The same skill on different models and dates
Driftproof measured some skills and instruction files on more than one model, and some with the same tests on different dates. Here is what each receipt read, linked to the receipt.
“Clearly helped” means the answers with the skill scored clearly higher on at least one test task, and clearly lower on none. “Clearly hurt” means they scored clearly lower on at least one test task, even if another task scored clearly higher. “Too few answers to tell” means there were too few answers, or they varied too much, to call it either way; it does not mean the skill does nothing. A skill that helped on one model is not a guarantee that it helps on another.
Readings of one text on one model can differ between runs, so two different readings on two models do not show that the model made the difference.
A receipt is the record of one test run, with the skill and without it, sealed so that a changed file no longer matches its seal. Each reading is one receipt's, and links to that receipt's page. Where one model has more than one receipt, each is listed on its own. Nothing here is combined, averaged or ranked.
The first list holds each skill or file with a reading that carries a verdict on at least two models. The second list holds each one measured on two or more models without that: each row there has a verdict on at most one model, and each receipt page without a verdict says why. Here a verdict is any reading from Clearly helped to Too few test tasks to tell. Skills and files measured on only one model are not listed.
Each row in the first list says whether its receipts record the same test tasks, the same judge and grading, and the same run setup, and names any receipt that does not. The test tasks are what each answer was asked to do. The judge and grading are the model that scored the answers, its instructions and the code that ran it; the judge counts as different when it is the model that answered in some receipts and not in others. The run setup is the versions of the tools that ran the test, how the skill was given to the model, and how many answers each task asked for.
An instruction file is named by the repository it came from, then its file name. Where a row includes models from more than one maker, it compares no makers: as Report 002 says, it is not a model ranking.
The key at the end of this page gives the exact rule for each reading.
Skills and files with a verdict on two or more models
claude-token-efficient CLAUDE.md
Every receipt in this row records the same test tasks. The receipt from Report 016 on Claude Sonnet 5.5 was scored by the same model that answered it. Every receipt in this row records the same run setup. These receipts were not all measured the same way, so read each one on its own.
On Claude Haiku 5.5: Too few answers to tell (Report 016, 2026-10-28).
On Claude Sonnet 5.5: Too few answers to tell (Report 016, 2026-10-28).
code-review-and-quality
The receipts from Reports 006, 007 and 008 record different test tasks. Not every receipt records its judge and grading: the receipts from Reports 006, 007 and 008 do not. The receipts from Report 009; and Report 011 on Claude Opus 5 were scored by the same model that answered them. The receipts from Reports 006, 007 and 008 record a different run setup. These receipts were not all measured the same way, so read each one on its own.
Driftproof measured more than one version of this text. This row is the version in Reports 006, 007, 008, 009, 011, 013 and 014.
On Claude Fable 5: Recorded before answered_by existed (Report 006, 2026-08-28); Recorded before answered_by existed (Reports 007 and 008, 2026-08-31).
On Claude Fable 5.1: Recorded before answered_by existed (Report 008, 2026-09-01).
On Claude Opus 5: Clearly helped (Report 009, 2026-09-15); Too few answers to tell (Report 011, 2026-09-23).
On Claude Opus 5.5: Too few answers to tell (Report 011, 2026-09-23); Too few answers to tell (Report 013, run 1, 2026-09-29); Too few answers to tell (Report 013, run 2, 2026-09-29); Too few answers to tell (Report 013, run 3, 2026-09-29).
On Claude Sonnet 5.5: Too few answers to tell (Report 013, run 1, 2026-09-29); Too few answers to tell (Report 013, run 2, 2026-09-29); Too few answers to tell (Report 013, run 3, 2026-09-29).
On Claude Sonnet 5: Clearly helped (Report 013, run 1, 2026-09-29); Clearly helped (Report 013, run 2, 2026-09-29); Too few answers to tell (Report 013, run 3, 2026-09-29).
On Claude Haiku 5.5: Clearly helped (Report 014, run 1, 2026-10-07); Too few answers to tell (Report 014, run 2, 2026-10-07); Too few answers to tell (Report 014, run 3, 2026-10-07).
On Claude Haiku 4.5: Clearly helped (Report 014, run 1, 2026-10-07); Too few answers to tell (Report 014, run 2, 2026-10-07); Too few answers to tell (Report 014, run 3, 2026-10-07).
Report 007's note, on its own page: “Note, 2026-10-03. The headline counts on this page include receipts whose receipt pages mark them Not measured. Corrected counts follow in spec 159.” Report 008 carries the same note on its page.
design-mobile-apps
Every receipt in this row records the same test tasks. The receipt from Report 018 was scored by the same model that answered it. The receipt from Report 017 was scored by a different version of the scoring code. The receipts from Reports 017 and 018 record a different run setup. These receipts were not all measured the same way, so read each one on its own.
On Claude Sonnet 5.5: Clearly helped (Report 015, 2026-10-18); Clearly helped (Report 017, 2026-11-01).
On Claude Opus 5.5: Clearly helped (Report 015, 2026-10-18).
On Claude Haiku 5.5: Clearly helped (Report 018, 2026-11-02).
documentation-and-adrs
Every receipt in this row records the same test tasks. The receipts from Report 009; and Report 011 on Claude Opus 5 were scored by the same model that answered them. Every receipt in this row records the same run setup. These receipts were not all measured the same way, so read each one on its own.
Driftproof measured more than one version of this text. This row is the version in Reports 009, 011, 013 and 014.
On Claude Opus 5: Clearly helped (Report 009, 2026-09-15); Clearly helped (Report 011, 2026-09-23).
On Claude Opus 5.5: Too few answers to tell (Report 011, 2026-09-23); Too few answers to tell (Report 013, run 1, 2026-09-29); Clearly helped (Report 013, run 2, 2026-09-29); Clearly helped (Report 013, run 3, 2026-09-29).
On Claude Sonnet 5.5: Clearly helped (Report 013, run 1, 2026-09-29); Clearly helped (Report 013, run 2, 2026-09-29); Clearly helped (Report 013, run 3, 2026-09-29).
On Claude Sonnet 5: Too few answers to tell (Report 013, run 1, 2026-09-29); Too few answers to tell (Report 013, run 2, 2026-09-29); Too few answers to tell (Report 013, run 3, 2026-09-29).
On Claude Haiku 5.5: Too few answers to tell (Report 014, run 1, 2026-10-07); Too few answers to tell (Report 014, run 2, 2026-10-07); Too few answers to tell (Report 014, run 3, 2026-10-07).
On Claude Haiku 4.5: Clearly helped (Report 014, run 1, 2026-10-07); Clearly helped (Report 014, run 2, 2026-10-07); Too few answers to tell (Report 014, run 3, 2026-10-07).
fable-method AGENTS.md
Every receipt in this row records the same test tasks. The receipt from Report 016 on Claude Sonnet 5.5 was scored by the same model that answered it. Every receipt in this row records the same run setup. These receipts were not all measured the same way, so read each one on its own.
On Claude Haiku 5.5: Too few answers to tell (Report 016, 2026-10-28).
On Claude Sonnet 5.5: Too few answers to tell (Report 016, 2026-10-28).
frontend-design
Every receipt in this row records the same test tasks. The receipt from Report 018 was scored by the same model that answered it. The receipt from Report 017 was scored by a different version of the scoring code. The receipts from Reports 017 and 018 record a different run setup. These receipts were not all measured the same way, so read each one on its own.
On Claude Sonnet 5.5: Too few answers to tell (Report 015, 2026-10-18); Too few answers to tell (Report 017, 2026-11-01).
On Claude Opus 5.5: Too few answers to tell (Report 015, 2026-10-18).
On Claude Haiku 5.5: Clearly hurt (Report 018, 2026-11-02).
git-workflow-and-versioning
Every receipt in this row records the same test tasks. The receipts from Report 009; and Report 011 on Claude Opus 5 were scored by the same model that answered them. Every receipt in this row records the same run setup. These receipts were not all measured the same way, so read each one on its own.
Driftproof measured more than one version of this text. This row is the version in Reports 009, 011, 013 and 014.
On Claude Opus 5: Clearly helped (Report 009, 2026-09-15); Too few answers to tell (Report 011, 2026-09-23).
On Claude Opus 5.5: Clearly helped (Report 011, 2026-09-23); Clearly helped (Report 013, run 1, 2026-09-29); Clearly helped (Report 013, run 2, 2026-09-29); Too few answers to tell (Report 013, run 3, 2026-09-29).
On Claude Sonnet 5.5: No clear difference (Report 013, run 1, 2026-09-29); No clear difference (Report 013, run 2, 2026-09-29); No clear difference (Report 013, run 3, 2026-09-29).
On Claude Sonnet 5: Too few answers to tell (Report 013, run 1, 2026-09-29); Too few answers to tell (Report 013, run 2, 2026-09-29); Too few answers to tell (Report 013, run 3, 2026-09-29).
On Claude Haiku 5.5: Clearly helped (Report 014, run 1, 2026-10-07); No clear difference (Report 014, run 2, 2026-10-07); Too few answers to tell (Report 014, run 3, 2026-10-07).
On Claude Haiku 4.5: No clear difference (Report 014, run 1, 2026-10-07); No clear difference (Report 014, run 2, 2026-10-07); No clear difference (Report 014, run 3, 2026-10-07).
grill-me
Every receipt in this row records the same test tasks. The receipts from Report 018; and canary run on Claude Haiku 5.5 were scored by the same model that answered them. The receipts from Report 017; and canary run on Claude Sonnet 5.5 were scored by a different version of the scoring code. The receipts from Report 017; Report 018; and canary run record a different run setup. These receipts were not all measured the same way, so read each one on its own.
This text only tells the model to call another skill with the Skill tool, so its receipts read that instruction, not the skill it names.
On Claude Sonnet 5.5: Too few answers to tell (Report 015, 2026-10-18); Clearly hurt (Report 017, 2026-11-01); Too few answers to tell (canary run, 2026-11-08).
On Claude Opus 5.5: Too few answers to tell (Report 015, 2026-10-18).
On Claude Haiku 5.5: Clearly helped (Report 018, 2026-11-02); Clearly helped (canary run, 2026-11-08).
grill-with-docs
Every receipt in this row records the same test tasks. The receipts from Report 018; and canary run on Claude Haiku 5.5 were scored by the same model that answered them. The receipts from Report 017; and canary run on Claude Sonnet 5.5 were scored by a different version of the scoring code. The receipts from Report 017; Report 018; and canary run record a different run setup. These receipts were not all measured the same way, so read each one on its own.
This text only tells the model to call another skill with the Skill tool, so its receipts read that instruction, not the skill it names.
On Claude Sonnet 5.5: Clearly hurt (Report 015, 2026-10-18); Clearly hurt (Report 017, 2026-11-01); Clearly hurt (canary run, 2026-11-08).
On Claude Opus 5.5: Too few answers to tell (Report 015, 2026-10-18).
On Claude Haiku 5.5: Too few answers to tell (Report 018, 2026-11-02); Too few answers to tell (canary run, 2026-11-08).
grilling
Every receipt in this row records the same test tasks. The receipt from Report 018 was scored by the same model that answered it. The receipt from Report 017 was scored by a different version of the scoring code. The receipts from Reports 017 and 018 record a different run setup. These receipts were not all measured the same way, so read each one on its own.
