Report 014: Claude Haiku 5.5 on release day, three skills

What did Report 014 find?

What we tested. Claude Haiku 5.5 came out on 7 Oct. We ran it and Claude Haiku 4.5 on three skills, three times each.
What we found. On Claude Haiku 5.5 no skill clearly helped in two or more of three repeats. Documentation clearly helped Claude Haiku 4.5 in two of three; on Haiku 5.5 all three had too few answers to tell, so these runs cannot say whether it still helps. Code review and git workflow did not clearly help either model in two or more repeats.
Repeats matter. Four of six model and task pairs changed reading between repeats.
What it doesn't show. One task per skill and few answers: nothing about other work, and no model ranking.

Report 014: Release drift, claude-haiku-4-5, claude-haiku-5-5

Type Release drift report

What moved Claude Haiku 5.5 on release day, three skills

Models claude-haiku-4-5 claude-haiku-5-5

Headline Claude Haiku 5.5 on the three skills and cases of Reports 009, 011 and 013, beside Claude Haiku 4.5, each run three times.

Receipts 18 linked receipts

Cite this

@techreport{driftproof-014,
  title  = {Report 014: Claude Haiku 5.5 on release day, three skills},
  author = {Driftproof},
  year   = {2026},
  type   = {Release drift report},
  note   = {claude-haiku-4-5, claude-haiku-5-5},
  url    = {https://driftproofhq.com/reports/014/}
}

Release drift report. It measures a new model on the same skills and tasks as earlier reports, beside the model before it, with Report 013’s Claude Sonnet 5.5 and Claude Opus 5.5 receipts as context.

Which skills help Claude Haiku 5.5?

Claude Haiku 5.5 on the three skills and cases of Reports 009, 011 and 013, beside Claude Haiku 4.5, each run three times. On Claude Haiku 5.5 no skill clearly helped in two or more of the three runs. Helped Claude Haiku 4.5; on Claude Haiku 5.5 these runs had too few answers to tell. documentation-and-adrs: clearly helped Claude Haiku 4.5 in two of three runs (one had too few answers to tell), and Claude Haiku 5.5 in zero of three runs (three had too few answers to tell). Did not clearly help either model in two or more of the three runs. code-review-and-quality: clearly helped Claude Haiku 4.5 in one of three runs (two had too few answers to tell), and Claude Haiku 5.5 in one of three runs (two had too few answers to tell). git-workflow-and-versioning: clearly helped Claude Haiku 4.5 in zero of three runs (three showed no clear difference), and Claude Haiku 5.5 in one of three runs (one showed no clear difference, one had too few answers to tell).

A skill counts as helping a model here when it clearly helped it in at least two of the three runs. Every figure below is read from a receipt or a file published beside it, and each block names its files. This report measures what a skill adds to each model on these cases. It makes no claim about any model’s coding ability.

task	model	run 1	run 2	run 3	clearly helped
code review	Claude Haiku 5.5	Clearly helped	Too few answers to tell	Too few answers to tell	one of three
code review	Claude Haiku 4.5	Clearly helped	Too few answers to tell	Too few answers to tell	one of three
code review	Claude Sonnet 5.5 (Report 013)	Too few answers to tell	Too few answers to tell	Too few answers to tell	zero of three
code review	Claude Opus 5.5 (Report 013)	Too few answers to tell	Too few answers to tell	Too few answers to tell	zero of three
git workflow	Claude Haiku 5.5	Clearly helped	No clear difference	Too few answers to tell	one of three
git workflow	Claude Haiku 4.5	No clear difference	No clear difference	No clear difference	zero of three
git workflow	Claude Sonnet 5.5 (Report 013)	No clear difference	No clear difference	No clear difference	zero of three
git workflow	Claude Opus 5.5 (Report 013)	Clearly helped	Clearly helped	Too few answers to tell	two of three
documentation	Claude Haiku 5.5	Too few answers to tell	Too few answers to tell	Too few answers to tell	zero of three
documentation	Claude Haiku 4.5	Clearly helped	Clearly helped	Too few answers to tell	two of three
documentation	Claude Sonnet 5.5 (Report 013)	Clearly helped	Clearly helped	Clearly helped	three of three
documentation	Claude Opus 5.5 (Report 013)	Too few answers to tell	Clearly helped	Clearly helped	two of three

Clearly helped: in that run the spread of the model’s scores with the skill did not overlap its spread without it, and the averages were at least 0.05 apart.

No clear difference: there were enough answers to see a gap of 0.05, and the spreads overlapped or the gap was smaller.

Too few answers to tell: the scores were too spread out, or too few, to see a gap of 0.05 either way. It does not mean the skill did nothing.

What did each model score, without the skill and with it?

Under each model’s name, one row per run, runs one to three from the top. The hollow dot is a model’s average score without the skill, the filled dot its average with the skill, and the line between them is what the skill added. Each panel spans scores from zero at the left edge to one at the right. The two Claude Haiku rows were run for this report; the Claude Sonnet 5.5 and Claude Opus 5.5 rows are Report 013’s.

What do the runs show?

Which skills still help Claude Haiku 5.5?

Still help: none. Stopped helping, as far as these runs can tell: none. Helped Claude Haiku 4.5, too few answers to tell on Claude Haiku 5.5: documentation-and-adrs. Started helping: none. Did not clearly help either model in two or more of the three runs: code-review-and-quality, git-workflow-and-versioning.

