H-AIF Gates G1–G9

These gates produce receipts. They map onto the existing P0..P8 ladder in CLAUDE.md; they never redefine it. The mapping lives in ledgers/HIERARCHICAL-AIF-GATE-TO-EXISTING-P-LADDER-MAP.md.

Gate detail

H-AIF-G1 — repo + frozen evidence integrity → P0

Four sub-checks: live remote HEAD matches local (git ls-remote, not a local cache); working tree pristine; audits/phase-c + audits/phase-d byte-identical to the Phase-1 anchor 4fcba6c; hierarchical-aif/ separate. Baseline: 250 files in reports/frozen-evidence-baseline.sha256. Re-run this at the start of every session.

H-AIF-G2 — correction notice + defect ledger

Correct the record before fixing or rerunning. Letting a known-invalid claim stand unqualified is claim laundering. Receipts: reports/CORRECTION-NOTICE-TO-REVIEWER.md, reports/CORRECTION-NOTICE-ISSUED.md, ledgers/HIERARCHICAL-AIF-DEFECT-LEDGER.md.

H-AIF-G3 — failing tests before fixes → P1

Reproducer tests precede repair. Where strict red-then-green was not achieved, say so explicitly rather than presenting it as compliance (see reports/H-AIF-G4-RUNNER-FIX-REPORT.md §5).

H-AIF-G4 — runner fixes + resource reclassification → P1, P0

Fixes live in hierarchical-aif/src/; the frozen runners keep their defects so historical results stay reproducible. Receipts: reports/H-AIF-G4-RUNNER-FIX-REPORT.md, reports/RESOURCE-BOUND-RECLASSIFICATION.md.

H-AIF-G5 — corrected full B4 reruns → P3/P6 per cell

Run order chosen by epistemic value, not cost. Each cell needs a prediction record committed before launch, full frozen N (or a prospective reduced protocol written first), corrected artifact names that never overwrite originals, and a report with hashes.

Order: C02 (highest epistemic value) → C10 → C11 (replaces the withdrawn U4) → C01.

H-AIF-G6 — gap audit + isolated implementation → P1, P6 gap

The audit found F, G, policy layer, priors E, precision Π, and up/down messages absent from the science pipeline; all nine B3 models are MLE density fits. The F-side model is built under the scope ruling, and has since been scored under H-AIF-G7 — verdict NOT_ESTABLISHED.

H-AIF-G7 — control + adversarial comparison → P3, P6

The F-side candidate must be scored on the same frozen split and scoring rule against CONTROL_CURRENT and the adversarial baselines, with a CI-bound verdict. Executed 2026-07-22 — verdict NOT_ESTABLISHED; reproduces the frozen M7 to 2.5e-7 nats.

H-AIF-G8 — raw archive / transfer → P2, P4

Raw MAT archive absent → NOT_LOCATED_RAW_ARCHIVE. No independent dataset → P4 and P7 remain NOT_ESTABLISHED. Never infer raw confirmation from the packaged event JSON.

H-AIF-G9 — ledgers, ladder, docs, handoff

Produces no new evidence by itself. Records mapped status changes only.

Gate rules

1. A gate result may update a P-level only if: the existing P-level definition is named; the source artifact is named; the claim scope is named; the falsifier is carried; partial/not-run/negative states are preserved; and the update does not redefine the P-level.
2. Gates may lower a level on receipts. Withdrawing C11 U4 lowered P6 for that scope. That is the gate working.
3. No gate may move P8. P8 is conjunctive and currently FULL_PARITY = false, with P4 transfer the first unsatisfied level.