saver-audit measures the caveman engine on users' own logs: method check?

[VladUZH]

Hi! I built saver-audit (npx saver-audit, MIT, fully offline). It reads a developer's local Claude Code and Codex logs, shows where their tokens went (priced with prompt-cache costs), and replays recorded tool outputs through token savers using the user's own installed copy. It never bundles your code (BSL respected).

• caveman engine (bin 1.1.7): every recorded tool output of 500+ tokens goes through the engine, as your proxy would send it. On my last 30 days: $13.32 of $4,209 (0.3%), labelled "replayed". Outputs under 500 tokens count as unchanged; that leaves out about 5%.
• caveman skill: you don't publish a measurement, so it's "modeled" from JetBrains' −8.5% output figure plus the SKILL.md tokens in every prompt: about 2.3%.

Code: https://github.com/VladUZH/saver-audit/blob/main/src/savers/builtin/caveman-engine.json

Is that a fair way to run the engine? And if you have your own measurement of the skill's output cut, I'd rather cite that than JetBrains'. I'm sharing results publicly this week and will update them if you spot a problem.