A result that had too few answers to tell does not show that a skill stopped working. The two Claude Haiku arms also differ in more than the model: Claude Code applied its per-turn effort to every call of Claude Haiku 5.5 and to no call of Claude Haiku 4.5.

code-review-and-quality: Claude Haiku 5.5 averaged 0.821, 0.813 and 0.823 without the skill and 0.891, 0.861 and 0.770 with it, so the skill added +0.070, +0.048 and -0.053 across the three runs. Claude Haiku 4.5 averaged 0.339, 0.540 and 0.547 without it and 0.677, 0.628 and 0.566 with it, an addition of +0.339, +0.088 and +0.019.

git-workflow-and-versioning: Claude Haiku 5.5 averaged 0.300, 0.849 and 0.613 without the skill and 0.849, 0.839 and 0.848 with it, so the skill added +0.549, -0.010 and +0.235 across the three runs. Claude Haiku 4.5 averaged 0.834, 0.838 and 0.839 without it and 0.832, 0.828 and 0.813 with it, an addition of -0.002, -0.010 and -0.026.

documentation-and-adrs: Claude Haiku 5.5 averaged 0.561, 0.589 and 0.483 without the skill and 0.597, 0.500 and 0.542 with it, so the skill added +0.036, -0.089 and +0.059 across the three runs. Claude Haiku 4.5 averaged 0.626, 0.579 and 0.605 without it and 0.763, 0.772 and 0.739 with it, an addition of +0.138, +0.193 and +0.134.

How often did the skills clearly help each model?

Out of nine results for each model (three tasks, three runs), the skills clearly helped Claude Haiku 5.5 in two of nine, Claude Haiku 4.5 in three of nine, Claude Sonnet 5.5 in three of nine and Claude Opus 5.5 in four of nine. Claude Sonnet 5.5 and Claude Opus 5.5 are Report 013’s receipts, not run again, and these counts set the four models side by side without ranking them.

Why run every test three times?

Repeat runs of the same test did not always agree: six of the twelve model and task pairs changed reading between runs. Most of those changes were between clearly helped and too few answers to tell; only one pair, git workflow on Claude Haiku 5.5, read two answers that conflict, clearly helped in one run and no clear difference in another.

What are the limits, and what differs from Reports 009, 011 and 013?

What does this report not show?

Each skill was tested on one task only. Nothing here says how these skills do on other tasks, or how these models do at coding in general.

Each arm drew between 3 and 10 answers in a run, and the grader scored every answer three times. That is few answers. Where a result had too few answers to tell, the scores were too spread out, or too few, to see a gap of 0.05.

Every answer was scored by claude-opus-5 as the grader, as in Reports 009, 011 and 013. Our other reports use a different grader, set by the judge policy, so compare these scores with Reports 009, 011 and 013 only.

A result reads one model’s receipt: what the skill added to the model that answered, in that run. The page counts how many runs clearly helped each model and sets the counts side by side; it does not test whether one model’s lift differs from another’s.

What differs from the earlier reports?

Claude Haiku 5.5 came out on 7 Oct 2026, as the vendor’s release notes date it (the record, release-date.json, is published beside this page). The first run began on 7 Oct 2026 at 20:17 UTC.

The model before it is Claude Haiku 4.5. Claude Sonnet 5.5 and Claude Opus 5.5 are Report 013’s receipts, not run again: they ran on the same Claude Code, 2.1.284, on 29 Sep 2026.

Every test ran three times. No model was given an effort or thinking setting. Claude Code applied its own per-turn effort in every call of Claude Haiku 5.5 and no call of Claude Haiku 4.5.

Prices come from the vendor’s pricing page as saved on the day of the run: 0.1 and 0.5 dollars per million input and output tokens for Claude Haiku 5.5 (the vendor’s first rate, for the shorter prompts every call here used), and 1 and 5 for Claude Haiku 4.5. Prices change the estimated cost only, never a score.

A safety limit allowed up to 480 model calls per run directory. It was set for up to twelve skill runs, in case the comparison models were run again; each run made six.

Evidence and files

How the runs were set up, in full: departures and what stayed fixed

What departs from Report 013

The same Claude Code. Every call of the two Claude Haiku arms ran on Claude Code 2.1.284, the version Report 013 pinned, installed at the npm integrity the run record states. 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 prices are this run’s own. The runner’s registry had no row for claude-haiku-5-5. This run used a copy of it with that row added at 0.1 and 0.5 dollars per million input and output tokens, and claude-haiku-4-5-20251001 at 1 and 5, each read from the vendor’s pricing page on the day of the run (docs-pricing-snapshot-2026-10-08.json). The rows are this run’s only; the product’s registry, config/models.json, does not carry claude-haiku-5-5, and each receipt’s pricing_snapshot.source names that file, config/models.json, although its prices are this run’s copy. Prices change the estimated cost below and no score.
Two comparison arms are Report 013’s. claude-sonnet-5-5 and claude-opus-5-5 were not run again. Their rows are the receipts Report 013 published, run by run.

What stayed fixed

The inputs are the earlier reports’. The same SKILL.md bytes (skill.content_hash, the same in all of this report’s receipts and Report 013’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 theirs. Driftproof runner 0.10.1 on the claude-cli surface, with the flags their runs 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 theirs. claude-opus-5, with the grading template 82586d1e44f8 in every receipt.

The reading. 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. A skill clearly helped when the with-skill band and the without-skill band do not overlap and the averages differ by at least the 0.05 effect floor. A case that does not separate is read under spec 035’s rule: when the spreads and draws could not have resolved a shift of the floor’s size, it reads too few answers to tell, and otherwise no clear difference. Neither is evidence that the skill does nothing.

