# Consolidation after the Opus 5.5 critique

GPT-6-Astra at high reasoning produced the initial assessment. A separate Claude Opus 5.5 session at high reasoning critiqued its attribution and judgments, using the original reviews, pinned source and saved probes. The consolidating evaluator checked the proposed changes against the cited evidence. Both evaluator models appear in the comparison, and both saw the model labels. This consultation improves the audit trail; it is not blinded ground truth.

All 200 original reviews remain unchanged. No review was rerun. The exact consultation prompt, command, returned model labels, full local session, verbatim critique and proposals are preserved with the experiment. The public prompt generalizes local access paths; the public critique and proposals preserve the critic's text. The critique describes the report as it existed when the consultation began, before the interim publication disclosed the assessor identities.

## Accepted changes

| Item | Decision and parent check | Effect |
| --- | --- | --- |
| CA-ANSI-ONLY | The original GPT-5.6-Sol r2 Finding 1 explicitly says an escape-only column emits nothing. Apply the same embedded-finding rule already applied to Opus. | 5.6 Sol B +1; runtime +1. |
| ST-ABI | Independently inspected the local 32-bit glibc disassembly: lseek64 uses syscall 140, `_llseek`. The pinned configuration enables 64-bit file offsets while the new test requires literal `lseek`. This establishes the mismatch without a full 32-bit CPython build. | C → B, tests/tooling: 5.6 +1, Astra +1, Opus +4. |
| ST-POSIXSPAWN | Reproduced rejection of the base syscall filter by strace 6.8. Read the pinned test and unchanged subprocess dispatch. A local Python 3.14.7 trace still uses `CLONE_VFORK` via posix_spawn with `_USE_VFORK=False`; the parent mock does not configure the child. The changed filter enables a test that can pass without exercising its intended implementation. | C → B, tests/tooling: Opus +3. |
| ST-VFORK-COVERAGE | Clarified that the removed negative case was nominal coverage on the checked environment, since the old filter skipped it. Retain B for removing an explicitly intended integration check while enabling the suite. | Wording only. |
| DJ-PARTITION | Confirmed the subject excludes partition columns from the JSON object unless `WRITE_PARTITION_COLUMNS` is enabled. Add that prerequisite to the verdict. | Wording only. |
| Tentative attribution | Checked the five cited Uncertain-section mentions. The same passages also contain Opus r4 ST-PARSER and Opus r2 ST-SENTINEL mentions; recorded those too. | Seven added tentative entries; no score effect. |
| Interpretation | Removed the claim that 6 Sol missed the Argon2 residue concern in every run: r4 raised it tentatively, and C is not a penalty. Added the fact that Opus's remaining macro review (r2) incorrectly dismissed the issue. | Clearer calibration; no score effect. |

ST-POSIXSPAWN remains a separate family from ST-VFORK-COVERAGE. Fixing the positive test's child configuration does not restore the removed negative integration case. A coarser grouping is reasonable: merging them would reduce Opus B from 57 to 54 and unique A+B families from 23 to 22, with the other models unchanged. The public tables retain the original two-family structure.

The 32-bit result is a dependency-disassembly and pinned-source argument, not a successfully executed 32-bit CPython test. The strace/posix_spawn checks use the recorded local tool versions and host Python, not a build of the pinned revision. Statements about the base filter are bounded to that checked environment; this consolidation does not assert that every historical strace version behaved identically. The `_llseek` mismatch is sufficient for ST-ABI; the additional `statx` claim and musl behavior are not required for the accepted decision.