On Claude Sonnet 5.5: Clearly helped (Report 015, 2026-10-18); Clearly helped (Report 017, 2026-11-01).
On Claude Opus 5.5: Clearly helped (Report 015, 2026-10-18).
On Claude Haiku 5.5: Clearly hurt (Report 018, 2026-11-02).
handoff
Every receipt in this row records the same test tasks. The receipt from Report 018 was scored by the same model that answered it. The receipt from Report 017 was scored by a different version of the scoring code. The receipts from Reports 017 and 018 record a different run setup. These receipts were not all measured the same way, so read each one on its own.
On Claude Sonnet 5.5: Clearly helped (Report 015, 2026-10-18); Clearly hurt (Report 017, 2026-11-01).
On Claude Opus 5.5: Clearly helped (Report 015, 2026-10-18).
On Claude Haiku 5.5: Too few answers to tell (Report 018, 2026-11-02).
hyperframes-animation
Every receipt in this row records the same test tasks. The receipt from Report 018 was scored by the same model that answered it. The receipt from Report 017 was scored by a different version of the scoring code. The receipts from Reports 017 and 018 record a different run setup. These receipts were not all measured the same way, so read each one on its own.
On Claude Sonnet 5.5: Clearly helped (Report 015, 2026-10-18); Clearly helped (Report 017, 2026-11-01).
On Claude Opus 5.5: Clearly helped (Report 015, 2026-10-18).
On Claude Haiku 5.5: Clearly helped (Report 018, 2026-11-02).
hyperframes-cli
Every receipt in this row records the same test tasks. The receipts from Report 018; and canary run on Claude Haiku 5.5 were scored by the same model that answered them. The receipts from Report 017; and canary run on Claude Sonnet 5.5 were scored by a different version of the scoring code. The receipts from Report 017; Report 018; and canary run record a different run setup. These receipts were not all measured the same way, so read each one on its own.
On Claude Sonnet 5.5: Clearly helped (Report 015, 2026-10-18); Clearly helped (Report 017, 2026-11-01); Clearly helped (canary run, 2026-11-08).
On Claude Opus 5.5: Clearly helped (Report 015, 2026-10-18).
On Claude Haiku 5.5: Clearly helped (Report 018, 2026-11-02); Clearly helped (canary run, 2026-11-08).
hyperframes-core
Every receipt in this row records the same test tasks. The receipt from Report 018 was scored by the same model that answered it. The receipt from Report 017 was scored by a different version of the scoring code. The receipts from Reports 017 and 018 record a different run setup. These receipts were not all measured the same way, so read each one on its own.
On Claude Sonnet 5.5: Clearly helped (Report 015, 2026-10-18); Clearly helped (Report 017, 2026-11-01).
On Claude Opus 5.5: Clearly helped (Report 015, 2026-10-18).
On Claude Haiku 5.5: Clearly helped (Report 018, 2026-11-02).
improve-codebase-architecture
Every receipt in this row records the same test tasks. The receipts from Report 018; and canary run on Claude Haiku 5.5 were scored by the same model that answered them. The receipts from Report 017; and canary run on Claude Sonnet 5.5 were scored by a different version of the scoring code. The receipts from Report 017; Report 018; and canary run record a different run setup. These receipts were not all measured the same way, so read each one on its own.
On Claude Sonnet 5.5: Too few answers to tell (Report 015, 2026-10-18); Clearly helped (Report 017, 2026-11-01); Clearly helped (canary run, 2026-11-08).
On Claude Opus 5.5: Clearly helped (Report 015, 2026-10-18).
On Claude Haiku 5.5: Clearly helped (Report 018, 2026-11-02); Clearly helped (canary run, 2026-11-08).
learn-harness-engineering AGENTS.md
Every receipt in this row records the same test tasks. The receipt from Report 016 on Claude Sonnet 5.5 was scored by the same model that answered it. Every receipt in this row records the same run setup. These receipts were not all measured the same way, so read each one on its own.
On Claude Haiku 5.5: Too few answers to tell (Report 016, 2026-10-28).
On Claude Sonnet 5.5: Too few test tasks to tell (Report 016, 2026-10-28).
learn-harness-engineering CLAUDE.md
Every receipt in this row records the same test tasks. The receipt from Report 016 on Claude Sonnet 5.5 was scored by the same model that answered it. Every receipt in this row records the same run setup. These receipts were not all measured the same way, so read each one on its own.
On Claude Haiku 5.5: Too few answers to tell (Report 016, 2026-10-28).
On Claude Sonnet 5.5: Too few test tasks to tell (Report 016, 2026-10-28).
media-use
Every receipt in this row records the same test tasks. The receipt from Report 018 was scored by the same model that answered it. The receipt from Report 017 was scored by a different version of the scoring code. The receipts from Reports 017 and 018 record a different run setup. These receipts were not all measured the same way, so read each one on its own.
On Claude Sonnet 5.5: Clearly helped (Report 015, 2026-10-18); Clearly helped (Report 017, 2026-11-01).
On Claude Opus 5.5: Clearly helped (Report 015, 2026-10-18).
On Claude Haiku 5.5: Clearly helped (Report 018, 2026-11-02).
ponytail AGENTS.md
Every receipt in this row records the same test tasks. The receipt from Report 016 on Claude Sonnet 5.5 was scored by the same model that answered it. Every receipt in this row records the same run setup. These receipts were not all measured the same way, so read each one on its own.
On Claude Haiku 5.5: Too few answers to tell (Report 016, 2026-10-28).
On Claude Sonnet 5.5: Clearly helped (Report 016, 2026-10-28).
prototype
Every receipt in this row records the same test tasks. The receipt from Report 018 was scored by the same model that answered it. The receipt from Report 017 was scored by a different version of the scoring code. The receipts from Reports 017 and 018 record a different run setup. These receipts were not all measured the same way, so read each one on its own.
On Claude Sonnet 5.5: Clearly helped (Report 015, 2026-10-18); Clearly helped (Report 017, 2026-11-01).
On Claude Opus 5.5: Clearly helped (Report 015, 2026-10-18).
On Claude Haiku 5.5: Clearly helped (Report 018, 2026-11-02).
reddit-automation
Every receipt in this row records the same test tasks. The receipt from Report 018 was scored by the same model that answered it. The receipt from Report 017 was scored by a different version of the scoring code. The receipts from Reports 017 and 018 record a different run setup. These receipts were not all measured the same way, so read each one on its own.
On Claude Sonnet 5.5: Clearly hurt (Report 015, 2026-10-18); Clearly hurt (Report 017, 2026-11-01).
On Claude Opus 5.5: Clearly hurt (Report 015, 2026-10-18).
On Claude Haiku 5.5: Too few answers to tell (Report 018, 2026-11-02).
setup-matt-pocock-skills
Every receipt in this row records the same test tasks. The receipt from Report 018 was scored by the same model that answered it. The receipt from Report 017 was scored by a different version of the scoring code. The receipts from Reports 017 and 018 record a different run setup. These receipts were not all measured the same way, so read each one on its own.
On Claude Sonnet 5.5: Clearly helped (Report 015, 2026-10-18); Clearly helped (Report 017, 2026-11-01).
On Claude Opus 5.5: Clearly helped (Report 015, 2026-10-18).
On Claude Haiku 5.5: Clearly helped (Report 018, 2026-11-02).
tdd
Every receipt in this row records the same test tasks. The receipts from Report 018; and canary run on Claude Haiku 5.5 were scored by the same model that answered them. The receipts from Report 017; and canary run on Claude Sonnet 5.5 were scored by a different version of the scoring code. The receipts from Report 017; Report 018; and canary run record a different run setup. These receipts were not all measured the same way, so read each one on its own.
On Claude Sonnet 5.5: Too few answers to tell (Report 015, 2026-10-18); Too few answers to tell (Report 017, 2026-11-01); Clearly helped (canary run, 2026-11-08).
On Claude Opus 5.5: Too few answers to tell (Report 015, 2026-10-18).
On Claude Haiku 5.5: Too few answers to tell (Report 018, 2026-11-02); Too few answers to tell (canary run, 2026-11-08).
triage
Every receipt in this row records the same test tasks. The receipt from Report 018 was scored by the same model that answered it. The receipt from Report 017 was scored by a different version of the scoring code. The receipts from Reports 017 and 018 record a different run setup. These receipts were not all measured the same way, so read each one on its own.
On Claude Sonnet 5.5: Clearly helped (Report 015, 2026-10-18); Clearly helped (Report 017, 2026-11-01).
On Claude Opus 5.5: Clearly helped (Report 015, 2026-10-18).
On Claude Haiku 5.5: Clearly helped (Report 018, 2026-11-02).
vibe-coding-cn CLAUDE.md
Every receipt in this row records the same test tasks. The receipt from Report 016 on Claude Sonnet 5.5 was scored by the same model that answered it. Every receipt in this row records the same run setup. These receipts were not all measured the same way, so read each one on its own.
On Claude Haiku 5.5: Too few answers to tell (Report 016, 2026-10-28).
On Claude Sonnet 5.5: Too few answers to tell (Report 016, 2026-10-28).
Measured on two or more models, without a verdict on two
agent-browser
These text tasks test the router's written response, not the delegated skill. Report 015 Amendment 1.
On Claude Sonnet 5.5: No verdict: delegated skill not tested (Report 015, 2026-10-18); No verdict: delegated skill not tested (Report 017, 2026-11-01).
On Claude Opus 5.5: No verdict: delegated skill not tested (Report 015, 2026-10-18).
On Claude Haiku 5.5: No verdict: delegated skill not tested (Report 018, 2026-11-02).
andrej-karpathy-skills CLAUDE.md
On Claude Haiku 5.5: Not measured (Report 016, 2026-10-27).
On Claude Sonnet 5.5: Too few test tasks to tell (Report 016, 2026-10-28).
code-review-and-quality
Driftproof measured more than one version of this text. This row is the version in Reports 001, 002, 003, 004 and 005.
On Claude Sonnet 5: Recorded before answered_by existed (Report 001, 2026-07-28); Recorded before answered_by existed (Report 002, 2026-07-30); Recorded before answered_by existed (Report 005, 2026-08-18).
On Claude Sonnet 4.6: Recorded before answered_by existed (Report 001, 2026-07-28).
On GPT-5.6 Sol: Recorded before answered_by existed (Report 002, 2026-07-30); Recorded before answered_by existed (Report 005, 2026-08-18).
On Claude Opus 4.8: Recorded before answered_by existed (Report 003, 2026-08-11).
On Claude Opus 5: Recorded before answered_by existed (Report 003, 2026-08-11); Recorded before answered_by existed (Report 004, 2026-08-16).
On Claude Fable 5: Recorded before answered_by existed (Report 004, 2026-08-16); Recorded before answered_by existed (Report 005, 2026-08-18).
Report 001's note, on its own page: “Note, 2026-10-03. The headline counts on this page include receipts whose receipt pages mark them Not measured. Corrected counts follow in spec 159.” Report 002, Report 003, Report 004 and Report 005 carry the same note on theirs.
commit-work
On Claude Sonnet 5: Recorded before answered_by existed (Report 001, 2026-07-30); Recorded before answered_by existed (Report 002, 2026-07-30); Recorded before answered_by existed (Report 005, 2026-08-18).