The figures behind each reading: every run’s averages, spreads and answers

Every run, without the skill and with it

task	model	run	without the skill: average ± spread (answers)	with the skill: average ± spread (answers)	what the skill added	result
`code-review-and-quality`	Claude Haiku 5.5	run 1	0.821 ± 0.051 (4 answers)	0.891 ± 0.012 (3 answers)	+0.070	Clearly helped
`code-review-and-quality`	Claude Haiku 5.5	run 2	0.813 ± 0.076 (4 answers)	0.861 ± 0.068 (4 answers)	+0.048	Too few answers to tell
`code-review-and-quality`	Claude Haiku 5.5	run 3	0.823 ± 0.031 (3 answers)	0.770 ± 0.158 (10 answers)	-0.053	Too few answers to tell
`code-review-and-quality`	Claude Haiku 4.5	run 1	0.339 ± 0.035 (3 answers)	0.677 ± 0.044 (4 answers)	+0.339	Clearly helped
`code-review-and-quality`	Claude Haiku 4.5	run 2	0.540 ± 0.172 (7 answers)	0.628 ± 0.112 (6 answers)	+0.088	Too few answers to tell
`code-review-and-quality`	Claude Haiku 4.5	run 3	0.547 ± 0.190 (4 answers)	0.566 ± 0.111 (4 answers)	+0.019	Too few answers to tell
`code-review-and-quality`	Claude Sonnet 5.5	run 1	0.878 ± 0.027 (3 answers)	0.887 ± 0.025 (3 answers)	+0.009	Too few answers to tell
`code-review-and-quality`	Claude Sonnet 5.5	run 2	0.849 ± 0.044 (3 answers)	0.900 ± 0.012 (3 answers)	+0.051	Too few answers to tell
`code-review-and-quality`	Claude Sonnet 5.5	run 3	0.861 ± 0.025 (3 answers)	0.896 ± 0.002 (3 answers)	+0.034	Too few answers to tell
`code-review-and-quality`	Claude Opus 5.5	run 1	0.831 ± 0.066 (4 answers)	0.907 ± 0.015 (3 answers)	+0.076	Too few answers to tell
`code-review-and-quality`	Claude Opus 5.5	run 2	0.857 ± 0.027 (3 answers)	0.894 ± 0.017 (3 answers)	+0.038	Too few answers to tell
`code-review-and-quality`	Claude Opus 5.5	run 3	0.713 ± 0.211 (5 answers)	0.903 ± 0.015 (3 answers)	+0.190	Too few answers to tell
`git-workflow-and-versioning`	Claude Haiku 5.5	run 1	0.300 ± 0.000 (3 answers)	0.849 ± 0.008 (3 answers)	+0.549	Clearly helped
`git-workflow-and-versioning`	Claude Haiku 5.5	run 2	0.849 ± 0.012 (3 answers)	0.839 ± 0.014 (3 answers)	-0.010	No clear difference
`git-workflow-and-versioning`	Claude Haiku 5.5	run 3	0.613 ± 0.293 (7 answers)	0.848 ± 0.004 (3 answers)	+0.235	Too few answers to tell
`git-workflow-and-versioning`	Claude Haiku 4.5	run 1	0.834 ± 0.016 (3 answers)	0.832 ± 0.002 (3 answers)	-0.002	No clear difference
`git-workflow-and-versioning`	Claude Haiku 4.5	run 2	0.838 ± 0.015 (3 answers)	0.828 ± 0.004 (3 answers)	-0.010	No clear difference
`git-workflow-and-versioning`	Claude Haiku 4.5	run 3	0.839 ± 0.014 (3 answers)	0.813 ± 0.003 (3 answers)	-0.026	No clear difference
`git-workflow-and-versioning`	Claude Sonnet 5.5	run 1	0.860 ± 0.009 (3 answers)	0.860 ± 0.009 (3 answers)	+0.000	No clear difference
`git-workflow-and-versioning`	Claude Sonnet 5.5	run 2	0.850 ± 0.007 (3 answers)	0.863 ± 0.000 (3 answers)	+0.013	No clear difference
`git-workflow-and-versioning`	Claude Sonnet 5.5	run 3	0.857 ± 0.009 (3 answers)	0.863 ± 0.009 (3 answers)	+0.007	No clear difference
`git-workflow-and-versioning`	Claude Opus 5.5	run 1	0.508 ± 0.287 (8 answers)	0.861 ± 0.005 (3 answers)	+0.353	Clearly helped
`git-workflow-and-versioning`	Claude Opus 5.5	run 2	0.485 ± 0.278 (9 answers)	0.857 ± 0.000 (3 answers)	+0.371	Clearly helped
`git-workflow-and-versioning`	Claude Opus 5.5	run 3	0.579 ± 0.322 (4 answers)	0.856 ± 0.010 (3 answers)	+0.276	Too few answers to tell
`documentation-and-adrs`	Claude Haiku 5.5	run 1	0.561 ± 0.042 (3 answers)	0.597 ± 0.006 (3 answers)	+0.036	Too few answers to tell
`documentation-and-adrs`	Claude Haiku 5.5	run 2	0.589 ± 0.005 (3 answers)	0.500 ± 0.087 (5 answers)	-0.089	Too few answers to tell
`documentation-and-adrs`	Claude Haiku 5.5	run 3	0.483 ± 0.035 (3 answers)	0.542 ± 0.065 (5 answers)	+0.059	Too few answers to tell
`documentation-and-adrs`	Claude Haiku 4.5	run 1	0.626 ± 0.044 (3 answers)	0.763 ± 0.092 (5 answers)	+0.138	Clearly helped
`documentation-and-adrs`	Claude Haiku 4.5	run 2	0.579 ± 0.053 (4 answers)	0.772 ± 0.063 (5 answers)	+0.193	Clearly helped
`documentation-and-adrs`	Claude Haiku 4.5	run 3	0.605 ± 0.084 (7 answers)	0.739 ± 0.092 (6 answers)	+0.134	Too few answers to tell
`documentation-and-adrs`	Claude Sonnet 5.5	run 1	0.572 ± 0.025 (3 answers)	0.781 ± 0.141 (6 answers)	+0.208	Clearly helped
`documentation-and-adrs`	Claude Sonnet 5.5	run 2	0.576 ± 0.025 (3 answers)	0.813 ± 0.094 (7 answers)	+0.238	Clearly helped
`documentation-and-adrs`	Claude Sonnet 5.5	run 3	0.574 ± 0.036 (3 answers)	0.847 ± 0.015 (3 answers)	+0.272	Clearly helped
`documentation-and-adrs`	Claude Opus 5.5	run 1	0.534 ± 0.039 (3 answers)	0.650 ± 0.102 (6 answers)	+0.116	Too few answers to tell
`documentation-and-adrs`	Claude Opus 5.5	run 2	0.550 ± 0.033 (3 answers)	0.850 ± 0.007 (3 answers)	+0.300	Clearly helped
`documentation-and-adrs`	Claude Opus 5.5	run 3	0.518 ± 0.083 (6 answers)	0.840 ± 0.020 (3 answers)	+0.322	Clearly helped

