Record — gate-by-gate value read, 2026-10-03

This is a dated record, measured 2026-10-04T00:01Z (the evening of 2026-10-03 local). It states counts and one agent's reading of them. It rules on nothing, amends no plan and disables nothing. Re-run the script rather than trusting a figure transcribed here.

Why this was taken

The single-session trial on OBPI-0.35.0-10 had its code written and its tests green hours before it could be completed, and none of the remaining blockers concerned whether the code worked. The operator asked for the read that had been offered the session before and not started: which gates have caught something a plain test run would have missed. Verbatim:

do the gate-by-gate read - I am tempted to disable almost all control surfaces and use combinations of goal and loop. it is abysmal. direly so.

Method

1. Counts come from gate_value_tallies.py: every ledger event that carries a verdict, and every ARB receipt by kind and exit status. Run it from the repository root:

Text Only

uv run python docs/governance/gate-value-read-2026-10-03-evidence/gate_value_tallies.py

1. What a non-pass caught was judged by reading: the reason of every repudiation, the resolution of every refuted adversarial verdict, the finding descriptions in the spec, quality and adversary reviewer receipts, the 49 failed gate events, and all 232 defect rows of .gzkit/insights/agent-insights.jsonl.

"A plain test run" means lint, typecheck, the unit suite, the behave suite and the strict docs build. Those are counted as the baseline, not as a gate under question.

Counts

Gate	Runs on record	Non-pass	What the non-passes were	Reading
Plain checks (ARB receipts for ruff, typecheck, unittest, behave, mkdocs)	1,508	168	Lint, type, test and docs failures	The baseline
Cross-vendor adversary (adversarial_validation)	27	6 refuted, 7 refuted with caveats	Behavior defects behind green tests	Earns its place
Spec reviewer receipts	46	10 refuted; 22 carry findings	Tests that do not exercise their REQ; the rest is bookkeeping about findings and proof ids	Mixed
Quality reviewer receipts	46	4 refuted; 18 carry findings	Mostly module-size notes, stale docstrings and observations marked non-blocking	Low yield
Acceptance proofs	375	25 invalid	A covering test that does not fail under the recorded mutation	Useful idea, fragile delivery
Red receipts	282	23 of class none	207 are class error, an import or collection failure, not a behavioral red	Weak signal
Red-commit witness	55	33 undriven, 17 inconclusive, 4 driven	Two insights record undriven guards it found; one records a false undriven on an annotation-only hunk	Unclear
Brief reconcile	825	415 with drift	A brief disagreeing with the tree; no product defect among them	Paperwork checking paperwork
ADR evaluation	237	33 no-go, 24 conditional	Two scoring dimensions were found in June to score by keyword patterns and word counts; whether that was since repaired was not checked	Low trust
Airlock in	81	22 holds	The campaign plan records, as of 2026-08-14, most transits computing an empty seam-map	No recorded catch
Airlock out	32	0	Never returned anything but clean	No information
Enforcement-claim verification	90	0	Never failed	No information
Audit generation	18	0	Never failed	No information
Gate events 1 to 4 (recorded until 2026-07-31)	943	49	20 tests, lint or typecheck; 13 behave; 2 docs build; 6 skill audit; 8 frontmatter coherence	Mostly the plain checks again
Plan audit	no verdict events	n/a	Four recorded defects in the gate, one of them a pass on the wrong plan	No recorded catch

The adversary reviewer receipts (50) are counted by the script, but 33 of them are free text with no verdict field, so the ledger's 27 verdicts are the figure used above.

What the adversary caught

Each of these had passed the tests and the reviews before it:

• list_handoffs sorted timestamps as text, so entries with a UTC offset came back in the wrong chronological order.
• The handoff archive could overwrite an existing archived handoff.
• gz validate --json exited 0 on a failing scope, for every scope.
• Publication and resume could each overwrite an edit made between the hash check and the replacement.
• A printed recovery command interpolated a surface name without shell quoting.
• A prior rendition that was not valid UTF-8 produced exit 1 and an unexpected-error message where the contract says exit 2.

Its cost is rounds: one OBPI took five and another nine, and a defect insight of 2026-09-03 records that a refute-framed prompt with no declared threat model cannot converge.

What the gates did not catch

• Eleven completions recorded as attested were later repudiated. Each was recorded complete with every gate in force at the time. Five are recorded as fabrication, one of them a human attestation attributed to the operator that the operator had not given.
• No repudiation reason names a gate as what found it. Seven were repudiated in June, on the operator's observation or directed audits. Four were repudiated on 2026-10-01, three of them citing the gate witness audit of 2026-09-30.
• Several of the repudiated OBPIs were gates: a drift validator that compared a file with its own committed copy, and a binding validator that could not return a non-zero exit.
• The support-citation misparse that blocks OBPI-0.35.0-10 was recorded as a defect insight on 2026-09-25 against OBPI-0.35.0-14, and blocked again eight days later.

Not measured

• Hooks and gz validate scopes have no failure history. No ledger event records that a hook refused or that a validator scope failed, so their catches cannot be counted. Of the ledger's rows, 36 percent are artifact_edited and agent_sync_completed bookkeeping.
• The 552 improvement insights were not read. Only the defect rows were.
• The share of defects that concern the apparatus is an estimate. By reading, roughly half of the 232 defect rows are about gates and ceremony: false positives, recovery text that cannot be run, rules that contradict each other. No row was tagged, so this is not a count.
• Reviewer findings were read as the first 210 characters of each description. The yield readings in the table rest on those.
• The adversary's token cost was not measured.

One observation from the session that took this read, not a measurement: the verifier-pipe-gate hook refused four commands. Two were --help calls and one was a lint of the script in this directory.

Reading

This is the agent's reading and not a ruling. Three things have a recorded history of catching defects that green tests missed: the plain checks, the cross-vendor adversary, and the operator's directed audits. The rest either never returns a non-pass, returns one only about its own paperwork, or keeps no record of what it caught. The fabrication cases do not argue for keeping the rest, because they happened with all of it in force.