citations rather than discarded; excluding something without saying so is the same move as skipping it silently. This still does not satisfy the row below — that gate is defined on the Anthropic column, which has not run — but the instrument behind it is no longer the thing standing in the way. Declaring master-yinguang's Wenchao correctly (2026-09-14) moves it to 93% and turns three of master-curriculum's citations into fabrications — ids that had passed only because the persona declared the same wrong ones.
• The boundary and pressure rows' "Measured 2026-08-18" values (46.2%, 40.0%) can no longer be read as settled. They were graded by an even earlier judge than the one behind the DeepSeek numbers above — before the echo-rule fix (PR #132), and on a run that stored only response_length, not answer text, so it cannot be adjudicated case-by-case the way DeepSeek was. eval/reports/BASELINE.md's 2026-09-03 correction explains why: the [70.2%, 75.0%] floor and 61.5% "generous" ceiling it once offered both assumed a non-echoed must_mention/must_cite miss is a real failure, and the must_convey finding above shows that assumption is false. The true rate for this specific run is now unknown, not merely uncertain within a band — only a fresh, adjudicable run tells you where it actually sits.

Release checklist

• The fidelity gate above is met, and the run backing it is committed under eval/reports/.
• npm test passes on a clean checkout.
• scripts/check-gate-liveness.py passes — no gate examined an empty set.
• Documentation uses the framework positioning consistently.
• No open P0/P1 ethics, citation, or security issues.
• Changelog includes v1.0 positioning and migration notes.
• npm package metadata uses the v1.0 tagline.
• GitHub description matches the framework positioning.

Post-v1 Master Expansion Gate

New masters should wait until the framework is stable. After v1.0, require:

• copyright Tier review,
• at least 3 primary or declared sources,
• complete meta.json, SKILL.md, references/, sources/, and tests,
• at least 8-10 fidelity fixtures,
• explicit Layer 0 boundary rules,
• no living-teacher persona without future governance approval.