What the skill added is a model’s average with the skill minus its average without it. The result is the receipt’s own reading of that, as its receipt page shows it. Claude Sonnet 5.5 and Claude Opus 5.5 are Report 013’s receipts, run by run.

The run record: calls, costs and integrity

Run record

The runs. Each run went arm by arm, one skill at a time, code-review-and-quality first in each. The command file writes a status line after each skill run. The two arms of a run were started separately, claude-haiku-5-5 first and claude-haiku-4-5-20251001 beside it, into the same run directory, so a run’s status lines are the two arms’ together. The runs overlapped in time: the claude-haiku-4-5-20251001 arm of one run ran beside the claude-haiku-5-5 arm of the next. Each receipt’s run.date_utc is written by the runner and is not read here as a start or a finish.

run 1, 20261007T201714Z: six status lines, each with exit status 0 and one receipt, the last at 2026-10-07T21:09:35Z; 160 calls; 0 unmeasured draws.
run 2, 20261007T203747Z: six status lines, each with exit status 0 and one receipt, the last at 2026-10-07T21:36:53Z; 200 calls; 0 unmeasured draws.
run 3, 20261007T205604Z: six status lines, each with exit status 0 and one receipt, the last at 2026-10-07T22:01:08Z; 232 calls; 0 unmeasured draws.

The caps. A guard in front of Claude Code allowed 480 calls per run and 240 seconds per call, and stops a run when a call reports any model other than the one it asked for; no run stopped.

The cost. Metered spend, as the runner reports it, was 0.00 dollars in each skill run’s console output: 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.

run	arm	calls	generation, estimated USD	judge, estimated USD	total, estimated USD
run 1	`claude-haiku-5-5`	76	0.06	5.11	5.16
run 1	`claude-haiku-4-5-20251001`	84	0.45	5.42	5.87
run 2	`claude-haiku-5-5`	88	0.07	5.96	6.03
run 2	`claude-haiku-4-5-20251001`	112	0.63	7.54	8.16
run 3	`claude-haiku-5-5`	124	0.09	8.45	8.54
run 3	`claude-haiku-4-5-20251001`	108	0.58	7.12	7.70
all three runs			1.87	39.60	41.47

Integrity. Each receipt validates with its receipt hash verified, and each sidecar names its receipt’s hash. The raw call records carry the SKILL.md text and stay unpublished; each run’s list of their sha256 is published.

Published evidence

The files this report makes public. For each run, the six receipts with their summaries, surface sidecars and the runner’s console output, the run record, its registry copy, the status lines and the sha256 of every raw call record; the log of the three runs; the pricing snapshot; the release date record; and the call guard. The pricing snapshot records the sha256 of the page it was read from; that page is not published. Claude Sonnet 5.5 and Claude Opus 5.5 are read from Report 013’s evidence, which is published with that report.

