Release drift report
Report 014: Claude Haiku 5.5 on release day, three skills
claude-haiku-4-5 claude-haiku-5-5
2026-10-07
Claude Haiku 5.5 on the three skills and cases of Reports 009, 011 and 013, beside Claude Haiku 4.5, each run three times.
Reports
Every report Driftproof has published, newest first. 13 reports, 179 receipts, all of them dated and hash-verified.
Release drift report
Report 013: Claude Sonnet 5.5, the day after its release
claude-opus-5-5 claude-sonnet-5 claude-sonnet-5-5
2026-09-29
With the skill, Claude Sonnet 5.5 was never clearly higher or lower than Claude Opus 5.5 in nine results, and five of them had too few answers to tell.
Release drift report
Report 011: Claude Opus 5.5, the day after its release, three skills
claude-opus-5 claude-opus-5-5
2026-09-23
Claude Opus 5.5 on the three skills and cases of Report 009, one case each, beside a fresh Claude Opus 5 arm on the same harness.
Instrument comparison report
Report 010: repeated ADR evaluations across two Claude Code configurations
claude-opus-5
2026-09-21
One skill’s behavioural eval from its own repository, run ten times with the same skill text, the same eval case and the same executor model, five times under each of two Claude Code configurations.
Instrument comparison report
Report 009: three skills under two eval harnesses
claude-opus-5
2026-09-15
Three skills from one plugin, one case each, measured by Claude Code’s native plugin eval and by Driftproof, with the same SKILL.md bytes, task prompts and rubrics.
Release drift report
Report 008: the skill stabilises the floor, not the ceiling
claude-fable-5-1
2026-09-02
No separation was detected under the rule in the with_skill arm of either cell across the release.
Instrument re-measurement report
Report 007: the baseline is the noisy arm
claude-fable-5 claude-sonnet-5
2026-08-31
Three skills re-measured under generation sampling, with a corrected instrument.
Revision drift report
Report 006: the skill text moves, the substrate holds still
claude-fable-5 claude-sonnet-5
2026-08-28
Three cells, one per skill upstream had revised since Report #005 pinned it, and none of the three returned a verdict.
Value report
Report 005: what does a skill cost to run?
claude-fable-5 claude-sonnet-5 gpt-5.6-sol
2026-08-18
In 20 of 30 skill × substrate pairs at least one case cleared the effect floor with separated bands: 18 with an improving case (5 of them also with a regressing one), 2 with only a regressing case.
Capability-gap report
Report 004: does encoded expertise still lift output on the frontier tier?
claude-fable-5 claude-opus-5
2026-08-16
3 durable · 5 tier-dependent · 0 regresses · 2 no effect
Release drift report
Report 003: do skill verdicts hold across a model release?
claude-opus-4-8 claude-opus-5
2026-08-12
2 regressed · 4 improved · 0 mixed · 4 with no separation detected over 10 skills.
Substrate durability report
Report 002: does a skill's benefit hold across substrates?
claude-sonnet-5 gpt-5.6-sol
2026-07-31
Cross-provider skill durability, not a model ranking.
Release drift report
Report 001: do skill verdicts hold across a model release?
claude-sonnet-4-6 claude-sonnet-5
2026-07-30
9 of 10 skills showed a separation detected under the rule on this model pair.
The seven report types
All seven use the same rule: a verdict is claimed only when the two ranges do not overlap and the move clears the floor. They differ in what moves underneath the skill, or, in the value report, in which axes are measured.
Release drift report
Asks whether a skill still helps when a newer version of the model underneath it comes out. Example: Report 014.
Instrument comparison report
Keeps the skills and their tasks the same and measures them with two instruments, two different test tools, saying what each can and cannot show. Example: Report 010.
Instrument re-measurement report
Changes neither, and runs again tests an earlier report published, with a corrected way of measuring. Example: Report 007.
Revision drift report
Keeps the model and the tools the same and changes only the skill's own text. Example: Report 006.
Value report
Keeps the model the same and asks what a skill costs to run, beside whether it helps. Example: Report 005.
Capability-gap report
Asks whether a skill still helps on a stronger model from the same company. Example: Report 004.
Substrate durability report
Asks whether a skill helps just as much on a different company's model, reached through that company's own command-line tool. Example: Report 002.
Writing
What happens to agent skills when the model changes: the launch essay, revised to read all thirteen reports together on what a moving model does to encoded expertise, and on what a corrected instrument did to three published results.