Report 011: Claude Opus 5.5, the day after its release, three skills

• One case per skill. Each receipt carries a suite of 1, 1, 1 case, in the order code-review-and-quality, git-workflow-and-versioning, documentation-and-adrs, the cases Report 009 used. Nothing on this page describes how any of the three skills behaves on any other task, or how Claude Opus 5.5 behaves in general.
• Few draws. Each arm drew between 3 and 10 generations, each judged 3 times, and each case’s band is the spread across its draws. With one case there is no spread across cases, and each receipt records its combined uncertainty on the lift as absent, single_case.
• Not Report 009’s harness, for either arm. Report 009 ran on Claude Code 2.1.272. That version cannot call this model: an earlier start of this run stopped at its first call with the API’s answer version 2.1.280 or newer is required, and produced no receipt. Both arms of this run therefore used Claude Code 2.1.280, and each receipt’s sidecar records the version every one of its calls reported (see Run record). The first table holds that harness fixed and moves the model. It is not Report 009’s harness, so neither table puts a Claude Opus 5.5 figure beside a Report 009 figure.
• The harness comparison moves more than Claude Code. The second table sets Report 009’s Claude Opus 5 receipts beside this run’s Claude Opus 5 receipts. Between them the Claude Code version changed, and so did the host: Report 009’s Claude Code binary was the darwin-arm64 build, and this run’s the linux-x64 build. So did the date, and every generation is a fresh draw. The runner, the judge, the grading template, the SKILL.md bytes, the cases, the draw rule and the caps are the same (see Setup). A separation in that table could not be put down to the Claude Code version alone.
• The judge is one of the two models compared. Every draw in both tables was judged by claude-opus-5, Report 009’s judge, which is also the model of the older arm. This report does not measure whether that judge scores its own model’s text differently from another model’s. It is also not the judge the judge policy fixes for Driftproof’s reports, the departure Report 009 made, so these figures are comparable with Report 009 and not with Reports 001 to 008.
• No reasoning effort was set. None of the 204 raw call records’ argv carries an effort setting, the guard adds none (guard.py), and no receipt records one (none recorded). Whatever each model’s default was on this surface applied. The tables read as the same surface with a different model, not as the same effort.
• A verdict reads the with-skill arms only. The runner’s comparison of two receipts sets one receipt’s with-skill band against the other’s, per case. The baseline and lift columns are context: no verdict on this page is read from them, however far apart they are.
• Verification levels. All six receipts of this run are TESTED, TESTED, TESTED, TESTED, TESTED, TESTED, each answered by a model the surface attested (answered_by.kind model, model, model, model, model, model).

Read from: docs/reports/011/evidence/code-review-and-quality-claude-opus-5-5-2026-09-23.json; docs/reports/011/evidence/git-workflow-and-versioning-claude-opus-5-5-2026-09-23.json; docs/reports/011/evidence/documentation-and-adrs-claude-opus-5-5-2026-09-23.json; docs/reports/011/evidence/code-review-and-quality-claude-opus-5-2026-09-23.json; docs/reports/011/evidence/git-workflow-and-versioning-claude-opus-5-2026-09-23.json; docs/reports/011/evidence/documentation-and-adrs-claude-opus-5-2026-09-23.json; three-skill-comparison/version-preflight.json (in Report 009’s bundle, not published; its sha256 is in docs/reports/009/evidence/three-skill-comparison--SHA256SUMS); docs/reports/011/evidence/run-20260923T062806Z--run-record.json; docs/reports/011/evidence/code-review-and-quality-claude-opus-5-5-2026-09-23.surface.json; driftproof-calls/001/command.json (not published: a raw call record; its sha256 is in docs/reports/011/evidence/run-20260923T062806Z--driftproof-calls.SHA256SUMS); docs/reports/011/evidence/run-20260923T062806Z--driftproof-calls.SHA256SUMS; docs/reports/011/evidence/guard.py.

The inputs are Report 009’s. The same SKILL.md bytes (skill.content_hash, the same in this run’s two receipts and Report 009’s for each skill: code-review-and-quality 13d360d7f786, git-workflow-and-versioning 91c8c72654ee, documentation-and-adrs b67a9f07ed10) and the same suites (suite.suite_hash: code-review-and-quality 5d729f885294, git-workflow-and-versioning 4e74150753d0, documentation-and-adrs 6cce54e7d9d0).