docs/reports/014/evidence/docs-pricing-snapshot-2026-10-08.json
sha256 a6973f82f458dce2f784ccef13b677c9294423c66be482dd7346542ae768b891
docs/reports/014/evidence/guard.py
sha256 1afe55847e1e11a988a783ae6dd9dfa0af24df058dbe1fc3423e465f8eeb8ef4
docs/reports/014/evidence/release-date.json
sha256 15829c0f162165b58b76535cfa22f952f4da772238ba01f61b6a06f25a1cadcc
docs/reports/014/evidence/run-1/code-review-and-quality-claude-haiku-4-5-20251001-2026-10-07.json
sha256 7409c0b08102f58df8e9e04498e601bbc8d2456d42ac4eb9f752dc10e635faa9
docs/reports/014/evidence/run-1/code-review-and-quality-claude-haiku-4-5-20251001-2026-10-07.stdout.txt
sha256 327eb16cfa0f501ec09633a46bd6247fa388aa4f1053b928c06baf53f122e370
docs/reports/014/evidence/run-1/code-review-and-quality-claude-haiku-4-5-20251001-2026-10-07.summary.md
sha256 fa0b245ce830b5ba4f65abed3bee03167410601f7872e322a23b55ff66e02249
docs/reports/014/evidence/run-1/code-review-and-quality-claude-haiku-4-5-20251001-2026-10-07.surface.json
sha256 6d5eba5bddd7878d51eab39d6ad5dd274ff940ac52093bd5519c1fa4c34ecbde
docs/reports/014/evidence/run-1/code-review-and-quality-claude-haiku-5-5-2026-10-07.json
sha256 17dc654342887aed1d26a27bf22c1a6ca7f32696ff1339e077aab43b737f5f5b
docs/reports/014/evidence/run-1/code-review-and-quality-claude-haiku-5-5-2026-10-07.stdout.txt
sha256 dd6ee57b1197a5b2c148f88f88df0c0d03fc09bf096efe60b1117d4d48b93cbb
docs/reports/014/evidence/run-1/code-review-and-quality-claude-haiku-5-5-2026-10-07.summary.md
sha256 148c559aed7bc74d73b11727a8ef373ca07e4dd282a0c2269a19d547690e6e78
docs/reports/014/evidence/run-1/code-review-and-quality-claude-haiku-5-5-2026-10-07.surface.json
sha256 4f833c704c44b4c4b708013e34211fabad7233981fc485b24ac62ded6252dfd2
docs/reports/014/evidence/run-1/documentation-and-adrs-claude-haiku-4-5-20251001-2026-10-07.json
sha256 f38cffe80c42f5c38d17beab33ffec0357b04b678cd297c8a73094025df9d4a6
docs/reports/014/evidence/run-1/documentation-and-adrs-claude-haiku-4-5-20251001-2026-10-07.stdout.txt
sha256 0989e8e8623596c0f73f79d02e31229a1e46966615ae2b8029206efeba9dae1e
docs/reports/014/evidence/run-1/documentation-and-adrs-claude-haiku-4-5-20251001-2026-10-07.summary.md
sha256 e22f925ae199a52ef2542241ffb092ca6935cb1f36af77b8125c31860d9d8fd4
docs/reports/014/evidence/run-1/documentation-and-adrs-claude-haiku-4-5-20251001-2026-10-07.surface.json
sha256 91d1b2c45d1063104a343a2233446eceeb2958586cfb5da4f66e9b5f071f114b
docs/reports/014/evidence/run-1/documentation-and-adrs-claude-haiku-5-5-2026-10-07.json
sha256 624f4514a53a68a5d7c55ef8d6e262db59ed476b3ecc30ef5b0f5c8132917131
docs/reports/014/evidence/run-1/documentation-and-adrs-claude-haiku-5-5-2026-10-07.stdout.txt
sha256 b7b7ef164089cd61e658b66e06b0ed1961826ad4aab9637ccdbf4829d7188f32
docs/reports/014/evidence/run-1/documentation-and-adrs-claude-haiku-5-5-2026-10-07.summary.md
sha256 a73f272fb51d26e14c5246724856b9e6aabb83319bfec722c9bfd35171d84c57
docs/reports/014/evidence/run-1/documentation-and-adrs-claude-haiku-5-5-2026-10-07.surface.json
sha256 17b2bf0e5d753287321fc530b3d4d8368014fa95418cf2a4ba24cc807df3e80b
docs/reports/014/evidence/run-1/driftproof-calls.SHA256SUMS
sha256 7cc6fec3a1c94de10ebad31bd4b05a2403e54f4d865ef93b862954b55b19ed03
docs/reports/014/evidence/run-1/git-workflow-and-versioning-claude-haiku-4-5-20251001-2026-10-07.json
sha256 a8b7b5f1eaa7153f7c2308056cd512a521709be79adde2f3500a28e7868759f6
docs/reports/014/evidence/run-1/git-workflow-and-versioning-claude-haiku-4-5-20251001-2026-10-07.stdout.txt
sha256 6bd42bae48c2a617e8bb26ecef8be2737f3e533868d793ad612b1aad9a5adee1
docs/reports/014/evidence/run-1/git-workflow-and-versioning-claude-haiku-4-5-20251001-2026-10-07.summary.md
sha256 5fc938c36594834408eadbeeb70ef1bb8964898616aab92db6b433304b4dc4af
docs/reports/014/evidence/run-1/git-workflow-and-versioning-claude-haiku-4-5-20251001-2026-10-07.surface.json
sha256 a1df7136549b85922cd1fb8efb9e3282268c0725cfb814cf74bf6e94694de1bb
docs/reports/014/evidence/run-1/git-workflow-and-versioning-claude-haiku-5-5-2026-10-07.json
sha256 192e83c467239cfa603d579366c9518b74783933ac38ed56a7f1f8e7be0a4926
docs/reports/014/evidence/run-1/git-workflow-and-versioning-claude-haiku-5-5-2026-10-07.stdout.txt
sha256 5115150ba880d9084f9b20c3f5961ecb91872c42bbb6f2aeb58da0de2cbf510a
docs/reports/014/evidence/run-1/git-workflow-and-versioning-claude-haiku-5-5-2026-10-07.summary.md
sha256 46a8e9ff52d67e3d9eebb1523011f316a08fe38f824e7c903eedc778ff19df2e
docs/reports/014/evidence/run-1/git-workflow-and-versioning-claude-haiku-5-5-2026-10-07.surface.json
sha256 d4499e9a01089fb21dc68c04585bea4d27cf0f18f0d63769394e10a0298e8798
docs/reports/014/evidence/run-1/registry.json
sha256 26e6f9ff9bd1b182987b48c69d6a079c3f13a1d6ca4c21f63ed274c96ddef4c9
docs/reports/014/evidence/run-1/run-record.json
