docs: reference-quality package checklist (phase 4.4)

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PDKeXvpT6tENSvyQGLV1Uq

docs/reference-standard.md

@@ -0,0 +1,47 @@
+# Reference-quality package — the checklist
+
+Derived from the artificial-intelligence-engineer reference run (2026-07-09,
+branch `quality/reference-standard`). A package is reference quality when
+every box below holds; `qa/lint_package.py` enforces the machine-checkable
+subset as the publish gate.
+
+## Market evidence
+- [ ] Every stored ad passed the relevance gate (blocklist + gemma3 judge
+      against the ESCO description); rejects logged in `rejected-ads.jsonl`
+- [ ] Corpus ≥ 100 relevant ads total, or the report says so explicitly
+- [ ] Regional sections (US / UK / EU-DACH); regions under 30 ads state
+      "insufficient evidence" instead of ranking noise
+- [ ] No foreign-role tools (designer/marketing stack) in any top-10 ranking
+
+## Agent-skill mapping
+- [ ] Every entry tiered `core` or `adjacent`; core = touches a top-20 market
+      hard skill or an essential ESCO competence
+- [ ] Max 10 entries per source repo; overflow only in the audit log
+- [ ] Per-source attribution complete: repo URL, commit, license, retrieval
+      date; links point at a pinned commit (`tree/<sha>`)
+- [ ] Nothing copied from sources — name, one-line summary, link only
+
+## SKILL.md
+- [ ] Intro is one grammatically clean paragraph (rewrite cached in
+      `data/rewrites/`); raw ESCO text reachable via the sources footer
+- [ ] Essential competences complete — never a silently cut alphabetical list
+- [ ] "Hot technologies" derived from the post-gate market ranking; omitted
+      below 30 gated ads; never an alphabetical O*NET dump
+- [ ] O*NET proxy caveat present wherever the crosswalk match is not exact
+- [ ] Curated content (core workflow, intro) lives in `overlays/<slug>/` and
+      survives regeneration
+
+## Counts & provenance
+- [ ] Every displayed number traces to generated `stats.json` — nothing
+      hard-coded in UI or READMEs
+- [ ] `PROVENANCE.md` present with all five source categories and pinned
+      colors; manifest carries esco_uri, onet_soc (+ match type), collar
+- [ ] Raw fetch data and extraction outputs exist on disk and were never
+      edited or deleted
+
+## Process
+- [ ] `qa/lint_package.py <slug>` exits clean
+- [ ] Both eval suites at ship level (`evals/ad-relevance` ≥ 27/30,
+      `evals/skill-tiering` ≥ 18/20)
+- [ ] Regeneration is a single command (`pipeline/regen_flagship.py <slug>`)
+      and reproduces the package from stored data