Record — gate witness audit, 2026-09-30

This is a dated record, measured 2026-09-30. It states measurements and their methods; it amends no plan, rules on nothing and repudiates nothing. The figures are the record of one pass, not a live value. Re-run the evidence scripts rather than transcribing a number from this file. Evidence and re-run instructions: artifacts/audits/gate-witness-audit-2026-09-30/ (raw reports, scripts and data; see its README.md).

Why this was taken

The v0.34.8 release closed a set of gates that had reported green without holding. The operator asked three questions in the release session:

• Did any earlier completion go through one of those gaps while it was open?
• Why are such gates so pervasive?
• Which greens can be believed?

This record holds the four measurements taken to answer them. Their conclusions are routed to GHI #1154 (test shape), GHI #1155 (enrollment) and the v0.34.8 release notes' Known issues.

Reading this record

Two claims in the same session were stated without measurement and then retracted ("most of the release", and "almost nothing fed the gates a known-bad input"). The measurements below answer different questions over different populations, so their numbers are not comparable with each other:

Measurement	Population	Question
1. Attestation audit	Completions, closeouts and releases inside each gap's open window	Did the gap let something through?
2. Pre-fix test shape	The 12 gates already known to be hollow, selected because they failed	What did their tests feed them?
3. Gate map	The 98 registered enforcement claims at c0aa6467b	Does each control depend on its gate?
4. Registered at parent	The same 12 hollow gates, at each fix commit's parent	Were they under a registered control?

A finding marked verified was re-read against the ledger, receipts, code or git in the main session. Reported means it comes from a delegated agent's report and was not re-checked.

1. Attestation audit (reports auditA–auditD)

Each gap's open window was taken from its fix commit. The population was every completion, closeout or release recorded inside it. Each item was classified EXPLOITED, EXPOSED-CLEAN or UNDETERMINABLE, with evidence.

Verified:

• OBPI-0.0.24-04 (heavy lane) completed with a failed unittest receipt on record: arb-step-unittest-e3b0f66d…, uv run -m unittest -q, exit 1, 18 minutes before completion. Its attestation (ledger line 4524) cites no unittest receipt. (#889)
• OBPI-0.0.22-05: its full-suite receipt f0aba782… shows exit 1, five minutes before completion. The attestation cites a scoped run and mentions "Out-of-scope drift filed as GHI #359". (#889)
• OBPI-0.0.28-03: the key_proof pairs a scoped unittest command with receipt a7295197…, which recorded the full suite (exit 0). (#942)
• Three completions on a standing refutation, each disclosed in the attested record: OBPI-0.35.0-02 (ledger L15504, "The adversary's own check was NOT re-run"), OBPI-0.34.0-03 (L14160, REQ-03/04 tests "still bypassable"), and OBPI-0.34.0-05 (L14463, "Unresolved and disclosed, not waived"). (#960)
• OBPI-0.35.0-09: at completion (2026-08-21T08:23:56Z), the latest red receipt for REQs 01/02/04/05/06/10 was none. The error receipts that now stand were written 10 minutes later. OBPI-0.34.0-02: REQ-02-05 had no red receipt before completion. (#849)
• The #1057 example OBPI-0.0.26-01: commit 32adb990c modified commands/adr_promote.py, events.py and ledger.py, none of them in the brief's Allowed Paths. That allowlist is itself hedged ("or wherever the evaluate command lives").
• mkdocs.yml:229 at d266be9ff^ carried links: not_found: ignore. (#803)
• 57a94bd58, a tests/ commit with no Task: trailer, reached origin through the --reuse-verified pre-push skip. The cause was read in code and fixed under #1017 (b22353885, c0aa6467b).

Reported, not re-checked: #889 284 completions (2 exploited, 56 clean, 226 undeterminable); #942 606 (1 / 1 / 604); #994 2 (0 / 2 / 0); #1093 53 (3 / 48 / 2); #959 7 clean; #960 20 (3 exploited / 7 clean / 10 undeterminable); #985 458 (0 / 1 / 457); #996 8 clean; #1057 307 replayable PASS receipts (56 exploited, 10 more on __init__.py only, 15 undeterminable, 223 clean); #849 22 (2 / 20 / 0); #927 599 fix( commits, all undeterminable; #995 7 clean; #1124 4 clean, 1 undeterminable; #803 every release since v0.3.1, by historical rebuild; #1017 280 trailer-less code commits reached origin.

