jev-rules

What this skill carries — grep references/ for any subject the names below miss: !d=${CLAUDE_SKILL_DIR}; command -v skill-toc >/dev/null 2>&1 && exec skill-toc "$d"; s=$HOME/.claude/skills/plugin-utils/bin/skill-toc; [ -x "$s" ] && exec "$s" "$d"; echo "(skill-toc unavailable: references and scripts are NOT listed here — install the plugin-utils plugin, or start a new session so its bin/ reaches PATH)"

A Jev rule is three parts: a deterministic extractor (evidence()), one proposition a judge answers with P(VIOLATED), and a calibration set that proves the two separate. The runner is skills/work/scripts/rule-check.ts; the calibrator is skills/work/scripts/rule-calibrate.ts; a rule blocks at p >= 0.85.

--wire exits 0 only after two consecutive rule-calibrate --runs 2 invocations pass on the final code, with real accepted cases in the set. Never by loosening a fixture, relabelling an accepted case without the user's ruling, or moving the bar. A rule that blocks accepted work costs the user every later round it fires on; a rule wired on noise blocks at random. Both are worse than no rule.

Iron Law: SCRIPT BEFORE JEV

A rule a regex, AST walk or count settles is a script leg in the workflow's check.sh, never a Jev rule. Jev is for the judgement residue. DQ1 sat at 0.62–0.66 on its own violation and 0.78–0.91 on other rules' files until a script (ds-dq.py check_dq1) took it. T-NARRATE's accepted Texas notes held 0.50–0.61 over seven invocations; its screen-phrase lexicon, as the workshop NAR check, separates every case of its set.

The process

triage ──► ALREADY MECHANICAL / COUNTED-ADVISORY / LENS ONLY ──► stop (name the owner)
   │
   ├──► SCRIPTABLE ──► script leg in check.sh (diff-scoped to added lines)
   │
   └──► JEV RULE ──► new-rule.ts <set> <ID> ──► extractor ──► proposition ──► twins
                                                                              │
                     real accepted cases (+ legacy-base ok/bad pairs) ◄───────┘
                                   │
                     new-rule.ts --wire <set> <ID>
                        exit 0 ──► wire into the workflow (new set only) ──► trim the lens
                        exit 4 ──► diagnose (references/diagnosis.md) ──► fix ──► --wire again
                        twice failed on accepted work, rule text right ──► ask the user, or PARK

1. Triage every rule into one bucket — references/triage.md. Script before Jev.
2. Scaffold: bun ${CLAUDE_SKILL_DIR}/scripts/new-rule.ts <set> <ID> [--layout files|diff] creates the module in <set>/uncalibrated/, identical placeholder twins, the manifest entry and a stub test under tests/jev-rules/. It refuses to overwrite. --manifest <path> targets another plugin's calibration set (teaching's carries a root key).
3. Extractor — references/extractor.md: deterministic, bounded, file:line spans, the counts and booleans the proposition reads, diff-scoped via changed, never the whole file.
4. Proposition — references/proposition.md: ONE checkable claim, YES = VIOLATED, the state field it reads, every exemption named.
5. Twins and real cases — references/fixtures.md. The stub test stays red until the twins differ.
6. Wire: bun ${CLAUDE_SKILL_DIR}/scripts/new-rule.ts --wire <set> <ID>. Exit 0 moved the module; 2 = Jev unavailable (never a pass); 3 = a placeholder, TODO or empty case remains; 4 = calibration failed or a cross-rule hit (--accept-cross only when one defect rightly breaks both rules).
7. A new set also needs its workflow's ruleChecks, RULE_DIRS + ruleSetFor in hooks/jev/rules.ts, and a lens trim — references/calibration.md § Wiring.

When --wire exits 0, start the next JEV row of the triage table at once; never stop after one rule.

Facts

• Rewording the question is the first fix, never the fixture. MOCK's violating twin went 0.51 → 0.92 when the proposition named the state field (every_assertion_on_mock); the fixture was not touched.
• A whole-file or raw-search state does not separate. 9 of the 10 ds rules were unwired (213223f6) until each extractor reported a defect count (e9c530ea).
• Accepted legacy work is scored on changed lines only. DQ4/DQ6 scored 0.94–1.0 on 4 of 7 accepted scripts; diff-scoped with legacy-base pairs, bad 1.00 and ok 0.00 (9dd3be43).
• A pass at the bar is noise; fix the state, not the bar. T-STORY's charter case sat at 0.84–0.89 until the extractor withheld the diagram labels that supplied the insight its comment lacked and gave closed-list flags instead: 0.99–1.00 (b3db6917). T-TAKEAWAY ranged 0.41–0.61 across 8 runs until its state stopped showing the slide body, whose lines stated the claim the label lacked: 0.93–0.98. --wire adds a third invocation within 0.03 of a bar.
• Jev is the backend. strands-decider v19 separated 0 of 18 rules (p stayed in 0.17–0.83; margin +0.105 against Jev's +0.564).

Red flags — STOP

About to	Why wrong	Do instead
About to edit a twin after seeing its score	the calibration then measures your edit, not the rule	add a corrected vio2/sat2 beside it; the old twin stays
About to relabel an accepted real case violating	the user accepted it; only they rule it wrong	ask; record the ruling in the manifest source and the docstring, then add a cured twin
About to hand the extractor's whole file to the state	the judge guesses, as on 9 of 10 ds rules	extract the spans and the count the proposition reads
About to move a module out of uncalibrated/ by hand	it skips the two-invocation gate	new-rule.ts --wire
About to change blockAt or the manifest criterion for one rule	every rule's evidence was measured against 0.85 / 0.5	fix the question or the extractor, or park
About to try a fourth wording for a rule that failed two rounds on accepted work	T-ECHO never moved; T-TRANSITION wired only after the user ruled the bridge is the first and last sentence across the break	ask the user which is right, or park with the numbers in the docstring