Does Accessibility audit help? Tested on accessibility tasks

Accessibility audit, from rampstackco/claude-skills. Its accessibility audit tasks were run with the skill loaded and without it. Each time it is the same task, run twice.

What we tested

Whether loading this skill helps a model find accessibility problems on a page.

What counts as helping

The skill has to find at least 10 percentage points more of the planted problems than the model finds without it, on the same 10 test pages.

How we scored it

We planted 48 known problems across the 10 test pages and counted how many the model named. The score is the share it found, from 0 to 1.

Without to with, per model

Via API tested via API

Gemini 3.1 Flash Lite First run, with and without the skill

0.633
0.728

No measured effect

Gemini 3.1 Flash Lite Wider run, three ways

0.673
0.728

No measured effect

Claude Haiku 4.5 First run, with and without the skill

0.410
0.699

Holds

GPT-5 mini First run, with and without the skill

0.693
0.689

No measured effect

GPT-5 mini Wider run, three ways

0.529
0.626

No measured effect

deterministic pass rate, 0 to 1

In Claude Code tested in Claude Code

Claude Fable 5.1 Fable 5.1, single attempt

0.875
0.935

No measured effect

Claude Fable 5.1 Wider run, three ways

0.855
0.895

No measured effect

Claude Fable 5 Second run, every task twice

0.800 without and with

No measured effect

Claude Opus 5 First run, single attempt

0.890
0.870

No measured effect

Claude Opus 5 Second run, every task twice

0.890
0.840

No measured effect

Claude Opus 5 Wider run, three ways

0.910
0.830

No measured effect

Claude Sonnet 5 Wider run, three ways

0.838
0.804

No measured effect

Claude Sonnet 5 Second run, every task twice

0.835
0.777

No measured effect

Claude Haiku 4.5 Second run, every task twice

0.458
0.690

Holds

Claude Haiku 4.5 Wider run, three ways

0.455
0.628

No measured effect

Claude Fable 5 First run, single attempt

served model mismatch; 1 usable pair

Withheld

deterministic pass rate, 0 to 1

A picture of the per-model numbers, drawn from the same results. The tables are the source. Grey is the score without. Colour is the score with. The bracket shows how much the difference could move if we ran it again. Rows are ordered by the score with. A model measured at two versions keeps its versions next to each other. The two test methods are reported separately and never averaged. Where the two scores are the same, the number is printed once at the end of the pair. A bracket wider than the axis is drawn to the edge with its cap omitted; the table gives its bounds.
Show per-task detail

Every item, without and with, per model

Via API tested via API

  • Claude Haiku 4.5
  • GPT-5 mini
  • Gemini 3.1 Flash Lite
  • Gemini 3.1 Flash Lite
  • GPT-5 mini

deterministic pass rate, 0 to 1

Hover or focus an item to read its task and every model’s two values.

Task-paired: these item means are what this cell's interval is built from.

The item means behind this plot
Per-item arm means. Without is the unaided arm, with is the treated arm.
ItemClaude Haiku 4.5 GPT-5 mini Gemini 3.1 Flash Lite Gemini 3.1 Flash Lite GPT-5 mini
Task without with without with without with without with without with
a10 0.250 0.500 0.250 0.500 0.250 0.750 0.250 0.750 0.250 0.500
a06 0.400 0.600 0.400 0.400 0.400 0.600 0.400 0.600 0.400 0.400
a08 0.000 0.667 0.333 0.667 0.667 1.000 0.667 1.000 0.667 0.667
a03 0.400 0.600 0.600 0.400 0.200 0.200 0.600 0.200 0.600 0.600
a05 0.600 0.600 0.000 0.000 0.800 0.600 0.800 0.600 0.800 0.600
a01 0.500 0.875 0.625 0.875 0.750 0.875 0.750 0.875 0.625 0.875
a09 0.333 1.000 0.667 1.000 0.667 0.667 0.667 0.667 1.000 0.667
a04 0.200 0.400 1.000 1.000 0.600 1.000 0.600 1.000 1.000 1.000
a02 0.667 1.000 0.667 0.667 1.000 0.833 1.000 0.833 0.833 0.833
a07 0.750 0.750 0.750 0.750 1.000 0.750 1.000 0.750 0.750 0.750
A picture of the per-item means behind the per-model numbers. The tables are the source. Each model draws two lines over the same task set: a dashed line through its unaided scores and a solid line through its treated ones. Items are ordered by the mean unaided score across the models that measured them, lowest first. The two instruments are reported separately and never averaged.