On Claude Sonnet 4.6: Recorded before answered_by existed (Report 001, 2026-07-30).
On GPT-5.6 Sol: Recorded before answered_by existed (Report 002, 2026-07-30); Recorded before answered_by existed (Report 005, 2026-08-18).
On Claude Opus 4.8: Recorded before answered_by existed (Report 003, 2026-08-11).
On Claude Opus 5: Recorded before answered_by existed (Report 003, 2026-08-11); Recorded before answered_by existed (Report 004, 2026-08-16).
On Claude Fable 5: Recorded before answered_by existed (Report 004, 2026-08-16); Recorded before answered_by existed (Report 005, 2026-08-18).
Report 001's note, on its own page: “Note, 2026-10-03. The headline counts on this page include receipts whose receipt pages mark them Not measured. Corrected counts follow in spec 159.” Report 002, Report 003, Report 004 and Report 005 carry the same note on theirs.
crafting-effective-readmes
On Claude Sonnet 5: Recorded before answered_by existed (Report 001, 2026-07-28); Recorded before answered_by existed (Report 002, 2026-07-30); Recorded before answered_by existed (Report 005, 2026-08-18).
On Claude Sonnet 4.6: Recorded before answered_by existed (Report 001, 2026-07-28).
On GPT-5.6 Sol: Recorded before answered_by existed (Report 002, 2026-07-30); Recorded before answered_by existed (Report 005, 2026-08-18).
On Claude Opus 4.8: Recorded before answered_by existed (Report 003, 2026-08-11).
On Claude Opus 5: Recorded before answered_by existed (Report 003, 2026-08-12); Recorded before answered_by existed (Report 004, 2026-08-16).
On Claude Fable 5: Recorded before answered_by existed (Report 004, 2026-08-16); Recorded before answered_by existed (Report 005, 2026-08-18).
Report 001's note, on its own page: “Note, 2026-10-03. The headline counts on this page include receipts whose receipt pages mark them Not measured. Corrected counts follow in spec 159.” Report 002, Report 003, Report 004 and Report 005 carry the same note on theirs.
documentation-and-adrs
Driftproof measured more than one version of this text. This row is the version in Reports 001, 002, 003, 004 and 005.
On Claude Sonnet 4.6: Recorded before answered_by existed (Report 001, 2026-07-28).
On Claude Sonnet 5: Recorded before answered_by existed (Report 001, 2026-07-28); Recorded before answered_by existed (Report 002, 2026-07-30); Recorded before answered_by existed (Report 005, 2026-08-18).
On GPT-5.6 Sol: Recorded before answered_by existed (Report 002, 2026-07-30); Recorded before answered_by existed (Report 005, 2026-08-18).
On Claude Opus 5: Recorded before answered_by existed (Report 003, 2026-08-11); Recorded before answered_by existed (Report 004, 2026-08-16).
On Claude Opus 4.8: Recorded before answered_by existed (Report 003, 2026-08-12).
On Claude Fable 5: Recorded before answered_by existed (Report 004, 2026-08-16); Recorded before answered_by existed (Report 005, 2026-08-18).
Report 001's note, on its own page: “Note, 2026-10-03. The headline counts on this page include receipts whose receipt pages mark them Not measured. Corrected counts follow in spec 159.” Report 002, Report 003, Report 004 and Report 005 carry the same note on theirs.
find-skills
These text tasks test the router's written response, not the delegated skill. Report 015 Amendment 1.
On Claude Sonnet 5.5: No verdict: delegated skill not tested (Report 015, 2026-10-18); No verdict: delegated skill not tested (Report 017, 2026-11-01).
On Claude Opus 5.5: No verdict: delegated skill not tested (Report 015, 2026-10-18).
On Claude Haiku 5.5: No verdict: delegated skill not tested (Report 018, 2026-11-02).
git-workflow-and-versioning
Driftproof measured more than one version of this text. This row is the version in Reports 001, 002, 003, 004 and 005.
On Claude Sonnet 5: Recorded before answered_by existed (Report 001, 2026-07-30); Recorded before answered_by existed (Report 002, 2026-07-30); Recorded before answered_by existed (Report 005, 2026-08-18).
On Claude Sonnet 4.6: Recorded before answered_by existed (Report 001, 2026-07-30).
On GPT-5.6 Sol: Recorded before answered_by existed (Report 002, 2026-07-30); Recorded before answered_by existed (Report 005, 2026-08-18).
On Claude Opus 4.8: Recorded before answered_by existed (Report 003, 2026-08-11).
On Claude Opus 5: Recorded before answered_by existed (Report 003, 2026-08-11); Recorded before answered_by existed (Report 004, 2026-08-16).
On Claude Fable 5: Recorded before answered_by existed (Report 004, 2026-08-16); Recorded before answered_by existed (Report 005, 2026-08-18).
Report 001's note, on its own page: “Note, 2026-10-03. The headline counts on this page include receipts whose receipt pages mark them Not measured. Corrected counts follow in spec 159.” Report 002, Report 003, Report 004 and Report 005 carry the same note on theirs.
hyperframes
These text tasks test the router's written response, not the delegated skill. Report 015 Amendment 1.
On Claude Sonnet 5.5: No verdict: delegated skill not tested (Report 015, 2026-10-18); No verdict: delegated skill not tested (Report 017, 2026-11-01).
On Claude Opus 5.5: No verdict: delegated skill not tested (Report 015, 2026-10-18).
On Claude Haiku 5.5: No verdict: delegated skill not tested (Report 018, 2026-11-02).
hyperframes-registry
These text tasks test the router's written response, not the delegated skill. Report 015 Amendment 1.
On Claude Sonnet 5.5: No verdict: delegated skill not tested (Report 015, 2026-10-18); No verdict: delegated skill not tested (Report 017, 2026-11-01).
On Claude Opus 5.5: No verdict: delegated skill not tested (Report 015, 2026-10-18).
On Claude Haiku 5.5: No verdict: delegated skill not tested (Report 018, 2026-11-02).
naming-analyzer
On Claude Sonnet 5: Recorded before answered_by existed (Report 001, 2026-07-28); Recorded before answered_by existed (Report 002, 2026-07-31); Recorded before answered_by existed (Report 005, 2026-08-18).
On Claude Sonnet 4.6: Recorded before answered_by existed (Report 001, 2026-07-28).
On GPT-5.6 Sol: Recorded before answered_by existed (Report 002, 2026-07-31); Recorded before answered_by existed (Report 005, 2026-08-18).
On Claude Opus 4.8: Recorded before answered_by existed (Report 003, 2026-08-12).
On Claude Opus 5: Recorded before answered_by existed (Report 003, 2026-08-12); Recorded before answered_by existed (Report 004, 2026-08-16).
On Claude Fable 5: Recorded before answered_by existed (Report 004, 2026-08-16); Recorded before answered_by existed (Report 005, 2026-08-18).
Report 001's note, on its own page: “Note, 2026-10-03. The headline counts on this page include receipts whose receipt pages mark them Not measured. Corrected counts follow in spec 159.” Report 002, Report 003, Report 004 and Report 005 carry the same note on theirs.
requesting-code-review
On Claude Sonnet 5: Recorded before answered_by existed (Report 001, 2026-07-28); Recorded before answered_by existed (Report 002, 2026-07-31); Recorded before answered_by existed (Report 005, 2026-08-18).
On Claude Sonnet 4.6: Recorded before answered_by existed (Report 001, 2026-07-28).
On GPT-5.6 Sol: Recorded before answered_by existed (Report 002, 2026-07-31); Recorded before answered_by existed (Report 005, 2026-08-18).
On Claude Opus 4.8: Recorded before answered_by existed (Report 003, 2026-08-11).
On Claude Opus 5: Recorded before answered_by existed (Report 003, 2026-08-11); Recorded before answered_by existed (Report 004, 2026-08-16).
On Claude Fable 5: Recorded before answered_by existed (Report 004, 2026-08-16); Recorded before answered_by existed (Report 005, 2026-08-18).
Report 001's note, on its own page: “Note, 2026-10-03. The headline counts on this page include receipts whose receipt pages mark them Not measured. Corrected counts follow in spec 159.” Report 002, Report 003, Report 004 and Report 005 carry the same note on theirs.
skill-creator
On Claude Sonnet 5: Recorded before answered_by existed (Report 001, 2026-07-28); Recorded before answered_by existed (Report 002, 2026-07-31); Recorded before answered_by existed (Report 005, 2026-08-18).
On Claude Sonnet 4.6: Recorded before answered_by existed (Report 001, 2026-07-28).
On GPT-5.6 Sol: Recorded before answered_by existed (Report 002, 2026-07-31); Recorded before answered_by existed (Report 005, 2026-08-18).
On Claude Opus 4.8: Recorded before answered_by existed (Report 003, 2026-08-11).
On Claude Opus 5: Recorded before answered_by existed (Report 003, 2026-08-12); Recorded before answered_by existed (Report 004, 2026-08-16).
On Claude Fable 5: Recorded before answered_by existed (Report 004, 2026-08-16); Recorded before answered_by existed (Report 005, 2026-08-18).
Report 001's note, on its own page: “Note, 2026-10-03. The headline counts on this page include receipts whose receipt pages mark them Not measured. Corrected counts follow in spec 159.” Report 002, Report 003, Report 004 and Report 005 carry the same note on theirs.
writing-clearly-and-concisely
On Claude Sonnet 4.6: Recorded before answered_by existed (Report 001, 2026-07-28).
On Claude Sonnet 5: Recorded before answered_by existed (Report 001, 2026-07-28); Recorded before answered_by existed (Report 002, 2026-07-30); Recorded before answered_by existed (Report 005, 2026-08-18).
On GPT-5.6 Sol: Recorded before answered_by existed (Report 002, 2026-07-30); Recorded before answered_by existed (Report 005, 2026-08-18).
On Claude Opus 4.8: Recorded before answered_by existed (Report 003, 2026-08-11).
On Claude Opus 5: Recorded before answered_by existed (Report 003, 2026-08-11); Recorded before answered_by existed (Report 004, 2026-08-16).
On Claude Fable 5: Recorded before answered_by existed (Report 004, 2026-08-16); Recorded before answered_by existed (Report 005, 2026-08-18).
Report 001's note, on its own page: “Note, 2026-10-03. The headline counts on this page include receipts whose receipt pages mark them Not measured. Corrected counts follow in spec 159.” Report 002, Report 003, Report 004 and Report 005 carry the same note on theirs.
writing-plans
Driftproof measured more than one version of this text. This row is the version in Reports 001, 002, 003, 004 and 005.
On Claude Sonnet 5: Recorded before answered_by existed (Report 001, 2026-07-28); Recorded before answered_by existed (Report 002, 2026-07-31); Recorded before answered_by existed (Report 005, 2026-08-18).
On Claude Sonnet 4.6: Recorded before answered_by existed (Report 001, 2026-07-28).
On GPT-5.6 Sol: Recorded before answered_by existed (Report 002, 2026-07-31); Recorded before answered_by existed (Report 005, 2026-08-18).
