Driftproof

Reports

Every report Driftproof has published, newest first. 8 reports, 123 receipts, all of them dated and hash-verified.

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.

Band plot for the first cell of Report 005.

The six report types

All six 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's verdict holds when the model version underneath it changes. Example: Report 008.

Substrate durability report

Asks whether it holds across a change of substrate: a different model behind a different vendor CLI. Example: Report 002.

Capability-gap report

Asks whether it holds on a higher tier of the same provider. Example: Report 004.

Value report

Holds the substrate still and asks what a skill costs to run, alongside whether it helps. Example: Report 005.

Revision drift report

Holds everything underneath still and moves the skill's own text. Example: Report 006.

Instrument re-measurement report

Moves neither, and re-runs cells an earlier report published on a corrected measuring device. Example: Report 007.

Writing

Three model releases later: what actually happens to agent skills: the launch essay, revised to read all eight reports together on what a moving model does to encoded expertise, and on what a corrected instrument did to three published results.