Without to with, per model

In Claude Code tested in Claude Code

Claude Haiku 4.5 Expansion cohort (three arms)

No measured effect

Claude Haiku 4.5 Panel v2 (repeat sampling)

Holds

Claude Fable 5 Panel v2 (repeat sampling)

No measured effect

Claude Sonnet 5 Panel v2 (repeat sampling)

No measured effect

Claude Sonnet 5 Expansion cohort (three arms)

No measured effect

Claude Fable 5.1 Expansion cohort (three arms)

No measured effect

Claude Fable 5.1 Fable 5.1, one pass

No measured effect

Claude Opus 5 Skill cohort (one pass)

No measured effect

Claude Opus 5 Panel v2 (repeat sampling)

No measured effect

Claude Opus 5 Expansion cohort (three arms)

No measured effect

deterministic pass rate, 0 to 1

A picture of the per-model numbers, drawn from the same results. The tables are the source. Each row runs from the score without to the score with. The thin bar beneath it is the bracket, and it shows how much the difference could move if we ran it again. The two test methods are reported separately and never averaged. A bracket wider than the axis is drawn to the edge with its end cap omitted; the table gives its bounds.

The Claude Code runs committed their cells as arm means and intervals rather than as per-item values, so this panel is drawn without to with rather than item by item.

What exactly was tested, and how it was scored
Repository
rampstackco/claude-skills
Path
skills/accessibility-audit
Commit
0479242522549dfdb389bb9b7807ad4d6016ffb7
Content hash
7b564cd3147768a1e9775a739e55fe56248e996c3fcdab86c264fe7289644921
Date tested
2026-08-30

The pin is the whole of this skill's identity here. It resolves at https://github.com/rampstackco/claude-skills/tree/0479242522549dfdb389bb9b7807ad4d6016ffb7/skills/accessibility-audit, and the content hash is a sha256 over exactly the text the model was given with the skill loaded. Nothing else about the skill appears on this site.

S03-accessibility-audit

Loading skills/accessibility-audit from rampstackco/claude-skills improves outputs on accessibility audit tasks.

Pass criterion
With-arm mean coverage exceeds the without-arm by at least 0.10 on the same 10 fixture pages at the same model and settings.
Scale
unit. Defects found over defects seeded, so a [0,1] fraction, which is the unit scale.
Task pairs planned per model
10
Notes
48 defects seeded across ten pages, ledger at fixtures/accessibility/defect-ledger.json. False positives are counted on the record and never netted off coverage.

Injected context tokens

Injected context tokens, per arm
ArmContext charactersInjected context tokens
without00
with43,62010,239

The character count is exact: it is the length of the text the with arm is given, and the content hash above is a sha256 over that same text. The token figure is an estimate at 4.26 characters per token, the ratio the phase 1 run measured over 3,120 calls, and it is labelled an estimate until a run reports its own token counts. The without arm is given the identical prompt and nothing else, so its zero is a measurement rather than a missing value.

Verdict per model

Effect per model
ModelReadingOrbitWithout -> withEffect95 percent intervalTask pairsModel version returned
Claude Haiku 4.5Deterministic, scored against a committed answer keyStable0.4100 -> 0.6992+0.2892[0.1450, 0.4333]10 pairs, 20 of 20 recordsclaude-haiku-4-5-20251001
Gemini 3.1 Flash LiteDeterministic, scored against a committed answer keyIn free drift unclear. The model scored 63% without it.0.6333 -> 0.7275+0.0942[-0.0683, 0.2566]10 pairs, 20 of 20 recordsgemini-3.1-flash-lite
GPT-5 miniDeterministic, scored against a committed answer keyIn free drift unclear. The model scored 69% without it.0.6925 -> 0.6892-0.0033[-0.1119, 0.1052]10 pairs, 20 of 20 recordsgpt-5-mini-2025-08-07
Gemini 3.1 Flash LiteDeterministic, scored against a committed answer keyIn free drift unclear. The model scored 63% without it.0.6733 -> 0.7275+0.0542[-0.1349, 0.2433]10 pairs, 30 of 30 recordsgemini-3.1-flash-lite
GPT-5 miniDeterministic, scored against a committed answer keyIn free drift unclear. The model scored 57% without it.0.5292 -> 0.6258+0.0967[-0.0153, 0.2086]10 pairs, 30 of 30 recordsgpt-5-mini-2025-08-07
  • Stable: the treatment arm outscored the control arm by more than the threshold, and the interval excludes zero.
  • In free drift: no separation the design can resolve. Not evidence of no effect.

Cost

Cost per task
ModelArmTasks attemptedMean input tokensMean output tokensCost per taskCost ratioGrading cost per task pair
Claude Haiku 4.5 served claude-haiku-4-5-20251001without1060055$0.00087
Claude Haiku 4.5 served claude-haiku-4-5-20251001with1012,59868$0.0129414.83x
GPT-5 mini served gpt-5-mini-2025-08-07without10515115$0.00036
GPT-5 mini served gpt-5-mini-2025-08-07with1010,995115$0.002988.30x
Gemini 3.1 Flash Litewithout1052859$0.00022
Gemini 3.1 Flash Litewith1011,59567$0.0030013.61x

Defined in the metrics canon. Cost per task divides every dollar spent on an arm by the tasks attempted on it, including tasks whose call returned nothing, because a call that returned nothing was still billed.

Where the with arm did not win

Claude Haiku 4.5, Deterministic, scored against a committed answer key

Drew on 2 of 10 pairs: a05, a07.

Findings not in the answer key: 19 with the skill, 22 without. These are counted and never netted off coverage.

