Main idea

Nothing changes in the skill and its spending rises anyway. The dataset goes from 88 to 138 analyses, and every phase that reads it through costs more. The instruction file grows by half over successive adjustments, and that supplement is paid on every call.

On competitor_analyze, two implemented optimizations removed 186,000 units, and data growth took back 71,000 over the same period. An audited cost is therefore dated: it is a measurement taken on one scenario, at one moment, not a property of the skill. Hence the three moments when the audit is run again — at deployment, when the ceiling is hit faster than before, and periodically on the heaviest skill.

Why it matters

This prevents treating an optimization as settled: a realized gain can be absorbed without anyone noticing, and the skill becomes the prime suspect again a few months later.

It also makes the audit a repeated operation rather than a one-off project.

Nuances and limits

The drift is not always suffered: a skill that costs more because it processes more has not regressed. What degrades is the cost per unit of useful work, and the total does not show it.

Open questions

• Which indicator, brought back to the unit of work, would distinguish a drift from a simple rise in load?