The runner is Report 009’s. Driftproof runner 0.10.1 on the claude-cli surface, with the flags Report 009’s run used: at most 80 calls and 8.00 dollars of estimated spend per skill run. It puts the SKILL.md text in the prompt for the with-skill arm and gives the baseline arm the task alone, with no tools in either. It draws generations per arm until the spread settles or a maximum is reached, and the judge scores each draw against the case’s rubric on a continuous 0 to 1 scale.

The judge is Report 009’s. claude-opus-5, with the grading template 82586d1e44f8 in all three.

The rule. A case’s band is its mean across draws plus or minus the sample standard deviation across draws: a descriptive spread with no coverage probability. Two with-skill bands separate under the rule when they do not overlap and their means differ by at least the 0.05 effect floor. A case that does not separate is read under spec 035’s rule: when the two arms’ spreads and draws could not have resolved a shift of the floor’s size, it reads not enough draws to conclude at this effect floor, and otherwise no separation detected. Neither is evidence that nothing differs. The verdicts are the runner’s: driftproof diff over each pair, whose outputs are published below.

Read from: docs/reports/011/evidence/code-review-and-quality-claude-opus-5-5-2026-09-23.json; docs/reports/011/evidence/code-review-and-quality-claude-opus-5-2026-09-23.json; docs/reports/009/evidence/three-skill-comparison--driftproof--code-review-and-quality--receipts--code-review-and-quality-claude-opus-5-2026-09-15.json; docs/reports/011/evidence/git-workflow-and-versioning-claude-opus-5-5-2026-09-23.json; docs/reports/011/evidence/git-workflow-and-versioning-claude-opus-5-2026-09-23.json; docs/reports/009/evidence/three-skill-comparison--driftproof--git-workflow-and-versioning--receipts--git-workflow-and-versioning-claude-opus-5-2026-09-15.json; docs/reports/011/evidence/documentation-and-adrs-claude-opus-5-5-2026-09-23.json; docs/reports/011/evidence/documentation-and-adrs-claude-opus-5-2026-09-23.json; docs/reports/009/evidence/three-skill-comparison--driftproof--documentation-and-adrs--receipts--documentation-and-adrs-claude-opus-5-2026-09-15.json; docs/reports/011/evidence/code-review-and-quality-claude-opus-5-5-2026-09-23.stdout.txt; config.js.

Three figures on this run read low, and a low figure can be an instrument defect rather than an answer: a truncated generation, a refusal, a tool error, a timeout or a judge reply that did not parse. Report 007’s timeout defect is the precedent. Each draw behind these three figures was read, in the receipt and in the raw call record, before any verdict was written. None is an artefact, and no cell is excluded. The review is published as run-20260923T062806Z--artefact-review-20260923.md.

• Claude Opus 5.5 without the skill on git-workflow-and-versioning, 0.300. Draws that ended their turn, untruncated and measured: three of three. Every judge sample reads 0.3. Each draw is a complete commit message whose subject line begins Fix password-reset links with no conventional type prefix, and the rubric reads if the subject line does not use a conventional type prefix, cap at 0.3. The judge’s reason: Body strongly explains cause and user-facing impact, and subject is specific, but subject lacks the conventional `fix:` type prefix, triggering the 0.3 cap. A real answer, judged low.
• Claude Opus 5.5 without the skill on code-review-and-quality, 0.680. Draws that ended their turn, untruncated and measured: four of four, with draw means 0.860, 0.483, 0.500, 0.877. The two low draws are complete reviews that label the leaked key and the logged customer data Critical and put the other findings under category headings, Correctness among them, with no severity label; the judge says so, for example are categories, not severity labels. A real answer, judged low, with the spread between draws.
• Claude Opus 5.5 with the skill on documentation-and-adrs, 0.661. Draws that ended their turn, untruncated and measured: six of six, with draw means 0.483, 0.800, 0.600, 0.853, 0.697, 0.533. Each is a complete revised function. The low draws keep the expired-coupon TODO reworded rather than resolving or removing it, which the judge reads as a miss, for example left a reworded but still-bare TODO. A real answer, judged low.