Gemini 3.1 Flash Lite, Deterministic, scored against a committed answer key

Lost on 3 of 10 pairs: a07 (-0.2500), a05 (-0.2000), a02 (-0.1667).

Drew on 2 of 10 pairs: a03, a09.

Findings not in the answer key: 20 with the skill, 18 without. These are counted and never netted off coverage.

GPT-5 mini, Deterministic, scored against a committed answer key

Lost on 2 of 10 pairs: a09 (-0.3333), a05 (-0.2000).

Drew on 6 of 10 pairs: a02, a03, a04, a06, a07, a08.

Findings not in the answer key: 59 with the skill, 64 without. These are counted and never netted off coverage.

Gemini 3.1 Flash Lite, Deterministic, scored against a committed answer key

Lost on 4 of 10 pairs: a03 (-0.4000), a07 (-0.2500), a05 (-0.2000), a02 (-0.1667).

Drew on 1 of 10 pairs: a09.

Findings not in the answer key: 23 with the skill, 18 without. These are counted and never netted off coverage.

GPT-5 mini, Deterministic, scored against a committed answer key

Lost on 1 of 10 pairs: a03 (-0.2000).

Drew on 5 of 10 pairs: a02, a04, a05, a06, a07.

Findings not in the answer key: 67 with the skill, 69 without. These are counted and never netted off coverage.

In Claude Code

These cells are tested in Claude Code, on a subscription path with no API key. They are a second instrument: no figure here is averaged with one tested via API above. How the two were compared.

  • Claude Opus 5 · unclear, and both runs agree
  • Claude Fable 5 · unclear under repeat sampling
  • Claude Haiku 4.5 · helped under repeat sampling
  • Claude Sonnet 5 · unclear under repeat sampling
  • Claude Fable 5.1 · unclear in the version re-test

Skill cohort (one pass)

one pass over each committed item. Matrix hash 2707a06a. No pre-registration: this run predates the practice on this arm. Run report.

One row per model. Every figure is read from this run’s committed cells.
ModelNo skillWith skillDeltaIntervalPairsOrbitInjected tokensFalse positives per page
Claude Opus 50.89000.8700-0.020-0.143 to 0.10310In free drift16,69928 with, 15 without

Verdict withheld on this run

Claude Fable 5: served model mismatch; 1 usable pair. The plan served a different model on most of one arm, so these records are evidence about that other model and are scored into no cell. They are kept, and their tokens still count against what the run consumed.

Panel v2 (repeat sampling)

two passes over each committed item, under repeat sampling. Matrix hash 94bab960. Pre-registration · Run report.

One row per model. Every figure is read from this run’s committed cells.
ModelNo skillWith skillDeltaIntervalPairsOrbitInjected tokens
Claude Fable 50.80000.8000+0.0000.000 to 0.0002In free drift387,982
Claude Haiku 4.50.45830.6904+0.2320.088 to 0.37710Stable270,552
Claude Opus 50.89000.8400-0.050-0.178 to 0.07810In free drift371,812
Claude Sonnet 50.83500.7775-0.058-0.140 to 0.02510In free drift384,852
Thinking tokens, recorded per arm. Reported, never scored: no interval or Orbit on this page consults them.
ModelArmRecords that reasonedMean thinking tokensMost on one record
Claude Fable 5with skill20 of 201,4053,393
Claude Fable 5no skill38 of 38522965

Fable 5.1, one pass

one pass over each committed item. Matrix hash db373661. Pre-registration · Run report.

One row per model. Every figure is read from this run’s committed cells.
ModelNo skillWith skillDeltaIntervalPairsOrbitInjected tokensFalse positives per page
Claude Fable 5.10.87500.9350+0.060-0.024 to 0.14420In free driftnot measured12 with, 11 without

Expansion cohort (three arms)

one pass over each committed item. Matrix hash 0078cfa8. Pre-registration · Run report.

One row per model. Every figure is read from this run’s committed cells.
ModelNo skillWith skillDeltaIntervalPairsOrbitInjected tokensFalse positives per page
Claude Fable 5.10.85500.8950+0.040-0.012 to 0.09210In free driftnot measured: this run records the dose per arm, in the doses list19 with, 13 without
Claude Haiku 4.50.45500.6283+0.1730.030 to 0.31710In free driftnot measured: this run records the dose per arm, in the doses list18 with, 20 without
Claude Opus 50.91000.8300-0.080-0.200 to 0.04010In free driftnot measured: this run records the dose per arm, in the doses list29 with, 16 without
Claude Sonnet 50.83750.8042-0.033-0.111 to 0.04410In free driftnot measured: this run records the dose per arm, in the doses list46 with, 37 without

Every figure above is computed at build time from harness/results/runs-skills.jsonl and harness/results/ledger-skills.jsonl, both committed, by the frozen status_v1 rule and the metrics_v1 cost definitions. How a claim gets tested.

720 run records behind this page. Every verdict is computed at build time by the same frozen status_v1 rule that decides every other verdict on this site, and nothing here is written by hand.