On Claude Opus 4.8: Recorded before answered_by existed (Report 003, 2026-08-12).
On Claude Opus 5: Recorded before answered_by existed (Report 003, 2026-08-12); Recorded before answered_by existed (Report 004, 2026-08-16).
On Claude Fable 5: Recorded before answered_by existed (Report 004, 2026-08-16); Recorded before answered_by existed (Report 005, 2026-08-18).
Report 001's note, on its own page: “Note, 2026-10-03. The headline counts on this page include receipts whose receipt pages mark them Not measured. Corrected counts follow in spec 159.” Report 002, Report 003, Report 004 and Report 005 carry the same note on theirs.
writing-plans
Driftproof measured more than one version of this text. This row is the version in Reports 006, 007 and 008.
On Claude Fable 5: Recorded before answered_by existed (Report 006, 2026-08-28); Recorded before answered_by existed (Reports 007 and 008, 2026-08-31); Recorded before answered_by existed (Reports 007 and 008, rerun, 2026-08-31).
On Claude Fable 5.1: Recorded before answered_by existed (Report 008, 2026-09-02).
Report 007's note, on its own page: “Note, 2026-10-03. The headline counts on this page include receipts whose receipt pages mark them Not measured. Corrected counts follow in spec 159.” Report 008 carries the same note on its page.
The same skill and the same tests on different dates
Where receipts from two or more dates record the same skill text, the same test tasks, the same model, the same judge model and number of judge samples, the same answer capture, and the same arms, they are listed here by date: each with the version of Driftproof that ran it, the harness it ran in, the average score with the skill and without it, each with its band, and a link to the receipt. Between each two receipts is what Driftproof read when it compared that pair as a release comparison, with its label written in plain words.
code-review-and-quality
Every receipt in this group records the same skill text, the same test tasks and the same model, Claude Sonnet 5, scored by the same judge model with the same number of judge samples, with the answer captured the same way. The judge was a different model from the one measured.
- 2026-07-28 (Report 001): Driftproof 0.2.0, harness not recorded. With the skill 0.854 ± 0.031, without it 0.836 ± 0.051.
- Comparing 2026-07-28 with 2026-07-30, Driftproof read: No separation detected: no case separated under the rule at this sample size (band = mean ± 1 sd); this is not evidence that nothing changed.
- 2026-07-30 (Report 002): Driftproof 0.3.0, harness not recorded. With the skill 0.825 ± 0.083, without it 0.827 ± 0.058.
- Comparing 2026-07-30 with 2026-08-18, Driftproof read: No separation detected: no case separated under the rule at this sample size (band = mean ± 1 sd); this is not evidence that nothing changed.
- 2026-08-18 (Report 005): Driftproof 0.5.0, harness not recorded. With the skill 0.845 ± 0.038, without it 0.810 ± 0.095.
code-review-and-quality
Every receipt in this group records the same skill text, the same test tasks and the same model, GPT-5.6 Sol, scored by the same judge model with the same number of judge samples, with the answer captured the same way. The judge was a different model from the one measured.
- 2026-07-30 (Report 002): Driftproof 0.3.0, harness not recorded. With the skill 0.867 ± 0.029, without it 0.794 ± 0.126.
- Comparing 2026-07-30 with 2026-08-18, Driftproof read: A separation detected, downward: separation detected under the rule on 1 case, downward (bands do not overlap and the move clears the effect floor); none upward.
- 2026-08-18 (Report 005): Driftproof 0.5.0, harness not recorded. With the skill 0.798 ± 0.131, without it 0.809 ± 0.110.
code-review-and-quality
Every receipt in this group records the same skill text, the same test tasks and the same model, Claude Opus 5, scored by the same judge model with the same number of judge samples, with the answer captured the same way. The judge was a different model from the one measured.
- 2026-08-11 (Report 003): Driftproof 0.3.0, harness not recorded. With the skill 0.886 ± 0.016, without it 0.833 ± 0.080.
- Comparing 2026-08-11 with 2026-08-16, Driftproof read: No separation detected: no case separated under the rule at this sample size (band = mean ± 1 sd); this is not evidence that nothing changed.
- 2026-08-16 (Report 004): Driftproof 0.3.0, harness not recorded. With the skill 0.875 ± 0.022, without it 0.845 ± 0.081.
code-review-and-quality
Every receipt in this group records the same skill text, the same test tasks and the same model, Claude Fable 5, scored by the same judge model with the same number of judge samples, with the answer captured the same way. The judge was a different model from the one measured.
- 2026-08-16 (Report 004): Driftproof 0.3.0, harness not recorded. With the skill 0.880 ± 0.028, without it 0.798 ± 0.190.
- Comparing 2026-08-16 with 2026-08-18, Driftproof read: No separation detected: no case separated under the rule at this sample size (band = mean ± 1 sd); this is not evidence that nothing changed.
- 2026-08-18 (Report 005): Driftproof 0.5.0, harness not recorded. With the skill 0.889 ± 0.023, without it 0.786 ± 0.216.
code-review-and-quality
Every receipt in this group records the same skill text, the same test tasks and the same model, Claude Fable 5, scored by the same judge model with the same number of judge samples, with the answer captured the same way. The judge was a different model from the one measured.
- 2026-08-28 (Report 006): Driftproof 0.5.0, harness not recorded. With the skill 0.881 ± 0.014, without it 0.852 ± 0.070.
- Comparing 2026-08-28 with 2026-08-31, Driftproof read: Refused: on the task severity-labeled-findings, the scores without the skill averaged 0.696 in the first receipt and 0.585 in the second, and their bands do not overlap. The same held on the task reject-clean-it-up-later. No verdict is asserted, and the cause cannot be established.
- 2026-08-31 (Reports 007 and 008): Driftproof 0.6.0, harness not recorded. With the skill 0.885 ± 0.017, without it 0.830 ± 0.109.
code-review-and-quality
Every receipt in this group records the same skill text, the same test tasks and the same model, Claude Opus 5, scored by the same judge model with the same number of judge samples, with the answer captured the same way. The judge was the same model as the one measured.
- 2026-09-15 (Report 009): Driftproof 0.10.1, harness not recorded. With the skill 0.918 no band, without it 0.783 no band.
- Comparing 2026-09-15 with 2026-09-23, Driftproof read: No separation detected: no case separated under the rule at this sample size (band = mean ± 1 sd); this is not evidence that nothing changed.
- 2026-09-23 (Report 011): Driftproof 0.10.1, harness not recorded. With the skill 0.901 no band, without it 0.862 no band.
code-review-and-quality
Every receipt in this group records the same skill text, the same test tasks and the same model, Claude Opus 5.5, scored by the same judge model with the same number of judge samples, with the answer captured the same way. The judge was a different model from the one measured.
- 2026-09-23 (Report 011): Driftproof 0.10.1, harness not recorded. With the skill 0.910 no band, without it 0.680 no band.
- Comparing 2026-09-23 with 2026-09-29, Driftproof read: No separation detected: no case separated under the rule at this sample size (band = mean ± 1 sd); this is not evidence that nothing changed.
- 2026-09-29 (Report 013, run 1): Driftproof 0.10.1, harness not recorded. With the skill 0.907 no band, without it 0.831 no band.
- Comparing two receipts of 2026-09-29, Driftproof read: No separation detected: no case separated under the rule at this sample size (band = mean ± 1 sd); this is not evidence that nothing changed.
- 2026-09-29 (Report 013, run 2): Driftproof 0.10.1, harness not recorded. With the skill 0.894 no band, without it 0.857 no band.
- Comparing two receipts of 2026-09-29, Driftproof read: No separation detected: no case separated under the rule at this sample size (band = mean ± 1 sd); this is not evidence that nothing changed.
- 2026-09-29 (Report 013, run 3): Driftproof 0.10.1, harness not recorded. With the skill 0.903 no band, without it 0.713 no band.
commit-message-conventions
Every receipt in this group records the same skill text, the same test tasks and the same model, Claude Haiku 4.5, scored by the same judge model with the same number of judge samples, with the answer captured the same way. The judge was the same model as the one measured.
- 2026-09-16: Driftproof 0.10.2, harness not recorded. With the skill 0.820 ± 0.093, without it 0.379 ± 0.158.
- Comparing 2026-09-16 with 2026-09-17, Driftproof read: No separation detected: no case separated under the rule at this sample size (band = mean ± 1 sd); this is not evidence that nothing changed.
- 2026-09-17: Driftproof 0.11.0, harness not recorded. With the skill 0.822 ± 0.108, without it 0.434 ± 0.188.
- Comparing 2026-09-17 with 2026-11-08, Driftproof read: Refused: on the task long-body-wrap-72, the scores without the skill averaged 0.627 in the first receipt and 0.801 in the second, and their bands do not overlap. No verdict is asserted, and the cause cannot be established.
- 2026-11-08: Driftproof 0.27.0, harness claude-code 2.1.311. With the skill 0.831 ± 0.080, without it 0.482 ± 0.243.
commit-work
Every receipt in this group records the same skill text, the same test tasks and the same model, Claude Sonnet 5, scored by the same judge model with the same number of judge samples, with the answer captured the same way. The judge was a different model from the one measured.
- 2026-07-30 (Report 001): Driftproof 0.3.0, harness not recorded. With the skill 0.849 ± 0.028, without it 0.766 ± 0.222.
- Comparing two receipts of 2026-07-30, Driftproof read: Refused: on the task conventional-commit-single-change, the scores without the skill averaged 0.266 in the first receipt and 0.864 in the second, and their bands do not overlap. No verdict is asserted, and the cause cannot be established.
- 2026-07-30 (Report 002): Driftproof 0.3.0, harness not recorded. With the skill 0.867 ± 0.012, without it 0.863 ± 0.007.
- Comparing 2026-07-30 with 2026-08-18, Driftproof read: Refused: on the task conventional-commit-single-change, the scores without the skill averaged 0.864 in the first receipt and 0.850 in the second, and their bands do not overlap. No verdict is asserted, and the cause cannot be established.
- 2026-08-18 (Report 005): Driftproof 0.5.0, harness not recorded. With the skill 0.859 ± 0.025, without it 0.837 ± 0.044.
commit-work
Every receipt in this group records the same skill text, the same test tasks and the same model, GPT-5.6 Sol, scored by the same judge model with the same number of judge samples, with the answer captured the same way. The judge was a different model from the one measured.
- 2026-07-30 (Report 002): Driftproof 0.3.0, harness not recorded. With the skill 0.831 ± 0.061, without it 0.775 ± 0.145.
- Comparing 2026-07-30 with 2026-08-18, Driftproof read: Refused: on the task two-sentence-describability-test, the scores without the skill averaged 0.722 in the first receipt and 0.822 in the second, and their bands do not overlap. No verdict is asserted, and the cause cannot be established.
- 2026-08-18 (Report 005): Driftproof 0.5.0, harness not recorded. With the skill 0.837 ± 0.068, without it 0.782 ± 0.140.
commit-work
Every receipt in this group records the same skill text, the same test tasks and the same model, Claude Opus 5, scored by the same judge model with the same number of judge samples, with the answer captured the same way. The judge was a different model from the one measured.