One more draw, outside the three. In Claude Opus 5’s baseline on git-workflow-and-versioning, draw three of ten scored 0.00. The surface gives the model no tools, and the model wrote a Glob invocation out as text, followed by the words No files found., and gave no commit message. The judge scored that absence. It ended its turn, was not truncated, ran no tool and parsed, so it is the model’s own output and it is not excluded. It sits in a baseline arm, which no verdict here reads.

Read from: docs/reports/011/evidence/git-workflow-and-versioning-claude-opus-5-5-2026-09-23.json; driftproof-calls/041/stdout.jsonl (not published: a raw call record; its sha256 is in docs/reports/011/evidence/run-20260923T062806Z--driftproof-calls.SHA256SUMS); driftproof-calls/042/prompt.txt (not published: a raw call record; its sha256 is in docs/reports/011/evidence/run-20260923T062806Z--driftproof-calls.SHA256SUMS); docs/reports/011/evidence/code-review-and-quality-claude-opus-5-5-2026-09-23.json; driftproof-calls/017/stdout.jsonl (not published: a raw call record; its sha256 is in docs/reports/011/evidence/run-20260923T062806Z--driftproof-calls.SHA256SUMS); driftproof-calls/019/stdout.jsonl (not published: a raw call record; its sha256 is in docs/reports/011/evidence/run-20260923T062806Z--driftproof-calls.SHA256SUMS); docs/reports/011/evidence/documentation-and-adrs-claude-opus-5-5-2026-09-23.json; driftproof-calls/056/stdout.jsonl (not published: a raw call record; its sha256 is in docs/reports/011/evidence/run-20260923T062806Z--driftproof-calls.SHA256SUMS); docs/reports/011/evidence/git-workflow-and-versioning-claude-opus-5-2026-09-23.json; driftproof-calls/133/stdout.jsonl (not published: a raw call record; its sha256 is in docs/reports/011/evidence/run-20260923T062806Z--driftproof-calls.SHA256SUMS).

The run. Run 20260923T062806Z, one skill at a time, the Claude Opus 5.5 arm first and then the Claude Opus 5 arm, code-review-and-quality first in each. The command file writes a status line after each skill run; there are six, each with exit status 0 and one receipt, the last at 2026-09-23T07:08:39Z. Each receipt’s run.date_utc is written by the runner and is not read here as a start or a finish.

The caps. Per skill run, the runner’s own caps, as in Report 009 (see Setup). Across the whole run, a guard in front of Claude Code allows 240 calls in all and 240 seconds per call. The run used 204 calls, and the six receipts record 0 unmeasured draws.

The harness. Claude Code 2.1.280, installed at the npm integrity the run record states, with runner 0.10.1 at Report 009’s integrity. The runner writes no Claude Code version into a receipt, and a receipt is sealed, so each receipt has a sidecar, <receipt>.surface.json, bound to it by its receipt_hash and recording the version each of its calls reported; the table gives it.

The price. The runner’s registry had no row for claude-opus-5-5, so this run used a copy of it with one row added, at 4 and 20 dollars per million input and output tokens, read from the vendor’s pricing page on the day of the run (docs-pricing-snapshot-2026-09-23.json). The row is this run’s only; the product’s registry does not carry it.

The cost. Metered spend was 0.00 dollars: the claude-cli surface runs on a subscription. The estimated API-equivalent below is every draw’s generation and judge token counts, as each receipt records them, at the prices each receipt froze, with cached input counted at the full input price, as the runner counts it.

Integrity. Receipt hashes: claude-opus-5-5 code-review-and-quality 5c3ad19b40dd4b39, claude-opus-5-5 git-workflow-and-versioning ca69c3436d0dbfa5, claude-opus-5-5 documentation-and-adrs b924a435c70e0092, claude-opus-5 code-review-and-quality d7bd2fcd3dbea881, claude-opus-5 git-workflow-and-versioning 2ba4ef175d7f1caf, claude-opus-5 documentation-and-adrs 34107e18c90f3b20. Each validates with its receipt hash verified. The raw call records carry the SKILL.md text and stay unpublished; their sha256 are published.