## Totals

Counts are asserted family detections across repetitions, not unique bugs. Model order is fixed below. The pre-consultation column is the initial public report, which already included the earlier Opus ANSI correction from B 46 to 50.

| Model | A, unchanged | B before → after | C before → after | D, unchanged | B runtime / support, after | Unique A+B, after |
| --- | --- | --- | --- | --- | --- | --- |
| GPT-5.6-Sol | 16 | 24 → 26 | 22 → 21 | 6 | 12 / 14 | 16 |
| GPT-6-Sol | 17 | 28 → 28 | 12 → 12 | 5 | 14 / 14 | 16 |
| GPT-6-Astra | 18 | 30 → 31 | 6 → 5 | 6 | 17 / 14 | 18 |
| GPT-6.1-Sol | 18 | 32 → 32 | 12 → 12 | 3 | 18 / 14 | 16 |
| Opus 5.5 | 19 | 50 → 57 | 26 → 19 | 2 | 19 / 38 | 23 |

## Judgment calls retained, with sensitivity shown

A remains a landing judgment, not a claim of critical production severity. Whole test-module import failures and invalid generated code can warrant blocking despite narrow triggers. Retain ST-PROBE, RA-MACRO and DJ-PARTITION in A, but show how demoting the first two removes nearly all serious-count differences: 16/16/16/15/16. This sample does not support a stable ordering on serious bugs.

Keep SH-INIT in C because the destructive example supplies invalid arguments and explicitly requests database replacement. A bounded-defect interpretation is reasonable: moving it to B would make runtime B 16/17/17/21/19, putting 6.1 ahead on that measure. That sensitivity is more informative than only promoting it to A.

Keep CA-QUADRATIC in B: the base/subject algorithm change creates repeated prefix work; this is an established complexity regression even though a practical stall needs large input. FE-INPUT-BOUND remains C because it asks for a new input-size policy without the same established complexity regression. Keep SA-TEST-COST in B for repeated production-cost derivations in the default test suite, with its limited CI/resource impact stated. These are debatable boundaries: moving the former to C subtracts one Astra and two 6.1 B detections; moving the latter to C subtracts one 5.6 and four Opus B detections.

Keep FE-FALLBACK in C because SPEC's broad wording and README's narrower address-in-use scenario leave the intended error contract ambiguous. Keep CA-PADDING in D because the checked effect is pre-existing. The report presents alternatives as sensitivity analyses rather than retroactively changing the rubric to produce a preferred ranking.

Family granularity and buckets were adjudicated with model labels visible; they were not a preregistered scoring scheme. Asserted versus tentative follows the report section, not a single hedging word. The consultation checked every family at the evidence/claim level, but did not read every original report line by line or repeat every probe; its coverage table records those limits.

## Conclusion

Opus still shows the broadest confirmed coverage in this sample, primarily in tests, CI and documentation. Runtime counts and serious-count ordering are much closer and sensitive to reasonable classification choices. The separate critique strengthens these limited conclusions without making this a general model leaderboard.

## Publication consistency audit

A separate Opus 5.5 high-reasoning reviewer and a separate GPT-6-Astra high-reasoning reviewer checked the report package for internal consistency without re-evaluating the bug claims. Both independently reconciled the figures, charts, CSV, JSON, historical deltas, scenario calculations and fingerprints. Their instructions, verbatim outputs and resolutions are preserved locally. These are publication checks, not additional observations in the 200-review sample.

Their findings were resolved without changing scores: the A-count explanation now acknowledges 6.1’s single DuckDB miss; shell examples preserve the prompt’s terminal newline; the reference launcher hashes object type and executable bits; and the timeout history distinguishes 192 retained first attempts under the original cutoff from eight unbounded replacements. The initial selection/replacement policy may affect findings, not just cost or latency, and its effect was not measured. Numeric account IDs were removed from the public trace, redaction scope was clarified, and “remaining review (r2)” replaces the ambiguous “fourth review.”

The critic’s wording that A differences come “only” from two families was too strong: the shared core is 16/16/16/15/16. Its exploratory leave-one-case-out numbers use the pre-consultation baseline. Preserve that critique as received; use the consolidated ledger and clearly labeled scenario calculations for current figures. The old CA-ANSI-ONLY verdict’s historical publication-audit sentence is now in the correction narrative; its removal from the verdict changes no mechanism or attribution. DJ-PARTITION remains in A with its trigger narrowed; its optional demotion is not adopted.

One further arithmetic clarification in the preserved critique: its CA-ANSI-ONLY proposal says the 5.6 Sol r2 per-review B cell changes from 3 to 4. The actual CSV and ledgers show 2 to 3. The proposed +1 delta and all model-level totals are correct; the current CSV records 3.