sha256 26c79da1b7a18cf6782f4a425222f95295450958b771e4892e5ecbc5210537c7
docs/reports/014/evidence/run-1/status.jsonl
sha256 74a4ae83791d0998ae41670cd24f31a081a11259738bd2255b43510752e273fa
docs/reports/014/evidence/run-2/code-review-and-quality-claude-haiku-4-5-20251001-2026-10-07.json
sha256 197ee185344a64f4eefbe81c84f6d3f2cfd25cb1f40d8200594abcaaa3de2a5e
docs/reports/014/evidence/run-2/code-review-and-quality-claude-haiku-4-5-20251001-2026-10-07.stdout.txt
sha256 d96d57640231014b82ac107766d7040c58b247d89b597f6016947f8222eb21e8
docs/reports/014/evidence/run-2/code-review-and-quality-claude-haiku-4-5-20251001-2026-10-07.summary.md
sha256 3544b5fd3288718f3ee75a43cece10ac3f763283636d87fcab51765d5cf3f05f
docs/reports/014/evidence/run-2/code-review-and-quality-claude-haiku-4-5-20251001-2026-10-07.surface.json
sha256 3b55d471887de8bc5091ff7c63eb3296d9287aa1acf78b2ce7392f8e13f6f8af
docs/reports/014/evidence/run-2/code-review-and-quality-claude-haiku-5-5-2026-10-07.json
sha256 b0be9dc1a87d175c1c3ae90a2d39ab6ca222371b06ff6b1593b88b95f20b07b8
docs/reports/014/evidence/run-2/code-review-and-quality-claude-haiku-5-5-2026-10-07.stdout.txt
sha256 1af81b8d56ce170b939bb5fd5a222606db0d138b50cec524660fee77757cfbc0
docs/reports/014/evidence/run-2/code-review-and-quality-claude-haiku-5-5-2026-10-07.summary.md
sha256 4adba7180f251320c61c2ad19e1cb280c3be9bcb8164d5aa094a15c2b5d47f6f
docs/reports/014/evidence/run-2/code-review-and-quality-claude-haiku-5-5-2026-10-07.surface.json
sha256 eda9bca33655adb37aef32dc9a4a4ce538bbdf1cf3c075ff7430cd532038af1c
docs/reports/014/evidence/run-2/documentation-and-adrs-claude-haiku-4-5-20251001-2026-10-07.json
sha256 23a2da9f4a5608b88075e611f8f8dd5cd0392dafc29ab3bc13e111513e6e7bef
docs/reports/014/evidence/run-2/documentation-and-adrs-claude-haiku-4-5-20251001-2026-10-07.stdout.txt
sha256 91d7201f5a123e32ef24a6957999528bd1ef5c5ea25f962efa8f7ec2e6d48471
docs/reports/014/evidence/run-2/documentation-and-adrs-claude-haiku-4-5-20251001-2026-10-07.summary.md
sha256 c50c194fa48072220176fa584957f5d478a5bf7899988f6db714e6c5fb7bed74
docs/reports/014/evidence/run-2/documentation-and-adrs-claude-haiku-4-5-20251001-2026-10-07.surface.json
sha256 df673686ac363a76c6d9872583084cbd6339af838ab4734312d93dc63e8d5e05
docs/reports/014/evidence/run-2/documentation-and-adrs-claude-haiku-5-5-2026-10-07.json
sha256 52f94978ca46d0079778b177ca4ee77045644af7e18edb7bfaca8cb3ea7f5280
docs/reports/014/evidence/run-2/documentation-and-adrs-claude-haiku-5-5-2026-10-07.stdout.txt
sha256 9d0e219372326a1bed014be15901342cd62700c721cdbdebbb33ce470d2f3fe8
docs/reports/014/evidence/run-2/documentation-and-adrs-claude-haiku-5-5-2026-10-07.summary.md
sha256 ddb00e1544002143cd87989597dd9e4d579e44a93484f4b23a257146a255a0fa
docs/reports/014/evidence/run-2/documentation-and-adrs-claude-haiku-5-5-2026-10-07.surface.json
sha256 6ed54fd4bbcf6b07313076496a6bc70ca53a15afcdd2fb3b37c6d3cd57452586
docs/reports/014/evidence/run-2/driftproof-calls.SHA256SUMS
sha256 13501cc74cea72078f65acbe457cbb9bc2e45b1bfe2e98034f8f80681fb62337
docs/reports/014/evidence/run-2/git-workflow-and-versioning-claude-haiku-4-5-20251001-2026-10-07.json
sha256 c80f0213a59939018e74acecbab3600cedbda108e1e3123263561e549d160295
docs/reports/014/evidence/run-2/git-workflow-and-versioning-claude-haiku-4-5-20251001-2026-10-07.stdout.txt
sha256 9482c5995cd293b3e89ba35a63f3ad189dc8540671bd4507df45c725b52eebec
docs/reports/014/evidence/run-2/git-workflow-and-versioning-claude-haiku-4-5-20251001-2026-10-07.summary.md
sha256 93ec8d057a7e7483d456f8f4ffde347a7cde25643c515370ede93ee28536b7d6
docs/reports/014/evidence/run-2/git-workflow-and-versioning-claude-haiku-4-5-20251001-2026-10-07.surface.json
sha256 c272e4bd61d17e75e8300881ac9c8c5e8af6c4eae19a90f0b0ee972563a643f8
docs/reports/014/evidence/run-2/git-workflow-and-versioning-claude-haiku-5-5-2026-10-07.json
sha256 460f50a6dbb609b62fb21cf46438382b9f01cc55e45ce79a3d03e244a015e6cc
docs/reports/014/evidence/run-2/git-workflow-and-versioning-claude-haiku-5-5-2026-10-07.stdout.txt
sha256 65af35eb3eb9f0787b8cefc679afc2beb6f08230a4b1c1a2975cd55aae9820c1
docs/reports/014/evidence/run-2/git-workflow-and-versioning-claude-haiku-5-5-2026-10-07.summary.md
sha256 a289b2e3d8cca367d48eb41b44d6d7034f51002017e1587229a63e3a47010080
docs/reports/014/evidence/run-2/git-workflow-and-versioning-claude-haiku-5-5-2026-10-07.surface.json
sha256 01dd511100e34f15e2750880003477c1828e643fc1d70965d86d9fb0cf227eb7
docs/reports/014/evidence/run-2/registry.json
sha256 26e6f9ff9bd1b182987b48c69d6a079c3f13a1d6ca4c21f63ed274c96ddef4c9
docs/reports/014/evidence/run-2/run-record.json
sha256 9ffbf5b00f2e1a8cd8cb5e2440644a7c3e673b6538891e59e66224e065f4f82c
docs/reports/014/evidence/run-2/status.jsonl