Unconfirmable: OBPI-0.35.0-05's Demo runs gz content compose, and a composition_candidate_emitted row exists before its completion, but no record ties the row to the Demo. OBPI-0.0.65-03's demo handoff was never committed.

2. Pre-fix test shape (pretest-shape.md, plus three main-session reads)

For each of the 12 hollow gates, the tests at the fix commit's parent were read for the inputs they fed the gate. Main session: #889, #996, #995. Delegated: the other nine, with

960 and #803 spot-checked.

Only #803 was happy-path-only. Eleven had negative tests. What they lacked was an input present but false: in five the negatives were absence-only, in four a test asserted the defect itself, and in three the passing fixture was a stand-in the test declared sufficient ("{}" as a receipt, "Rationale here." as a walkthrough). Per-gate table: GHI #1154.

3. Gate map (gatemap.py, gatemap-final.json)

For each of the 98 registered enforcement claims at c0aa6467b, every guard statement in the gate functions the claim names was replaced with pass, one at a time, in a detached worktree. That claim's own control was then re-run through gzkit.mutation_witness: 662 mutants, 235 killed, 427 survived.

Result	Claims
Control fails when at least one guard is removed	92 (31 catch all removals, 8 at least half, 53 fewer than half)
Control passes with every guard removed	1 (population-controls-disclosed, an admit control whose refuse partner is load-bearing)
Not measured	5 (module-size, tautological-debt: the derived gate target is a helper and the real gate runs behind a subprocess; readiness-audit, tautological-debt-waived: the shared run_command cannot be mutated by the sweep; cli-usage-error-exit-two: no in-process function)

24 of the 98 claims pin the reason they expect (expect); 74 accept any failure. The bar is low: one caught removal counts as load-bearing. The 427 survivors are unreviewed. One read in the smoke run (if not drift: return []) was equivalent.

4. Registered at parent (registered_at_parent.py, .json)

The criterion, fixed before running: at the fix commit's parent, does any registered enforcement claim name, in its source_fn or gate_targets, a src/gzkit function the fix modified? Each parent's registry was dumped from its own code.

9 of 12 had no such claim: #889, #996, #995, #960, #959, #888, #932, #933, #1124. In five of them, a keyword pass found a nearby claim on the same surface naming a different function. #851 (session-green-gate) and #803 (docs-build) were registered. #1007 is a literal hit only because it shares fix commit 1edf9dc1b with #851.

Errata — report claims that did not survive checking

Report	Claim	Finding
auditC.md	Only 2 of the 4 survived hunks in the 73aaab093 receipt are driven by 57a94bd58	False: all four mutants fail that commit's tests, including 127-127, run separately
auditD.md	61eb84194 is a trailer-less code commit	False: it carries Task: TASK-deps-upgrade-ty-0.0.84
predecessor handoff and insight	The undriven guard in 92f64debc is the seen-dedupe	False: the receipt's 618-619 is the event-type filter; removing the dedupe is already killed
drafted release entry (#959)	None of the 13 completed refutations carried a resolution	False: all 13 refuted adversarial_validation rows carry a non-empty resolution
auditC.md	names the plan-audit verb with a hyphen (plan-audit)	No such verb: it is gz plan audit. Corrected in the report before it landed; the claim it carried is unchanged

What this record does not establish

• How many of today's gates (64 gz check steps, 100 gz validate scopes, 12 gz obpi precomplete checks, plus the completion and closeout gates) lack a registered control. That population was not mapped. Its enumeration is GHI #1155's contract.
• What caused the pattern. The 2026-09-24 integrity audit found production and tests in the same commit in 9 of 9 sampled OBPI increments, which is consistent with co-authoring but does not establish it.
• Whether any completion above should be repudiated. That is an operator ruling and has not been made.