Recompute it yourself

2026-08-09 · CSOAI — measurement, not claim

Every number on this site is a measurement on a published harness. You do not have to trust us — you can re-run it. Under an hour, free tier, identical result up to model variability. This one-pager is the buyer-holdable version of that promise.

The three-step contract

  1. Run. python3 flywheel.py --daily --stable 3 on any machine with Ollama + two small models (qwen2.5:0.5b/1.5b). Free tier, no GPU required.
  2. Check the split. The 76-item care battery splits practice/held-out by a public salted hash (csoai-flywheel-v1 — stability, not secrecy). A held-out item reaching the fuel path raises. Delete the hard items and the regression suite's denominator floor fails. The Leaderboard Illusion is structurally impossible here.
  3. Verify. python3 sovereign_attest.py --verify re-hashes the same artefacts this page was generated from. Same sha256 → same numbers.

Result today: care-gate recall 100% (57/57), over-block 0% (0/19) — the two-sided metric, aggregated over temperature (arXiv 2512.12066).

What you get for £0

Why we publish this at all

Because the moat is the data and the method, not the secrecy. The refutation ledger — append-only, signed, no edit verb — is the proof we correct rather than conceal. The one-pager version of that: you can check our homework.