Read from: docs/reports/011/evidence/run-20260923T062806Z--run-record.json; docs/reports/011/evidence/run-20260923T062806Z--status.jsonl; docs/reports/011/evidence/guard.py; docs/reports/011/evidence/code-review-and-quality-claude-opus-5-5-2026-09-23.json; docs/reports/011/evidence/git-workflow-and-versioning-claude-opus-5-5-2026-09-23.json; docs/reports/011/evidence/documentation-and-adrs-claude-opus-5-5-2026-09-23.json; docs/reports/011/evidence/code-review-and-quality-claude-opus-5-2026-09-23.json; docs/reports/011/evidence/git-workflow-and-versioning-claude-opus-5-2026-09-23.json; docs/reports/011/evidence/documentation-and-adrs-claude-opus-5-2026-09-23.json; docs/reports/011/evidence/code-review-and-quality-claude-opus-5-5-2026-09-23.surface.json; docs/reports/011/evidence/code-review-and-quality-claude-opus-5-5-2026-09-23.stdout.txt; docs/reports/011/evidence/git-workflow-and-versioning-claude-opus-5-5-2026-09-23.surface.json; docs/reports/011/evidence/documentation-and-adrs-claude-opus-5-5-2026-09-23.surface.json; docs/reports/011/evidence/code-review-and-quality-claude-opus-5-2026-09-23.surface.json; docs/reports/011/evidence/git-workflow-and-versioning-claude-opus-5-2026-09-23.surface.json; docs/reports/011/evidence/documentation-and-adrs-claude-opus-5-2026-09-23.surface.json; docs/reports/011/evidence/run-20260923T062806Z--driftproof-calls.SHA256SUMS.

Amendment 1, 2026-09-25: the answers graded again, with Claude Opus 5.5 as the judge

This amendment adds to the record above and changes none of its sentences, figures or tables. The limits above say every draw was judged by Claude Opus 5, the model of the older arm, and that this report does not measure whether that judge scores its own model’s text differently. Spec 044 graded the same 51 answers again, twice, and regenerated none of them: once with claude-opus-5 as the judge, as a control, and once with claude-opus-5-5. Both ran through the same command on the same runner, 0.11.2, so the judge model is the one difference between them. Each answer was judged 3 times, as in the published run, whose runner was 0.10.1. Spec 044 fixed its protocol before any call, including that the result would be published here whichever way it went.

No verdict changed. In each of the three cells, the label with Claude Opus 5.5 as the judge is the label under the control. Each cell below is the Claude Opus 5.5 arm’s with-skill mean minus the Claude Opus 5 arm’s, then the runner’s reading of the two with-skill bands, as in the first table above.

The one label that differs from the published one did not move with the judge. On git-workflow-and-versioning the control reads no separation detected, where this report published no separation detected; not enough draws to conclude at this effect floor. Claude Opus 5 judged both, so the judge model did not move it. Spec 044 names such a gap the code and re-sampling effect: the newer runner and a fresh set of judge samples. It does not split it further. With Claude Opus 5.5 as the judge the cell reads as it does under the control.

Answer by answer. An answer’s score is its judge’s mean over the samples, and its pass line is its case’s threshold, at or above counting as a pass. Of the 51 answers, 50 sat on the same side of the pass line with Claude Opus 5.5 as the judge as under the control, and the mean absolute difference between the two scores was 0.029. For scale, the control against the published grades, the same judge model grading twice: 50 of 51 on the same side, and a mean absolute difference of 0.013.

The one answer that crossed is one of Claude Opus 5’s three answers without the skill on code-review-and-quality. The control scored it 0.827 and Claude Opus 5.5 as the judge 0.633, a difference of -0.193, which takes it from above its pass line of 0.7 to below it. It sits in a baseline arm, which no verdict reads.

Did Claude Opus 5.5 favour its own model’s answers? Not on these answers. Set against the control answer by answer, its scores moved by -0.020 on average over Claude Opus 5.5’s 22 answers, and by -0.012 over Claude Opus 5’s 29. It scored both models’ answers lower than the control did, and its own model’s by more. On documentation-and-adrs, where the two models are furthest apart, the newer model’s with-skill mean minus the older’s is -0.144 with Claude Opus 5.5 judging and -0.117 under the control. This reads what happened on these answers. It is not a measurement of how Claude Opus 5.5 judges its own text in general.

The grading template. Every receipt, the published ones and the re-grades, records the grading template 82586d1e44f8, so the re-grades used the template the published grades used.