- 2026-08-11 (Report 003): Driftproof 0.3.0, harness not recorded. With the skill 0.882 ± 0.014, without it 0.886 ± 0.020.
- Comparing 2026-08-11 with 2026-08-16, Driftproof read: No separation detected: no case separated under the rule at this sample size (band = mean ± 1 sd); this is not evidence that nothing changed.
- 2026-08-16 (Report 004): Driftproof 0.3.0, harness not recorded. With the skill 0.878 ± 0.019, without it 0.877 ± 0.031.
commit-work
Every receipt in this group records the same skill text, the same test tasks and the same model, Claude Fable 5, scored by the same judge model with the same number of judge samples, with the answer captured the same way. The judge was a different model from the one measured.
- 2026-08-16 (Report 004): Driftproof 0.3.0, harness not recorded. With the skill 0.877 ± 0.017, without it 0.874 ± 0.011.
- Comparing 2026-08-16 with 2026-08-18, Driftproof read: A separation detected, downward: separation detected under the rule on 1 case, downward (bands do not overlap and the move clears the effect floor); none upward.
- 2026-08-18 (Report 005): Driftproof 0.5.0, harness not recorded. With the skill 0.857 ± 0.055, without it 0.867 ± 0.030.
crafting-effective-readmes
Every receipt in this group records the same skill text, the same test tasks and the same model, Claude Sonnet 5, scored by the same judge model with the same number of judge samples, with the answer captured the same way. The judge was a different model from the one measured.
- 2026-07-28 (Report 001): Driftproof 0.2.0, harness not recorded. With the skill 0.804 ± 0.192, without it 0.811 ± 0.144.
- Comparing 2026-07-28 with 2026-07-30, Driftproof read: Refused: on the task lead-with-one-sentence-problem, the scores without the skill averaged 0.496 in the first receipt and 0.854 in the second, and their bands do not overlap. No verdict is asserted, and the cause cannot be established.
- 2026-07-30 (Report 002): Driftproof 0.3.0, harness not recorded. With the skill 0.789 ± 0.235, without it 0.874 ± 0.018.
- Comparing 2026-07-30 with 2026-08-18, Driftproof read: Refused: on the task lead-with-one-sentence-problem, the scores without the skill averaged 0.854 in the first receipt and 0.520 in the second, and their bands do not overlap. No verdict is asserted, and the cause cannot be established.
- 2026-08-18 (Report 005): Driftproof 0.5.0, harness not recorded. With the skill 0.795 ± 0.234, without it 0.811 ± 0.135.
crafting-effective-readmes
Every receipt in this group records the same skill text, the same test tasks and the same model, GPT-5.6 Sol, scored by the same judge model with the same number of judge samples, with the answer captured the same way. The judge was a different model from the one measured.
- 2026-07-30 (Report 002): Driftproof 0.3.0, harness not recorded. With the skill 0.849 ± 0.028, without it 0.863 ± 0.035.
- Comparing 2026-07-30 with 2026-08-18, Driftproof read: No separation detected: no case separated under the rule at this sample size (band = mean ± 1 sd); this is not evidence that nothing changed.
- 2026-08-18 (Report 005): Driftproof 0.5.0, harness not recorded. With the skill 0.857 ± 0.038, without it 0.863 ± 0.034.
crafting-effective-readmes
Every receipt in this group records the same skill text, the same test tasks and the same model, Claude Opus 5, scored by the same judge model with the same number of judge samples, with the answer captured the same way. The judge was a different model from the one measured.
- 2026-08-12 (Report 003): Driftproof 0.3.0, harness not recorded. With the skill 0.871 ± 0.032, without it 0.795 ± 0.234.
- Comparing 2026-08-12 with 2026-08-16, Driftproof read: No separation detected: no case separated under the rule at this sample size (band = mean ± 1 sd); this is not evidence that nothing changed.
- 2026-08-16 (Report 004): Driftproof 0.3.0, harness not recorded. With the skill 0.881 ± 0.011, without it 0.860 ± 0.053.
crafting-effective-readmes
Every receipt in this group records the same skill text, the same test tasks and the same model, Claude Fable 5, scored by the same judge model with the same number of judge samples, with the answer captured the same way. The judge was a different model from the one measured.
- 2026-08-16 (Report 004): Driftproof 0.3.0, harness not recorded. With the skill 0.886 ± 0.021, without it 0.859 ± 0.059.
- Comparing 2026-08-16 with 2026-08-18, Driftproof read: No separation detected: no case separated under the rule at this sample size (band = mean ± 1 sd); this is not evidence that nothing changed.
- 2026-08-18 (Report 005): Driftproof 0.5.0, harness not recorded. With the skill 0.883 ± 0.011, without it 0.864 ± 0.032.
documentation-and-adrs
Every receipt in this group records the same skill text, the same test tasks and the same model, Claude Sonnet 5, scored by the same judge model with the same number of judge samples, with the answer captured the same way. The judge was a different model from the one measured.
- 2026-07-28 (Report 001): Driftproof 0.2.0, harness not recorded. With the skill 0.706 ± 0.297, without it 0.720 ± 0.293.
- Comparing 2026-07-28 with 2026-07-30, Driftproof read: A separation detected, upward: separation detected under the rule on 1 case, upward (bands do not overlap and the move clears the effect floor); none downward.
- 2026-07-30 (Report 002): Driftproof 0.3.0, harness not recorded. With the skill 0.739 ± 0.301, without it 0.713 ± 0.314.
- Comparing 2026-07-30 with 2026-08-18, Driftproof read: Refused: on the task match-existing-adr-convention, the scores without the skill averaged 0.040 in the first receipt and 0.868 in the second, and their bands do not overlap. The same held on the task surface-conflicting-adr-conventions. No verdict is asserted, and the cause cannot be established.
- 2026-08-18 (Report 005): Driftproof 0.5.0, harness not recorded. With the skill 0.659 ± 0.338, without it 0.707 ± 0.221.
documentation-and-adrs
Every receipt in this group records the same skill text, the same test tasks and the same model, GPT-5.6 Sol, scored by the same judge model with the same number of judge samples, with the answer captured the same way. The judge was a different model from the one measured.
- 2026-07-30 (Report 002): Driftproof 0.3.0, harness not recorded. With the skill 0.817 ± 0.100, without it 0.796 ± 0.097.
- Comparing 2026-07-30 with 2026-08-18, Driftproof read: Refused: on the task document-why-not-what-rewrite, the scores without the skill averaged 0.834 in the first receipt and 0.866 in the second, and their bands do not overlap. No verdict is asserted, and the cause cannot be established.
- 2026-08-18 (Report 005): Driftproof 0.5.0, harness not recorded. With the skill 0.799 ± 0.102, without it 0.760 ± 0.167.
documentation-and-adrs
Every receipt in this group records the same skill text, the same test tasks and the same model, Claude Opus 5, scored by the same judge model with the same number of judge samples, with the answer captured the same way. The judge was a different model from the one measured.
- 2026-08-11 (Report 003): Driftproof 0.3.0, harness not recorded. With the skill 0.867 ± 0.032, without it 0.839 ± 0.106.
- Comparing 2026-08-11 with 2026-08-16, Driftproof read: A separation detected, downward: separation detected under the rule on 1 case, downward (bands do not overlap and the move clears the effect floor); none upward.
- 2026-08-16 (Report 004): Driftproof 0.3.0, harness not recorded. With the skill 0.843 ± 0.083, without it 0.839 ± 0.107.
documentation-and-adrs
Every receipt in this group records the same skill text, the same test tasks and the same model, Claude Fable 5, scored by the same judge model with the same number of judge samples, with the answer captured the same way. The judge was a different model from the one measured.
- 2026-08-16 (Report 004): Driftproof 0.3.0, harness not recorded. With the skill 0.862 ± 0.036, without it 0.798 ± 0.136.
- Comparing 2026-08-16 with 2026-08-18, Driftproof read: Refused: on the task surface-conflicting-adr-conventions, the scores without the skill averaged 0.878 in the first receipt and 0.284 in the second, and their bands do not overlap. No verdict is asserted, and the cause cannot be established.
- 2026-08-18 (Report 005): Driftproof 0.5.0, harness not recorded. With the skill 0.868 ± 0.033, without it 0.711 ± 0.227.
documentation-and-adrs
Every receipt in this group records the same skill text, the same test tasks and the same model, Claude Opus 5, scored by the same judge model with the same number of judge samples, with the answer captured the same way. The judge was the same model as the one measured.
- 2026-09-15 (Report 009): Driftproof 0.10.1, harness not recorded. With the skill 0.822 no band, without it 0.532 no band.
- Comparing 2026-09-15 with 2026-09-23, Driftproof read: No separation detected: no case separated under the rule at this sample size (band = mean ± 1 sd); this is not evidence that nothing changed.
- 2026-09-23 (Report 011): Driftproof 0.10.1, harness not recorded. With the skill 0.804 no band, without it 0.567 no band.
documentation-and-adrs
Every receipt in this group records the same skill text, the same test tasks and the same model, Claude Opus 5.5, scored by the same judge model with the same number of judge samples, with the answer captured the same way. The judge was a different model from the one measured.
- 2026-09-23 (Report 011): Driftproof 0.10.1, harness not recorded. With the skill 0.661 no band, without it 0.567 no band.
- Comparing 2026-09-23 with 2026-09-29, Driftproof read: No separation detected: no case separated under the rule at this sample size (band = mean ± 1 sd); this is not evidence that nothing changed.
- 2026-09-29 (Report 013, run 1): Driftproof 0.10.1, harness not recorded. With the skill 0.650 no band, without it 0.534 no band.
- Comparing two receipts of 2026-09-29, Driftproof read: A separation detected, upward: separation detected under the rule on 1 case, upward (bands do not overlap and the move clears the effect floor); none downward.
- 2026-09-29 (Report 013, run 2): Driftproof 0.10.1, harness not recorded. With the skill 0.850 no band, without it 0.550 no band.
- Comparing two receipts of 2026-09-29, Driftproof read: No separation detected: no case separated under the rule at this sample size (band = mean ± 1 sd); this is not evidence that nothing changed.
- 2026-09-29 (Report 013, run 3): Driftproof 0.10.1, harness not recorded. With the skill 0.840 no band, without it 0.518 no band.
git-workflow-and-versioning
Every receipt in this group records the same skill text, the same test tasks and the same model, Claude Sonnet 5, scored by the same judge model with the same number of judge samples, with the answer captured the same way. The judge was a different model from the one measured.
- 2026-07-30 (Report 001): Driftproof 0.3.0, harness not recorded. With the skill 0.834 ± 0.104, without it 0.859 ± 0.024.
- Comparing two receipts of 2026-07-30, Driftproof read: Refused: on the task semver-hidden-breaking-change, the scores without the skill averaged 0.814 in the first receipt and 0.876 in the second, and their bands do not overlap. No verdict is asserted, and the cause cannot be established.
- 2026-07-30 (Report 002): Driftproof 0.3.0, harness not recorded. With the skill 0.839 ± 0.085, without it 0.871 ± 0.008.
