Harness improve

Harness-improve: changing the protocol without meta-neuroslop

BLUF: improvements to the harness are the highest-slop-risk artifact in this repo — plausible protocol text detached from incidents is neuroslop about neuroslop. Entry requires a live anchor; every change class has its own gate; a null measurement result is information, not failure.

Stage 0 — anchor or stop

Admissible anchors, in order of strength:

1. A reproduced incident — tool output from this or a documented session.
2. A structured harness-gap report (.github/ISSUE_TEMPLATE/harness-gap.yml).
3. A measurement — an eval run, an A/B result, a counted pattern across sessions.
4. An explicit owner decision.

«Would be nice» without an anchor → record as a ROADMAP row candidate or drop. Do not proceed to implementation on an unanchored idea: critique of the harness demands the same evidence bar as critique of code.

Stage 1 — verify against current main

• Pull main; confirm the gap still exists by reading the current files, not the report. Reports go stale: the 2026-06-11 issue revision found a «missing» acceptance kit that had been merged before the issue was written.
• Check ROADMAP.md and open issues for an existing row or tracker. A recorded decision (ADR, ROADMAP status, protocol doc) must be quoted before re-opening, and re-opening requires new evidence — not repeated association.

Stage 2 — classify the change and pick its gate

Class	Typical diff	Gate before merge
Mechanical	hooks/, scripts/	bash -n + python3 -m json.tool on JSON + the matching scripts/selftest_*.sh (extend it with the new behavior — a test of a hand-copied reimplementation is decorative coverage)
Behavioral text	invariants.txt, agents/, references/, CLAUDE.md	single-line invariant format (selftest_random_invariant.sh); role-subset inheritance ONLY via the eval-gated N8 path (ROADMAP)
Calibration	eval/criteria.md, judge prompts, role subsets, agent system prompts	eval run (label run-eval); reconcile every consumer of changed counts/contracts (ci_eval.py, docs — the 16-vs-17 drift class); expect judge noise — report null results as-is

All classes pass the critic gate at push (auto-critic.sh); protocol paths additionally pass the human-token gate (PR #59). One concern per branch and PR.

Stage 3 — implement the minimal diff

• Evolve existing mechanisms; a new subsystem needs an ADR first.
• ROADMAP bookkeeping follows the file's own convention: new row before work starts, or move to Shipped with the landing PR.
• The change description states its own Counter — the condition under which the new rule does not apply.

Stage 4 — measure and keep the deletion path open

• A calibration claim gets a before/after eval reference; publish the numbers including null (precedent: a candidate sentence measured null at judge-noise α 0.34–0.46 — that outcome was reported, not hidden).
• Every added rule carries a deletion path: rules that stop earning their keep are removed, not accumulated. The harness is not append-only.

Output contract

1. Anchor. 2. Verified gap vs main. 3. Class + gate. 4. Diff. 5. Verification/measurement outputs. 6. ROADMAP/issue bookkeeping. BLUF at every step.

Counter: when the owner explicitly orders an unanchored exploratory change («try it, we'll see»), the anchor requirement converts into a measurement obligation after landing — exploration is allowed; unmeasured persistence is not.

---

Provenance: codifies the pipeline executed manually in the 2026-06-11/12 issue revision (#55/#56, PR #57–#60): verify-against-main caught an already-done item; per-class gates landed across #44, #51–#53, #59.