The cost. The re-grades made 306 judge calls. At the runner’s estimated API-equivalent, every input token at the input price and the prices of the registry the run used, they cost 11.29 dollars: 7.12 for the control and 4.18 for Claude Opus 5.5, under a cap of 15 dollars set before the run.

The limits above still hold. One case per skill: each receipt’s suite holds one case, so nothing here describes how either model judges or is judged on any other task. Few draws: each arm has between 3 and 10 answers. And the judge question now points the other way: Claude Opus 5.5 is the model of the newer arm, so this amendment reads the newer model judging its own answers beside the older one’s, where the report above reads the older model judging its own. Each re-grade is a receipt of verification level TESTED.

The files. Spec 044’s re-judge receipts, each with its .regrade.json sidecar naming the published receipt it re-grades and its summary; the driftproof diff outputs the labels are read from; and the run’s cells.json, run-record.json, status.jsonl and projection.json. Each is a byte-identical copy of its twin under specs/044-report-011-regrade/evidence/regrade-20260924T070018Z/, which is not published. The answers file the re-grade read is not published either; the run record gives its sha256. Validate a receipt with npx driftproof validate <file>.

• docs/reports/011/amendment-1/code-review-and-quality-claude-opus-5-5-regrade-claude-opus-5-2026-09-24.json
  sha256 df043010694e7ae41fcbf0f8ef225e17cb09b28c669d06d15e65f486b7732a5f
• docs/reports/011/amendment-1/code-review-and-quality-claude-opus-5-5-regrade-claude-opus-5-2026-09-24.regrade.json
  sha256 8d96127bc842729586e3e067a59fb5141b422558b9676358a2d94748f894f405
• docs/reports/011/amendment-1/code-review-and-quality-claude-opus-5-5-regrade-claude-opus-5-2026-09-24.summary.md
  sha256 fc1611be69e08672a22be2084ee266f84e74b72496b2f96abbb306eff8371a68
• docs/reports/011/amendment-1/git-workflow-and-versioning-claude-opus-5-5-regrade-claude-opus-5-2026-09-24.json
  sha256 f32c7202091c7bdf605d7943f70cf19a823e875e2617ac8310a91e1493915689
• docs/reports/011/amendment-1/git-workflow-and-versioning-claude-opus-5-5-regrade-claude-opus-5-2026-09-24.regrade.json
  sha256 ea73e8595a448896228af2922e9f82b40b7c0f0ed7677e2d248882138d4d3d13
• docs/reports/011/amendment-1/git-workflow-and-versioning-claude-opus-5-5-regrade-claude-opus-5-2026-09-24.summary.md
  sha256 d9dca9d25e5cd80612425d2bc977f2b2b0ec227b00723e827afde544a5688372
• docs/reports/011/amendment-1/documentation-and-adrs-claude-opus-5-5-regrade-claude-opus-5-2026-09-24.json
  sha256 bef648903183098741988a1b618aea6d2267ec952ba77e45e209792e86bdea4d
• docs/reports/011/amendment-1/documentation-and-adrs-claude-opus-5-5-regrade-claude-opus-5-2026-09-24.regrade.json
  sha256 ce3eaf18dda08d7cb57d55234e84e60d13ac8d57151d061a1b39900f0350299d
• docs/reports/011/amendment-1/documentation-and-adrs-claude-opus-5-5-regrade-claude-opus-5-2026-09-24.summary.md
  sha256 45338908cf5c90e84506dcd3202306adfa0be0cbb82f949737c92a6aae3b1f36
• docs/reports/011/amendment-1/code-review-and-quality-claude-opus-5-regrade-claude-opus-5-2026-09-24.json
  sha256 d6969013b24472fd0863e3f11393ba118a762acdee53994237ae95d7713f9d38
• docs/reports/011/amendment-1/code-review-and-quality-claude-opus-5-regrade-claude-opus-5-2026-09-24.regrade.json