sha256 465c50a9aed0c9acb1e351d0f3c079e2b73c20420789dcbcc8bba9ca9f7f70af
docs/reports/014/evidence/run-3/code-review-and-quality-claude-haiku-4-5-20251001-2026-10-07.json
sha256 2f97943876d27abe43248c18eab0d9acdfe5b5aa8da2db416d26107a08ee519e
docs/reports/014/evidence/run-3/code-review-and-quality-claude-haiku-4-5-20251001-2026-10-07.stdout.txt
sha256 58d8c7da93b87e26e217f35fb4a2c57a16d18fc99ad486a5986e24dda7dfd9fc
docs/reports/014/evidence/run-3/code-review-and-quality-claude-haiku-4-5-20251001-2026-10-07.summary.md
sha256 39b8b84f674b7d8b9534b73f616874e1be137a8b75f0cd2d5186b56360f2a560
docs/reports/014/evidence/run-3/code-review-and-quality-claude-haiku-4-5-20251001-2026-10-07.surface.json
sha256 d534973ec68c817f9876cb636ffe7691c06a6cbfd9c02645c7c610ba5ba2c4c6
docs/reports/014/evidence/run-3/code-review-and-quality-claude-haiku-5-5-2026-10-07.json
sha256 6bc26b6843049c6687a0d289e937c233d5d81c0fd95b8374e7328df46d579e94
docs/reports/014/evidence/run-3/code-review-and-quality-claude-haiku-5-5-2026-10-07.stdout.txt
sha256 c2216ce72dabc23d322273269e956f0f23ef962a91d01947155a66c4e16845c1
docs/reports/014/evidence/run-3/code-review-and-quality-claude-haiku-5-5-2026-10-07.summary.md
sha256 00f104f22b214d0a095a20c874ffb3781500724401632bc99dca07942012b016
docs/reports/014/evidence/run-3/code-review-and-quality-claude-haiku-5-5-2026-10-07.surface.json
sha256 1ddbf3084806751e5de8f1b087e7512d80d95c6c6fc2088b52ec8aca4681b812
docs/reports/014/evidence/run-3/documentation-and-adrs-claude-haiku-4-5-20251001-2026-10-07.json
sha256 044f0d11596cb1d08361fdc3e6ec37c035fe7c3fade4dd2ba6adeee4e16d032b
docs/reports/014/evidence/run-3/documentation-and-adrs-claude-haiku-4-5-20251001-2026-10-07.stdout.txt
sha256 48c1c78ca007caa75b8fa31ee4e75ea6c154136f609ae1bf724eac2f995d3a4a
docs/reports/014/evidence/run-3/documentation-and-adrs-claude-haiku-4-5-20251001-2026-10-07.summary.md
sha256 fdda3694edf4c4e71f3c1d424406521671562fd28840376654443edd2c64c7f1
docs/reports/014/evidence/run-3/documentation-and-adrs-claude-haiku-4-5-20251001-2026-10-07.surface.json
sha256 3dc031a154c2d2570937240fc430c2c87e80050dc212f788a68b5f10db66d085
docs/reports/014/evidence/run-3/documentation-and-adrs-claude-haiku-5-5-2026-10-07.json
sha256 a2bd9689d91c7de05f97ae0c9ad9e4dc77105df7281ca7d493340b91db43b14f
docs/reports/014/evidence/run-3/documentation-and-adrs-claude-haiku-5-5-2026-10-07.stdout.txt
sha256 eee991cf780c1df506392afb24fcb3b759add62d75c7edfc143a128f99c02331
docs/reports/014/evidence/run-3/documentation-and-adrs-claude-haiku-5-5-2026-10-07.summary.md
sha256 18024a789cb36e5bf7faeb601dc8e5fa8eff2752a2e510a0d31731defffc0130
docs/reports/014/evidence/run-3/documentation-and-adrs-claude-haiku-5-5-2026-10-07.surface.json
sha256 776d3a87232779c8f05871b3dbbf837862d281591645dd5ebb9ad3018cff5d0f
docs/reports/014/evidence/run-3/driftproof-calls.SHA256SUMS
sha256 ebd8b9b2329e89098de0994259d90d3eba056d365a463c699efd96d091466114
docs/reports/014/evidence/run-3/git-workflow-and-versioning-claude-haiku-4-5-20251001-2026-10-07.json
sha256 ccf8b4e46ed8abdf1e0ff689a62220b5e0e99e5cb57f51fe8e9f094a0261d33b
docs/reports/014/evidence/run-3/git-workflow-and-versioning-claude-haiku-4-5-20251001-2026-10-07.stdout.txt
sha256 7b85daa3b055c228fca2ad595505e31678cd47a2594a212a9542fa8a8c59816a
docs/reports/014/evidence/run-3/git-workflow-and-versioning-claude-haiku-4-5-20251001-2026-10-07.summary.md
sha256 6d67a46f8b36b98d354d18a4b01d5f5e9cd53318b1d429c20993dcd9dc64b78e
docs/reports/014/evidence/run-3/git-workflow-and-versioning-claude-haiku-4-5-20251001-2026-10-07.surface.json
sha256 7dfd04928518183d061242edd27886c02df11dc0bdc618d25573cc749f63f194
docs/reports/014/evidence/run-3/git-workflow-and-versioning-claude-haiku-5-5-2026-10-07.json
sha256 30a75f51e905bf473ad8f117a0f0a9de745dce51e2454fea2f93e7e90a360043
docs/reports/014/evidence/run-3/git-workflow-and-versioning-claude-haiku-5-5-2026-10-07.stdout.txt
sha256 c63506ab79c4f303340d7ef7d6c531bc6fc5d1f8ed47fe1f8d836da632ee0d70
docs/reports/014/evidence/run-3/git-workflow-and-versioning-claude-haiku-5-5-2026-10-07.summary.md
sha256 c34d53078c4e9d8fb29010428d1e3948e585ecf8b4e689e0adfbd9b471986fd2
docs/reports/014/evidence/run-3/git-workflow-and-versioning-claude-haiku-5-5-2026-10-07.surface.json
sha256 f845e2e9339dbffb5ec6b984a466fadd8264808f16c9908b3cd50a3a8a2b570e
docs/reports/014/evidence/run-3/registry.json
sha256 26e6f9ff9bd1b182987b48c69d6a079c3f13a1d6ca4c21f63ed274c96ddef4c9
docs/reports/014/evidence/run-3/run-record.json
sha256 a29aa81dfae0ca07f44d2eb380e10437ca7523c60ed5b093341a9acca04f6c24
docs/reports/014/evidence/run-3/status.jsonl
sha256 86e6a1c39172affc9e8364732ac2f05f571306f09166dca51f15fb6aa85e93ce
docs/reports/014/evidence/runs-log.txt
sha256 3b4dee8f4daaccf3ef596fea0c8ee7cdf93336ccecc51f5d78935a62050f4776