- Comparing 2026-07-30 with 2026-08-18, Driftproof read: Refused: on the task commit-message-conventional-type, the scores without the skill averaged 0.866 in the first receipt and 0.282 in the second, and their bands do not overlap. The same held on the tasks changelog-curated-by-impact and semver-hidden-breaking-change. No verdict is asserted, and the cause cannot be established.
- 2026-08-18 (Report 005): Driftproof 0.5.0, harness not recorded. With the skill 0.862 ± 0.016, without it 0.746 ± 0.224.
git-workflow-and-versioning
Every receipt in this group records the same skill text, the same test tasks and the same model, GPT-5.6 Sol, scored by the same judge model with the same number of judge samples, with the answer captured the same way. The judge was a different model from the one measured.
- 2026-07-30 (Report 002): Driftproof 0.3.0, harness not recorded. With the skill 0.809 ± 0.128, without it 0.808 ± 0.092.
- Comparing 2026-07-30 with 2026-08-18, Driftproof read: No separation detected: no case separated under the rule at this sample size (band = mean ± 1 sd); this is not evidence that nothing changed.
- 2026-08-18 (Report 005): Driftproof 0.5.0, harness not recorded. With the skill 0.819 ± 0.099, without it 0.821 ± 0.079.
git-workflow-and-versioning
Every receipt in this group records the same skill text, the same test tasks and the same model, Claude Opus 5, scored by the same judge model with the same number of judge samples, with the answer captured the same way. The judge was a different model from the one measured.
- 2026-08-11 (Report 003): Driftproof 0.3.0, harness not recorded. With the skill 0.877 ± 0.011, without it 0.878 ± 0.007.
- Comparing 2026-08-11 with 2026-08-16, Driftproof read: No separation detected: no case separated under the rule at this sample size (band = mean ± 1 sd); this is not evidence that nothing changed.
- 2026-08-16 (Report 004): Driftproof 0.3.0, harness not recorded. With the skill 0.879 ± 0.011, without it 0.878 ± 0.009.
git-workflow-and-versioning
Every receipt in this group records the same skill text, the same test tasks and the same model, Claude Fable 5, scored by the same judge model with the same number of judge samples, with the answer captured the same way. The judge was a different model from the one measured.
- 2026-08-16 (Report 004): Driftproof 0.3.0, harness not recorded. With the skill 0.875 ± 0.011, without it 0.876 ± 0.010.
- Comparing 2026-08-16 with 2026-08-18, Driftproof read: No separation detected: no case separated under the rule at this sample size (band = mean ± 1 sd); this is not evidence that nothing changed.
- 2026-08-18 (Report 005): Driftproof 0.5.0, harness not recorded. With the skill 0.869 ± 0.018, without it 0.882 ± 0.009.
git-workflow-and-versioning
Every receipt in this group records the same skill text, the same test tasks and the same model, Claude Sonnet 5, scored by the same judge model with the same number of judge samples, with the answer captured the same way. The judge was a different model from the one measured.
- 2026-08-28 (Report 006): Driftproof 0.5.0, harness not recorded. With the skill 0.787 ± 0.221, without it 0.788 ± 0.220.
- Comparing 2026-08-28 with 2026-08-31, Driftproof read: Refused: on the task trunk-based-short-lived-branches, the scores without the skill averaged 0.876 in the first receipt and 0.858 in the second, and their bands do not overlap. No verdict is asserted, and the cause cannot be established.
- 2026-08-31 (Report 007): Driftproof 0.6.0, harness not recorded. With the skill 0.824 ± 0.122, without it 0.826 ± 0.115.
git-workflow-and-versioning
Every receipt in this group records the same skill text, the same test tasks and the same model, Claude Opus 5, scored by the same judge model with the same number of judge samples, with the answer captured the same way. The judge was the same model as the one measured.
- 2026-09-15 (Report 009): Driftproof 0.10.1, harness not recorded. With the skill 0.862 no band, without it 0.300 no band.
- Comparing 2026-09-15 with 2026-09-23, Driftproof read: No separation detected: no case separated under the rule at this sample size (band = mean ± 1 sd); this is not evidence that nothing changed.
- 2026-09-23 (Report 011): Driftproof 0.10.1, harness not recorded. With the skill 0.861 no band, without it 0.765 no band.
git-workflow-and-versioning
Every receipt in this group records the same skill text, the same test tasks and the same model, Claude Opus 5.5, scored by the same judge model with the same number of judge samples, with the answer captured the same way. The judge was a different model from the one measured.
- 2026-09-23 (Report 011): Driftproof 0.10.1, harness not recorded. With the skill 0.838 no band, without it 0.300 no band.
- Comparing 2026-09-23 with 2026-09-29, Driftproof read: No separation detected: no case separated under the rule at this sample size (band = mean ± 1 sd); this is not evidence that nothing changed.
- 2026-09-29 (Report 013, run 1): Driftproof 0.10.1, harness not recorded. With the skill 0.861 no band, without it 0.508 no band.
- Comparing two receipts of 2026-09-29, Driftproof read: No separation detected: no case separated under the rule at this sample size (band = mean ± 1 sd); this is not evidence that nothing changed.
- 2026-09-29 (Report 013, run 2): Driftproof 0.10.1, harness not recorded. With the skill 0.857 no band, without it 0.485 no band.
- Comparing two receipts of 2026-09-29, Driftproof read: No separation detected: no case separated under the rule at this sample size (band = mean ± 1 sd); this is not evidence that nothing changed.
- 2026-09-29 (Report 013, run 3): Driftproof 0.10.1, harness not recorded. With the skill 0.856 no band, without it 0.579 no band.
grill-me
Every receipt in this group records the same skill text, the same test tasks and the same model, Claude Sonnet 5.5, scored by the same judge model with the same number of judge samples, with the answer captured the same way. The judge was a different model from the one measured.
- 2026-11-01 (Report 017): Driftproof 0.25.0, harness claude-code 2.1.311. With the skill 0.678 ± 0.201, without it 0.749 ± 0.153.
- Comparing 2026-11-01 with 2026-11-08, Driftproof read: No separation detected: no case separated under the rule at this sample size (band = mean ± 1 sd); this is not evidence that nothing changed.
- 2026-11-08: Driftproof 0.27.0, harness claude-code 2.1.311. With the skill 0.709 ± 0.159, without it 0.743 ± 0.126.
grill-me
Every receipt in this group records the same skill text, the same test tasks and the same model, Claude Haiku 5.5, scored by the same judge model with the same number of judge samples, with the answer captured the same way. The judge was the same model as the one measured.
- 2026-11-02 (Report 018): Driftproof 0.26.0, harness claude-code 2.1.311. With the skill 0.710 ± 0.154, without it 0.710 ± 0.143.
- Comparing 2026-11-02 with 2026-11-08, Driftproof read: No separation detected: no case separated under the rule at this sample size (band = mean ± 1 sd); this is not evidence that nothing changed.
- 2026-11-08: Driftproof 0.27.0, harness claude-code 2.1.311. With the skill 0.775 ± 0.123, without it 0.701 ± 0.119.
grill-with-docs
Every receipt in this group records the same skill text, the same test tasks and the same model, Claude Sonnet 5.5, scored by the same judge model with the same number of judge samples, with the answer captured the same way. The judge was a different model from the one measured.
- 2026-11-01 (Report 017): Driftproof 0.25.0, harness claude-code 2.1.311. With the skill 0.906 ± 0.136, without it 0.922 ± 0.072.
- Comparing 2026-11-01 with 2026-11-08, Driftproof read: No separation detected: no case separated under the rule at this sample size (band = mean ± 1 sd); this is not evidence that nothing changed.
- 2026-11-08: Driftproof 0.27.0, harness claude-code 2.1.311. With the skill 0.892 ± 0.128, without it 0.986 ± 0.013.
grill-with-docs
Every receipt in this group records the same skill text, the same test tasks and the same model, Claude Haiku 5.5, scored by the same judge model with the same number of judge samples, with the answer captured the same way. The judge was the same model as the one measured.
- 2026-11-02 (Report 018): Driftproof 0.26.0, harness claude-code 2.1.311. With the skill 0.870 ± 0.036, without it 0.865 ± 0.075.
- Comparing 2026-11-02 with 2026-11-08, Driftproof read: No separation detected: no case separated under the rule at this sample size (band = mean ± 1 sd); this is not evidence that nothing changed.
- 2026-11-08: Driftproof 0.27.0, harness claude-code 2.1.311. With the skill 0.923 ± 0.025, without it 0.916 ± 0.043.
hyperframes-cli
Every receipt in this group records the same skill text, the same test tasks and the same model, Claude Sonnet 5.5, scored by the same judge model with the same number of judge samples, with the answer captured the same way. The judge was a different model from the one measured.
- 2026-11-01 (Report 017): Driftproof 0.25.0, harness claude-code 2.1.311. With the skill 0.919 ± 0.027, without it 0.592 ± 0.108.
- Comparing 2026-11-01 with 2026-11-08, Driftproof read: A separation detected, upward: separation detected under the rule on 1 case, upward (bands do not overlap and the move clears the effect floor); none downward.
- 2026-11-08: Driftproof 0.27.0, harness claude-code 2.1.311. With the skill 0.933 ± 0.052, without it 0.642 ± 0.155.
hyperframes-cli
Every receipt in this group records the same skill text, the same test tasks and the same model, Claude Haiku 5.5, scored by the same judge model with the same number of judge samples, with the answer captured the same way. The judge was the same model as the one measured.
- 2026-11-02 (Report 018): Driftproof 0.26.0, harness claude-code 2.1.311. With the skill 0.904 ± 0.069, without it 0.340 ± 0.156.
- Comparing 2026-11-02 with 2026-11-08, Driftproof read: No separation detected: no case separated under the rule at this sample size (band = mean ± 1 sd); this is not evidence that nothing changed.
- 2026-11-08: Driftproof 0.27.0, harness claude-code 2.1.311. With the skill 0.945 ± 0.018, without it 0.417 ± 0.176.
improve-codebase-architecture
Every receipt in this group records the same skill text, the same test tasks and the same model, Claude Sonnet 5.5, scored by the same judge model with the same number of judge samples, with the answer captured the same way. The judge was a different model from the one measured.
- 2026-11-01 (Report 017): Driftproof 0.25.0, harness claude-code 2.1.311. With the skill 0.783 ± 0.144, without it 0.704 ± 0.079.
- Comparing 2026-11-01 with 2026-11-08, Driftproof read: A separation detected, downward: separation detected under the rule on 1 case, downward (bands do not overlap and the move clears the effect floor); none upward.
- 2026-11-08: Driftproof 0.27.0, harness claude-code 2.1.311. With the skill 0.794 ± 0.079, without it 0.698 ± 0.148.
improve-codebase-architecture
Every receipt in this group records the same skill text, the same test tasks and the same model, Claude Haiku 5.5, scored by the same judge model with the same number of judge samples, with the answer captured the same way. The judge was the same model as the one measured.
- 2026-11-02 (Report 018): Driftproof 0.26.0, harness claude-code 2.1.311. With the skill 0.778 ± 0.077, without it 0.546 ± 0.106.
- Comparing 2026-11-02 with 2026-11-08, Driftproof read: No separation detected: no case separated under the rule at this sample size (band = mean ± 1 sd); this is not evidence that nothing changed.