  sha256 9de51bacf54ae7bb0d78dfbe979beafbe6c4b856fc0b6f020d724c64a6015a4e
• docs/reports/011/amendment-1/code-review-and-quality-claude-opus-5-regrade-claude-opus-5-2026-09-24.summary.md
  sha256 16575a4f571e408efcfa876a0afe698101aa3fb7b675eca1cd42a1b94d823555
• docs/reports/011/amendment-1/git-workflow-and-versioning-claude-opus-5-regrade-claude-opus-5-2026-09-24.json
  sha256 0d34ea176f1cd87ee8cda5c304bbb153b8e1cb82702c21462188644c38b28426
• docs/reports/011/amendment-1/git-workflow-and-versioning-claude-opus-5-regrade-claude-opus-5-2026-09-24.regrade.json
  sha256 815bee07e130d69f3f6c2fde802d1c3dfb5e2acf987ac99e802921e949082768
• docs/reports/011/amendment-1/git-workflow-and-versioning-claude-opus-5-regrade-claude-opus-5-2026-09-24.summary.md
  sha256 9028b388c3fa5e4466f632b18c20acff9a10ddcc6eadb5a74c1f250ab532a5ff
• docs/reports/011/amendment-1/documentation-and-adrs-claude-opus-5-regrade-claude-opus-5-2026-09-24.json
  sha256 2c2bb9c828c524b2a9f1823283d9187617094b3c47fb60873698f06ad989c723
• docs/reports/011/amendment-1/documentation-and-adrs-claude-opus-5-regrade-claude-opus-5-2026-09-24.regrade.json
  sha256 91ab761affe5930392e749de148c7aca9b9d2ea593cb5746cf2375964ac9e127
• docs/reports/011/amendment-1/documentation-and-adrs-claude-opus-5-regrade-claude-opus-5-2026-09-24.summary.md
  sha256 39714137ef0f4521af8dfc44729b7133c3c377c4c2970e3ab9cdc9c04f807a1c
• docs/reports/011/amendment-1/code-review-and-quality-claude-opus-5-5-regrade-claude-opus-5-5-2026-09-24.json
  sha256 387544ed43c00f0f1df33da5a610516dccda1a781df499c54aec31054c043eb4
• docs/reports/011/amendment-1/code-review-and-quality-claude-opus-5-5-regrade-claude-opus-5-5-2026-09-24.regrade.json
  sha256 d4e49a1f2000496e0cb042027db7f97356c909bda8204cbc90695893abae21f2
• docs/reports/011/amendment-1/code-review-and-quality-claude-opus-5-5-regrade-claude-opus-5-5-2026-09-24.summary.md
  sha256 49cae3cbdfdc11f485d50af8f4d047e3dc8929439412079412b29804071b6a87
• docs/reports/011/amendment-1/git-workflow-and-versioning-claude-opus-5-5-regrade-claude-opus-5-5-2026-09-24.json
  sha256 57bb02383afd57590d87472969d88ddedbbc87cb563d3b38d76a972d952596a9
• docs/reports/011/amendment-1/git-workflow-and-versioning-claude-opus-5-5-regrade-claude-opus-5-5-2026-09-24.regrade.json
  sha256 c09f2175064d1e5d3f38240bbd871cacb28753d45fb8a434bcef5993b5207486
• docs/reports/011/amendment-1/git-workflow-and-versioning-claude-opus-5-5-regrade-claude-opus-5-5-2026-09-24.summary.md
  sha256 ac4f79a7be72a6b18f7d5c21d67c3bc097073096fb8551b2212b3ab209a1a78b
• docs/reports/011/amendment-1/documentation-and-adrs-claude-opus-5-5-regrade-claude-opus-5-5-2026-09-24.json
  sha256 77881d1cbce73d2607f179d56ae0a686090a201c3b4dd1573fcb582d5b7a7a49
• docs/reports/011/amendment-1/documentation-and-adrs-claude-opus-5-5-regrade-claude-opus-5-5-2026-09-24.regrade.json
  sha256 3fc6afda93f8dc75204c517622ad23cff1f74e9c24cfd2659cb899f967b6b08e
• docs/reports/011/amendment-1/documentation-and-adrs-claude-opus-5-5-regrade-claude-opus-5-5-2026-09-24.summary.md
  sha256 6354bf29db06ad1a6fd4d7c677afee90e3c3d9ba286b86ea7f2e4e903eca9e04
• docs/reports/011/amendment-1/code-review-and-quality-claude-opus-5-regrade-claude-opus-5-5-2026-09-24.json
  sha256 700f277853d53954fa2cd4d6a821856c9bc5e79bdb8dd6915b6597bc18a24095
