# To Mike — final note

**Final score: 8.5/10** (final proposal in `proposal.md`; peak was 9/10 at Round 9).

**Why the loop stopped:** I ran 12 blind ask-colleague rounds (11 optimization iterations over your baseline), each rated by a fresh subagent that could not see prior scores. The last five ratings were 9 → 8.5 → 8.5 → 8.5 → 8.5 under a strict 10/10 bar (sign-off as-is, no reservations): the needle has been flat for five consecutive turns, so per your stopping rule I stopped. 10/10 was not hit — and after Round 2 the rater was already strict, so this is not a leniency artifact. Simulated Dan's remaining reservations are consistently operability (nine gates on a small team, no worked example block with hours yet) and residual precision leak (exact 100-point scores below the 20-point threshold), plus the pilot's named owner/reviewer are still placeholders for you to fill in before sending.

**Does the proposal still carry your concern?** Yes — it carries it, transformed. Your original form of the concern (no benchmark under 250 tasks, or researchers won't take it seriously) is not in the final proposal: the 250 gate is explicitly non-binding, and researcher acceptance is explicitly not the success criterion. What survives, and is stronger, is the core worry about **false precision**: it is now carried by the publication format itself (per-task-only scores, no aggregate or leaderboard with an on-block invalid-aggregation warning, actual task/tester/run counts printed at the score, calibration and disagreement shown, 30-day revealed use with self-report-vs-log labelling, point-scores suppressed on high-disagreement/single-run tasks, and a coverage claim an independent internal reviewer can fail). Mike's concern won on substance; the 250-task threshold and the researcher-credibility framing were traded away as the mechanism, because the persona's evidence says neither is what makes a small set credible.