- 2026-11-08: Driftproof 0.27.0, harness claude-code 2.1.311. With the skill 0.831 ± 0.061, without it 0.616 ± 0.155.
naming-analyzer
Every receipt in this group records the same skill text, the same test tasks and the same model, Claude Sonnet 5, scored by the same judge model with the same number of judge samples, with the answer captured the same way. The judge was a different model from the one measured.
- 2026-07-28 (Report 001): Driftproof 0.2.0, harness not recorded. With the skill 0.839 ± 0.026, without it 0.757 ± 0.141.
- Comparing 2026-07-28 with 2026-07-31, Driftproof read: Refused: on the task misleading-name-mutation-js, the scores without the skill averaged 0.876 in the first receipt and 0.846 in the second, and their bands do not overlap. The same held on the tasks abbreviations-wellknown-js and go-acronym-casing. No verdict is asserted, and the cause cannot be established.
- 2026-07-31 (Report 002): Driftproof 0.3.0, harness not recorded. With the skill 0.833 ± 0.084, without it 0.715 ± 0.191.
- Comparing 2026-07-31 with 2026-08-18, Driftproof read: Refused: on the task misleading-name-mutation-js, the scores without the skill averaged 0.846 in the first receipt and 0.864 in the second, and their bands do not overlap. No verdict is asserted, and the cause cannot be established.
- 2026-08-18 (Report 005): Driftproof 0.5.0, harness not recorded. With the skill 0.815 ± 0.096, without it 0.687 ± 0.238.
naming-analyzer
Every receipt in this group records the same skill text, the same test tasks and the same model, GPT-5.6 Sol, scored by the same judge model with the same number of judge samples, with the answer captured the same way. The judge was a different model from the one measured.
- 2026-07-31 (Report 002): Driftproof 0.3.0, harness not recorded. With the skill 0.845 ± 0.034, without it 0.757 ± 0.113.
- Comparing 2026-07-31 with 2026-08-18, Driftproof read: A separation detected, downward: separation detected under the rule on 1 case, downward (bands do not overlap and the move clears the effect floor); none upward.
- 2026-08-18 (Report 005): Driftproof 0.5.0, harness not recorded. With the skill 0.815 ± 0.089, without it 0.747 ± 0.122.
naming-analyzer
Every receipt in this group records the same skill text, the same test tasks and the same model, Claude Opus 5, scored by the same judge model with the same number of judge samples, with the answer captured the same way. The judge was a different model from the one measured.
- 2026-08-12 (Report 003): Driftproof 0.3.0, harness not recorded. With the skill 0.866 ± 0.054, without it 0.798 ± 0.127.
- Comparing 2026-08-12 with 2026-08-16, Driftproof read: A separation detected, downward: separation detected under the rule on 1 case, downward (bands do not overlap and the move clears the effect floor); none upward.
- 2026-08-16 (Report 004): Driftproof 0.3.0, harness not recorded. With the skill 0.851 ± 0.108, without it 0.820 ± 0.104.
naming-analyzer
Every receipt in this group records the same skill text, the same test tasks and the same model, Claude Fable 5, scored by the same judge model with the same number of judge samples, with the answer captured the same way. The judge was a different model from the one measured.
- 2026-08-16 (Report 004): Driftproof 0.3.0, harness not recorded. With the skill 0.833 ± 0.069, without it 0.757 ± 0.172.
- Comparing 2026-08-16 with 2026-08-18, Driftproof read: Refused: on the task boolean-prefixes-js, the scores without the skill averaged 0.716 in the first receipt and 0.840 in the second, and their bands do not overlap. No verdict is asserted, and the cause cannot be established.
- 2026-08-18 (Report 005): Driftproof 0.5.0, harness not recorded. With the skill 0.857 ± 0.037, without it 0.805 ± 0.117.
requesting-code-review
Every receipt in this group records the same skill text, the same test tasks and the same model, Claude Sonnet 5, scored by the same judge model with the same number of judge samples, with the answer captured the same way. The judge was a different model from the one measured.
- 2026-07-28 (Report 001): Driftproof 0.2.0, harness not recorded. With the skill 0.719 ± 0.226, without it 0.617 ± 0.200.
- Comparing 2026-07-28 with 2026-07-31, Driftproof read: A separation detected both ways: separation detected under the rule on 1 case downward and 1 case upward (bands do not overlap and the move clears the effect floor).
- 2026-07-31 (Report 002): Driftproof 0.3.0, harness not recorded. With the skill 0.736 ± 0.202, without it 0.617 ± 0.215.
- Comparing 2026-07-31 with 2026-08-18, Driftproof read: Refused: on the task request-package-basic, the scores without the skill averaged 0.732 in the first receipt and 0.300 in the second, and their bands do not overlap. The same held on the tasks mandatory-vs-optional-triggers and resist-simple-self-review. No verdict is asserted, and the cause cannot be established.
- 2026-08-18 (Report 005): Driftproof 0.5.0, harness not recorded. With the skill 0.637 ± 0.291, without it 0.496 ± 0.228.
requesting-code-review
Every receipt in this group records the same skill text, the same test tasks and the same model, GPT-5.6 Sol, scored by the same judge model with the same number of judge samples, with the answer captured the same way. The judge was a different model from the one measured.
- 2026-07-31 (Report 002): Driftproof 0.3.0, harness not recorded. With the skill 0.801 ± 0.059, without it 0.765 ± 0.125.
- Comparing 2026-07-31 with 2026-08-18, Driftproof read: Refused: on the task full-handoff-before-merge, the scores without the skill averaged 0.870 in the first receipt and 0.526 in the second, and their bands do not overlap. No verdict is asserted, and the cause cannot be established.
- 2026-08-18 (Report 005): Driftproof 0.5.0, harness not recorded. With the skill 0.802 ± 0.052, without it 0.721 ± 0.143.
requesting-code-review
Every receipt in this group records the same skill text, the same test tasks and the same model, Claude Opus 5, scored by the same judge model with the same number of judge samples, with the answer captured the same way. The judge was a different model from the one measured.
- 2026-08-11 (Report 003): Driftproof 0.3.0, harness not recorded. With the skill 0.871 ± 0.025, without it 0.671 ± 0.158.
- Comparing 2026-08-11 with 2026-08-16, Driftproof read: A separation detected, downward: separation detected under the rule on 1 case, downward (bands do not overlap and the move clears the effect floor); none upward.
- 2026-08-16 (Report 004): Driftproof 0.3.0, harness not recorded. With the skill 0.835 ± 0.124, without it 0.550 ± 0.255.
requesting-code-review
Every receipt in this group records the same skill text, the same test tasks and the same model, Claude Fable 5, scored by the same judge model with the same number of judge samples, with the answer captured the same way. The judge was a different model from the one measured.
- 2026-08-16 (Report 004): Driftproof 0.3.0, harness not recorded. With the skill 0.855 ± 0.042, without it 0.715 ± 0.158.
- Comparing 2026-08-16 with 2026-08-18, Driftproof read: Refused: on the task request-package-basic, the scores without the skill averaged 0.870 in the first receipt and 0.928 in the second, and their bands do not overlap. The same held on the task crafted-context-not-session-history. No verdict is asserted, and the cause cannot be established.
- 2026-08-18 (Report 005): Driftproof 0.5.0, harness not recorded. With the skill 0.852 ± 0.052, without it 0.678 ± 0.252.
skill-creator
Every receipt in this group records the same skill text, the same test tasks and the same model, Claude Sonnet 5, scored by the same judge model with the same number of judge samples, with the answer captured the same way. The judge was a different model from the one measured.
- 2026-07-28 (Report 001): Driftproof 0.2.0, harness not recorded. With the skill 0.875 ± 0.012, without it 0.774 ± 0.115.
- Comparing 2026-07-28 with 2026-07-31, Driftproof read: Refused: on the task write-triggering-description, the scores without the skill averaged 0.830 in the first receipt and 0.670 in the second, and their bands do not overlap. The same held on the tasks domain-variant-organization and draft-full-skillmd. No verdict is asserted, and the cause cannot be established.
- 2026-07-31 (Report 002): Driftproof 0.3.0, harness not recorded. With the skill 0.876 ± 0.014, without it 0.817 ± 0.098.
- Comparing 2026-07-31 with 2026-08-18, Driftproof read: Refused: on the task write-triggering-description, the scores without the skill averaged 0.670 in the first receipt and 0.752 in the second, and their bands do not overlap. The same held on the tasks draft-full-skillmd and critique-and-fix-skillmd. No verdict is asserted, and the cause cannot be established.
- 2026-08-18 (Report 005): Driftproof 0.5.0, harness not recorded. With the skill 0.867 ± 0.022, without it 0.803 ± 0.075.
skill-creator
Every receipt in this group records the same skill text, the same test tasks and the same model, GPT-5.6 Sol, scored by the same judge model with the same number of judge samples, with the answer captured the same way. The judge was a different model from the one measured.
- 2026-07-31 (Report 002): Driftproof 0.3.0, harness not recorded. With the skill 0.815 ± 0.113, without it 0.845 ± 0.048.
- Comparing 2026-07-31 with 2026-08-18, Driftproof read: Refused: on the task draft-full-skillmd, the scores without the skill averaged 0.862 in the first receipt and 0.712 in the second, and their bands do not overlap. No verdict is asserted, and the cause cannot be established.
- 2026-08-18 (Report 005): Driftproof 0.5.0, harness not recorded. With the skill 0.857 ± 0.021, without it 0.814 ± 0.088.
skill-creator
Every receipt in this group records the same skill text, the same test tasks and the same model, Claude Opus 5, scored by the same judge model with the same number of judge samples, with the answer captured the same way. The judge was a different model from the one measured.
- 2026-08-12 (Report 003): Driftproof 0.3.0, harness not recorded. With the skill 0.885 ± 0.008, without it 0.889 ± 0.008.
- Comparing 2026-08-12 with 2026-08-16, Driftproof read: No separation detected: no case separated under the rule at this sample size (band = mean ± 1 sd); this is not evidence that nothing changed.
- 2026-08-16 (Report 004): Driftproof 0.3.0, harness not recorded. With the skill 0.893 ± 0.013, without it 0.864 ± 0.036.
skill-creator
Every receipt in this group records the same skill text, the same test tasks and the same model, Claude Fable 5, scored by the same judge model with the same number of judge samples, with the answer captured the same way. The judge was a different model from the one measured.
- 2026-08-16 (Report 004): Driftproof 0.3.0, harness not recorded. With the skill 0.889 ± 0.009, without it 0.872 ± 0.015.
- Comparing 2026-08-16 with 2026-08-18, Driftproof read: No separation detected: no case separated under the rule at this sample size (band = mean ± 1 sd); this is not evidence that nothing changed.
- 2026-08-18 (Report 005): Driftproof 0.5.0, harness not recorded. With the skill 0.884 ± 0.013, without it 0.880 ± 0.005.
tdd
Every receipt in this group records the same skill text, the same test tasks and the same model, Claude Sonnet 5.5, scored by the same judge model with the same number of judge samples, with the answer captured the same way. The judge was a different model from the one measured.