Validate a receipt with npx driftproof validate <file>. Each copy here is byte-identical to its twin under specs/166-report-014-haiku-5-5/, and each sidecar names the receipt it belongs to and that receipt’s hash.

Receipts

Every receipt this report rests on, each one resolvable. Six for each run, one per model and skill, the same bytes as the evidence copies above.

code-review-and-quality@claude-haiku-5-5 (run 1): receipts/report-014/run-1/code-review-and-quality-claude-haiku-5-5-2026-10-07.json
git-workflow-and-versioning@claude-haiku-5-5 (run 1): receipts/report-014/run-1/git-workflow-and-versioning-claude-haiku-5-5-2026-10-07.json
documentation-and-adrs@claude-haiku-5-5 (run 1): receipts/report-014/run-1/documentation-and-adrs-claude-haiku-5-5-2026-10-07.json
code-review-and-quality@claude-haiku-4-5-20251001 (run 1): receipts/report-014/run-1/code-review-and-quality-claude-haiku-4-5-20251001-2026-10-07.json
git-workflow-and-versioning@claude-haiku-4-5-20251001 (run 1): receipts/report-014/run-1/git-workflow-and-versioning-claude-haiku-4-5-20251001-2026-10-07.json
documentation-and-adrs@claude-haiku-4-5-20251001 (run 1): receipts/report-014/run-1/documentation-and-adrs-claude-haiku-4-5-20251001-2026-10-07.json
code-review-and-quality@claude-haiku-5-5 (run 2): receipts/report-014/run-2/code-review-and-quality-claude-haiku-5-5-2026-10-07.json
git-workflow-and-versioning@claude-haiku-5-5 (run 2): receipts/report-014/run-2/git-workflow-and-versioning-claude-haiku-5-5-2026-10-07.json
documentation-and-adrs@claude-haiku-5-5 (run 2): receipts/report-014/run-2/documentation-and-adrs-claude-haiku-5-5-2026-10-07.json
code-review-and-quality@claude-haiku-4-5-20251001 (run 2): receipts/report-014/run-2/code-review-and-quality-claude-haiku-4-5-20251001-2026-10-07.json
git-workflow-and-versioning@claude-haiku-4-5-20251001 (run 2): receipts/report-014/run-2/git-workflow-and-versioning-claude-haiku-4-5-20251001-2026-10-07.json
documentation-and-adrs@claude-haiku-4-5-20251001 (run 2): receipts/report-014/run-2/documentation-and-adrs-claude-haiku-4-5-20251001-2026-10-07.json
code-review-and-quality@claude-haiku-5-5 (run 3): receipts/report-014/run-3/code-review-and-quality-claude-haiku-5-5-2026-10-07.json
git-workflow-and-versioning@claude-haiku-5-5 (run 3): receipts/report-014/run-3/git-workflow-and-versioning-claude-haiku-5-5-2026-10-07.json
documentation-and-adrs@claude-haiku-5-5 (run 3): receipts/report-014/run-3/documentation-and-adrs-claude-haiku-5-5-2026-10-07.json
code-review-and-quality@claude-haiku-4-5-20251001 (run 3): receipts/report-014/run-3/code-review-and-quality-claude-haiku-4-5-20251001-2026-10-07.json
git-workflow-and-versioning@claude-haiku-4-5-20251001 (run 3): receipts/report-014/run-3/git-workflow-and-versioning-claude-haiku-4-5-20251001-2026-10-07.json
documentation-and-adrs@claude-haiku-4-5-20251001 (run 3): receipts/report-014/run-3/documentation-and-adrs-claude-haiku-4-5-20251001-2026-10-07.json

Validate any of them with npx driftproof validate <file>. The comparison rows are Report 013’s receipts, listed on that report.