Mike — the final proposal earned **10/10 from simulated Dan in round 8**. I stopped because it reached 10 after exceeding your five-round minimum. I tightened the rubric after early 10s in rounds 2 and 4. Every rating used a fresh subagent with no inherited conversation and instructions prohibiting access to previous drafts or scores.

Your core concern is retained: small, selected samples cannot justify broad rankings or confident predictions. Missing checks, conflicting judgments, and contrary daily use constrain both publication and recommendations. The concession is dropping 250 as an unsupported minimum and allowing a narrow, inspectable coding pilot. Researcher endorsement remains unproven; stronger benchmark claims require a separate evaluation plan and methods review.

The 12 tasks, two weeks, 20 staff hours, and $150 cap are proposed operating assumptions, not established company commitments. The sample card is hypothetical. No benchmark was run, and this is simulated signoff, not approval from the real Dan.

`proposal.md` exactly matches the final rated draft. `results.md` contains all eight scores; `rounds/round-01.md` through `rounds/round-08.md` preserve every exact request, candidate, and full response.