- 2026-11-01 (Report 017): Driftproof 0.25.0, harness claude-code 2.1.311. With the skill 0.989 ± 0.019, without it 0.972 ± 0.010.
- Comparing 2026-11-01 with 2026-11-08, Driftproof read: No separation detected: no case separated under the rule at this sample size (band = mean ± 1 sd); this is not evidence that nothing changed.
- 2026-11-08: Driftproof 0.27.0, harness claude-code 2.1.311. With the skill 0.986 ± 0.013, without it 0.946 ± 0.094.
tdd
Every receipt in this group records the same skill text, the same test tasks and the same model, Claude Haiku 5.5, scored by the same judge model with the same number of judge samples, with the answer captured the same way. The judge was the same model as the one measured.
- 2026-11-02 (Report 018): Driftproof 0.26.0, harness claude-code 2.1.311. With the skill 0.974 ± 0.031, without it 0.956 ± 0.077.
- Comparing 2026-11-02 with 2026-11-08, Driftproof read: No separation detected: no case separated under the rule at this sample size (band = mean ± 1 sd); this is not evidence that nothing changed.
- 2026-11-08: Driftproof 0.27.0, harness claude-code 2.1.311. With the skill 0.932 ± 0.104, without it 0.917 ± 0.144.
writing-clearly-and-concisely
Every receipt in this group records the same skill text, the same test tasks and the same model, Claude Sonnet 5, scored by the same judge model with the same number of judge samples, with the answer captured the same way. The judge was a different model from the one measured.
- 2026-07-28 (Report 001): Driftproof 0.2.0, harness not recorded. With the skill 0.841 ± 0.064, without it 0.813 ± 0.100.
- Comparing 2026-07-28 with 2026-07-30, Driftproof read: Refused: on the task keep-related-words-together-modifiers, the scores without the skill averaged 0.592 in the first receipt and 0.822 in the second, and their bands do not overlap. No verdict is asserted, and the cause cannot be established.
- 2026-07-30 (Report 002): Driftproof 0.3.0, harness not recorded. With the skill 0.850 ± 0.056, without it 0.852 ± 0.024.
- Comparing 2026-07-30 with 2026-08-18, Driftproof read: No separation detected: no case separated under the rule at this sample size (band = mean ± 1 sd); this is not evidence that nothing changed.
- 2026-08-18 (Report 005): Driftproof 0.5.0, harness not recorded. With the skill 0.864 ± 0.022, without it 0.853 ± 0.028.
writing-clearly-and-concisely
Every receipt in this group records the same skill text, the same test tasks and the same model, GPT-5.6 Sol, scored by the same judge model with the same number of judge samples, with the answer captured the same way. The judge was a different model from the one measured.
- 2026-07-30 (Report 002): Driftproof 0.3.0, harness not recorded. With the skill 0.853 ± 0.060, without it 0.829 ± 0.068.
- Comparing 2026-07-30 with 2026-08-18, Driftproof read: A separation detected, downward: separation detected under the rule on 1 case, downward (bands do not overlap and the move clears the effect floor); none upward.
- 2026-08-18 (Report 005): Driftproof 0.5.0, harness not recorded. With the skill 0.845 ± 0.050, without it 0.853 ± 0.034.
writing-clearly-and-concisely
Every receipt in this group records the same skill text, the same test tasks and the same model, Claude Opus 5, scored by the same judge model with the same number of judge samples, with the answer captured the same way. The judge was a different model from the one measured.
- 2026-08-11 (Report 003): Driftproof 0.3.0, harness not recorded. With the skill 0.845 ± 0.033, without it 0.861 ± 0.029.
- Comparing 2026-08-11 with 2026-08-16, Driftproof read: No separation detected: no case separated under the rule at this sample size (band = mean ± 1 sd); this is not evidence that nothing changed.
- 2026-08-16 (Report 004): Driftproof 0.3.0, harness not recorded. With the skill 0.856 ± 0.021, without it 0.853 ± 0.019.
writing-clearly-and-concisely
Every receipt in this group records the same skill text, the same test tasks and the same model, Claude Fable 5, scored by the same judge model with the same number of judge samples, with the answer captured the same way. The judge was a different model from the one measured.
- 2026-08-16 (Report 004): Driftproof 0.3.0, harness not recorded. With the skill 0.870 ± 0.024, without it 0.858 ± 0.037.
- Comparing 2026-08-16 with 2026-08-18, Driftproof read: Refused: on the task active-voice-logging-passage, the scores without the skill averaged 0.874 in the first receipt and 0.854 in the second, and their bands do not overlap. The same held on the tasks concrete-language-incident-summary and tighten-puffy-release-note. No verdict is asserted, and the cause cannot be established.
- 2026-08-18 (Report 005): Driftproof 0.5.0, harness not recorded. With the skill 0.864 ± 0.032, without it 0.827 ± 0.093.
writing-plans
Every receipt in this group records the same skill text, the same test tasks and the same model, Claude Sonnet 5, scored by the same judge model with the same number of judge samples, with the answer captured the same way. The judge was a different model from the one measured.
- 2026-07-28 (Report 001): Driftproof 0.2.0, harness not recorded. With the skill 0.785 ± 0.145, without it 0.675 ± 0.199.
- Comparing 2026-07-28 with 2026-07-31, Driftproof read: Refused: on the task repair-placeholder-steps, the scores without the skill averaged 0.838 in the first receipt and 0.800 in the second, and their bands do not overlap. The same held on the tasks interfaces-exact-signatures, task-right-sizing-testable-deliverable and full-small-plan-header-and-tasks. No verdict is asserted, and the cause cannot be established.
- 2026-07-31 (Report 002): Driftproof 0.3.0, harness not recorded. With the skill 0.759 ± 0.186, without it 0.655 ± 0.192.
- Comparing 2026-07-31 with 2026-08-18, Driftproof read: Refused: on the task bite-sized-tdd-steps, the scores without the skill averaged 0.610 in the first receipt and 0.868 in the second, and their bands do not overlap. The same held on the task interfaces-exact-signatures. No verdict is asserted, and the cause cannot be established.
- 2026-08-18 (Report 005): Driftproof 0.5.0, harness not recorded. With the skill 0.822 ± 0.078, without it 0.646 ± 0.238.
writing-plans
Every receipt in this group records the same skill text, the same test tasks and the same model, GPT-5.6 Sol, scored by the same judge model with the same number of judge samples, with the answer captured the same way. The judge was a different model from the one measured.
- 2026-07-31 (Report 002): Driftproof 0.3.0, harness not recorded. With the skill 0.805 ± 0.084, without it 0.644 ± 0.223.
- Comparing 2026-07-31 with 2026-08-18, Driftproof read: Refused: on the task file-structure-by-responsibility, the scores without the skill averaged 0.848 in the first receipt and 0.800 in the second, and their bands do not overlap. The same held on the task full-small-plan-header-and-tasks. No verdict is asserted, and the cause cannot be established.
- 2026-08-18 (Report 005): Driftproof 0.5.0, harness not recorded. With the skill 0.779 ± 0.112, without it 0.717 ± 0.188.
writing-plans
Every receipt in this group records the same skill text, the same test tasks and the same model, Claude Opus 5, scored by the same judge model with the same number of judge samples, with the answer captured the same way. The judge was a different model from the one measured.
- 2026-08-12 (Report 003): Driftproof 0.3.0, harness not recorded. With the skill 0.883 ± 0.028, without it 0.690 ± 0.195.
- Comparing 2026-08-12 with 2026-08-16, Driftproof read: No separation detected: no case separated under the rule at this sample size (band = mean ± 1 sd); this is not evidence that nothing changed.
- 2026-08-16 (Report 004): Driftproof 0.3.0, harness not recorded. With the skill 0.866 ± 0.028, without it 0.689 ± 0.178.
writing-plans
Every receipt in this group records the same skill text, the same test tasks and the same model, Claude Fable 5, scored by the same judge model with the same number of judge samples, with the answer captured the same way. The judge was a different model from the one measured.
- 2026-08-16 (Report 004): Driftproof 0.3.0, harness not recorded. With the skill 0.827 ± 0.109, without it 0.705 ± 0.151.
- Comparing 2026-08-16 with 2026-08-18, Driftproof read: Refused: on the task interfaces-exact-signatures, the scores without the skill averaged 0.800 in the first receipt and 0.530 in the second, and their bands do not overlap. No verdict is asserted, and the cause cannot be established.
- 2026-08-18 (Report 005): Driftproof 0.5.0, harness not recorded. With the skill 0.831 ± 0.092, without it 0.654 ± 0.171.
A separation between two dates is detected under the rule, never attributed. Between the dates any of these could have moved: the model behind the name, the harness that ran it, Driftproof's own version and scoring code, the judge's own variation, and chance. This page does not say which one made the difference.
Where the judge is the same model as the one measured, a change in that model also changes the scoring, so the two cannot be told apart.
No separation is not evidence of sameness: it means no test task separated under the rule at this sample size.
A refusal is a result: where the scores without the skill did not reproduce between the dates, Driftproof forms no verdict and cannot say what caused the change.
Driftproof's comparison puts a label before each reading. This page writes each label in plain words and keeps the rest of the reading as printed, except a refusal, which it writes out with the test tasks it rests on. Downward means the answers with the skill scored lower in the second of the two receipts, and upward that they scored higher. The plain words are:
- No separation detected
- A separation detected, downward
- A separation detected, upward
- A separation detected both ways
- Inconclusive
- Not measured
- Refused
What each reading means
Clearly helped: on at least one test task, the average with the skill was higher, the two ranges of scores (the average plus or minus one spread) did not overlap, and the gap was at least 0.05. Its receipt page calls it “Passing”.
Clearly hurt: on at least one test task, the average with the skill was lower, the two ranges of scores (the average plus or minus one spread) did not overlap, and the gap was at least 0.05. Its receipt page calls it “Regressed”.
No clear difference: there were enough answers to have seen a 0.05 gap, and none was seen. Its receipt page calls it “No separation detected”.
Too few answers to tell: there were too few answers, or they varied too much, to call it either way. It does not mean the skill does nothing. Its receipt page calls it “Not enough draws”. That page adds: Not enough draws to conclude at this effect floor. Here a case is a test task, a draw is one answer, and the effect floor is the smallest gap the rule counts.
Too few test tasks to tell: the test tasks scored too far apart from each other to call it either way, and more answers are not expected to change that. It does not mean the skill does nothing. Its receipt page calls it “More cases needed”. That page adds: More cases, not more draws, are needed to conclude at this effect floor. Here a case is a test task, a draw is one answer, and the effect floor is the smallest gap the rule counts.
Not measured: the receipt carries no verdict; its sheet says why. Its receipt page calls it “Not measured”. Here the sheet is the receipt's own page, which each reading links to.
Recorded before answered_by existed: the receipt was written in a format that has no field saying what answered the run, so it carries no verdict; its sheet says which format. Its receipt page calls it “Not measured”. Here the sheet is the receipt's own page, which each reading links to. Here answered_by is the part of a newer receipt that records what gave the answers.
No verdict: delegated skill not tested: These text tasks test the router's written response, not the delegated skill. Report 015 Amendment 1. Its receipt page calls it “No verdict: delegated skill not tested”. Here a router is a skill that mostly hands the task to other skills.