• docs/reports/011/amendment-1/code-review-and-quality-claude-opus-5-regrade-claude-opus-5-5-2026-09-24.regrade.json
  sha256 7548999b166a9ec7af71bd69916179147d507f8a1989e749d8d7ee2dc853ee29
• docs/reports/011/amendment-1/code-review-and-quality-claude-opus-5-regrade-claude-opus-5-5-2026-09-24.summary.md
  sha256 3e636a4334d93962c3499775cb0a62051226252eeb677739c6679c4196ea2718
• docs/reports/011/amendment-1/git-workflow-and-versioning-claude-opus-5-regrade-claude-opus-5-5-2026-09-24.json
  sha256 bff8274ed9630c66a0cce40f6256ca55c1cbf3e8c56b8999c54c509788b935e3
• docs/reports/011/amendment-1/git-workflow-and-versioning-claude-opus-5-regrade-claude-opus-5-5-2026-09-24.regrade.json
  sha256 82886336b2becc99f7a3318f29fbaff5e1a2efb408d4f40a3ff5ec7b527b9503
• docs/reports/011/amendment-1/git-workflow-and-versioning-claude-opus-5-regrade-claude-opus-5-5-2026-09-24.summary.md
  sha256 f1e08d20e9b0b39bb752c1817c734f349d1e36de15bc8ad34137fbb2b52ff130
• docs/reports/011/amendment-1/documentation-and-adrs-claude-opus-5-regrade-claude-opus-5-5-2026-09-24.json
  sha256 6d441a37a68278d7fec8c8ebe0df8b8f3a2bff1ec34b75714e648a7276931c48
• docs/reports/011/amendment-1/documentation-and-adrs-claude-opus-5-regrade-claude-opus-5-5-2026-09-24.regrade.json
  sha256 7ee14d921cf2d53d37c0ee1b65c55858ae40dfd63cb658742334bc8c11c94863
• docs/reports/011/amendment-1/documentation-and-adrs-claude-opus-5-regrade-claude-opus-5-5-2026-09-24.summary.md
  sha256 3f063796004097e9adaf2815833c55190212a63915442bbb2a85c5dce73f87e2
• docs/reports/011/amendment-1/diff--model--code-review-and-quality--control.md
  sha256 590942c832f3222d67cdb268d896f5a8d4a42c240502ebb5412193acea1066f5
• docs/reports/011/amendment-1/diff--model--code-review-and-quality--original.md
  sha256 aeecaa1c4b93470033284e266933f26203b5ccc05349f2631d174b656137270c
• docs/reports/011/amendment-1/diff--model--code-review-and-quality--treatment.md
  sha256 b76a0a5e2880a6e097e93c90c2337b2ba16ea4075f5896bde5d1c8c65e48fa71
• docs/reports/011/amendment-1/diff--model--documentation-and-adrs--control.md
  sha256 a8510d11e217ac132a2adeffcc9234653dabce90953f214f383dc1921e7cfc77
• docs/reports/011/amendment-1/diff--model--documentation-and-adrs--original.md
  sha256 4c6a96a53f821aa16ba0e4a83bc765994372d002d6a2e6c1f39bbd1e09030eac
• docs/reports/011/amendment-1/diff--model--documentation-and-adrs--treatment.md
  sha256 02b9a619c33bcf174f59b0b5bb1a1457210d83392eb1bb65f17362aa7717f6ea
• docs/reports/011/amendment-1/diff--model--git-workflow-and-versioning--control.md
  sha256 c21978908dcc82f186bf50fc9d88a396dda9732a9393252a2b636a5795c06cdb
• docs/reports/011/amendment-1/diff--model--git-workflow-and-versioning--original.md
  sha256 1abedd3f837988f46b61e37b4108474898720402983e5fee087112efb0d31785
• docs/reports/011/amendment-1/diff--model--git-workflow-and-versioning--treatment.md
  sha256 53335d88436417d3243aa87ff9abb7f40c336486a67e43d48068d793bfc950e3
• docs/reports/011/amendment-1/cells.json
  sha256 5017438d26dd9e819eb1950cdcae42f52093e1ed58cfeedb24b590d77b2b46e4
• docs/reports/011/amendment-1/run-record.json
  sha256 22787089322d4c9ea33e8ce444d5de18d1d47b5a03aaea9cb822de02b2a22af4
• docs/reports/011/amendment-1/status.jsonl
  sha256 14905851df8f825d6279d3c16d8ab57527fbf5ac59cb795ed40f38d7a5c6f14c
• docs/reports/011/amendment-1/projection.json
  sha256 242a1d7693e358028b3affb9d9803b89acd3983f1a0e5cc71573a659cc749f70

