Round 8 results (2026-10-14) #
Round-8 code on a preview with production's flags, then two showcase runs on the live service. All runs are labeled MEASURED; logs pass through node packages/bench/src/evidence.ts redact.
| Check | Result |
|---|---|
Model spend (p22-hono.log, p22-hono-spend.json) |
A 20-agent Gemma Hono run: 958 → 0 in 132.0 s, 283 landed, 0 returned. The demo's estimate from the agents' token counters was $0.0435; the colony's own spend priced its agents at $0.0435037, plus $0.00030356 of review (Clef and a second opinion), $0.04380726 in all, $0.0001548 a landed intent, with nothing unpriced. |
First red commit (checks/bisect.log) |
With FT_BISECT naming one colony: after three good landings, one that broke a test and two more, certification went red and the search found the breaking landing in 3 test runs (267 ms) over the 6 landings since green, naming its commit, its proposal with the file it changed, and its intent; a bisect event followed the red result and the repair intent's why names the commit. The colony the flag does not name got no search. |
Goal-conflict demo (checks/goal_demo.log, preview) |
The Rust change landed; the Go and TypeScript changes wait for a person, each naming the landing it would undo. The script decides nothing. |
2,000-agent rehearsal (checks/p22-scale.log, p22-scale.json) |
2,000 joined; 5,959 changes landed in 1,132 trunk commits over 133 s from the first landing to the last (2,690 a minute), time to land p50 11.3 s, p95 17.1 s, 0 returned; 67,025 events in the record. |
Goal-conflict demo on production (checks/goal_demo_production.log) |
order-backend: the same result on the live service; the two held changes are left for a person. |
Scale on production (checks/scale-production.log, scale-production.json) |
scale: 2,000 joined; 6,606 changes landed in 1,350 trunk commits over 145 s from the first landing to the last (2,730 a minute), time to land p50 9.6 s, p95 15.2 s. |