Read from: docs/reports/011/amendment-1/cells.json; docs/reports/011/amendment-1/code-review-and-quality-claude-opus-5-5-regrade-claude-opus-5-2026-09-24.json; docs/reports/011/amendment-1/git-workflow-and-versioning-claude-opus-5-5-regrade-claude-opus-5-2026-09-24.json; docs/reports/011/amendment-1/documentation-and-adrs-claude-opus-5-5-regrade-claude-opus-5-2026-09-24.json; docs/reports/011/amendment-1/code-review-and-quality-claude-opus-5-regrade-claude-opus-5-2026-09-24.json; docs/reports/011/amendment-1/git-workflow-and-versioning-claude-opus-5-regrade-claude-opus-5-2026-09-24.json; docs/reports/011/amendment-1/documentation-and-adrs-claude-opus-5-regrade-claude-opus-5-2026-09-24.json; docs/reports/011/amendment-1/code-review-and-quality-claude-opus-5-5-regrade-claude-opus-5-5-2026-09-24.json; docs/reports/011/amendment-1/git-workflow-and-versioning-claude-opus-5-5-regrade-claude-opus-5-5-2026-09-24.json; docs/reports/011/amendment-1/documentation-and-adrs-claude-opus-5-5-regrade-claude-opus-5-5-2026-09-24.json; docs/reports/011/amendment-1/code-review-and-quality-claude-opus-5-regrade-claude-opus-5-5-2026-09-24.json; docs/reports/011/amendment-1/git-workflow-and-versioning-claude-opus-5-regrade-claude-opus-5-5-2026-09-24.json; docs/reports/011/amendment-1/documentation-and-adrs-claude-opus-5-regrade-claude-opus-5-5-2026-09-24.json; docs/reports/011/evidence/code-review-and-quality-claude-opus-5-5-2026-09-23.json; docs/reports/011/evidence/git-workflow-and-versioning-claude-opus-5-5-2026-09-23.json; docs/reports/011/evidence/documentation-and-adrs-claude-opus-5-5-2026-09-23.json; docs/reports/011/evidence/code-review-and-quality-claude-opus-5-2026-09-23.json; docs/reports/011/evidence/git-workflow-and-versioning-claude-opus-5-2026-09-23.json; docs/reports/011/evidence/documentation-and-adrs-claude-opus-5-2026-09-23.json; docs/reports/011/evidence/diff--model--code-review-and-quality.md; docs/reports/011/amendment-1/diff--model--code-review-and-quality--control.md; docs/reports/011/amendment-1/diff--model--code-review-and-quality--treatment.md; docs/reports/011/evidence/diff--model--git-workflow-and-versioning.md; docs/reports/011/amendment-1/diff--model--git-workflow-and-versioning--control.md; docs/reports/011/amendment-1/diff--model--git-workflow-and-versioning--treatment.md; docs/reports/011/evidence/diff--model--documentation-and-adrs.md; docs/reports/011/amendment-1/diff--model--documentation-and-adrs--control.md; docs/reports/011/amendment-1/diff--model--documentation-and-adrs--treatment.md; docs/reports/011/amendment-1/status.jsonl; docs/reports/011/amendment-1/run-record.json.

Amendment 2, 2026-10-10: the title, which said release day

This report was first published as Report 011: Claude Opus 5.5 on release day, three skills. The vendor’s news index lists Claude Opus 5.5 on 22 Sep 2026, and the run every figure on this page is read from, 20260923T062806Z, began on 23 Sep 2026 at 06:28 UTC, the day after. The title now says so: Report 011: Claude Opus 5.5, the day after its release, three skills. It reads so on this page, in its Markdown copy and in the site’s lists of reports.

This amendment changes the title and nothing else: no sentence, figure, table, verdict or file above it, and no receipt or badge. The news index was read on 10 Oct 2026. Its record, with the line quoted and the sha256 of the page, is published beside this page as release-date.json.