Start hereTry itAgentsArchitectureMechanismsResultsGlossaryRelated workSpikeChangelogContributingSecurity

Round 11 results (2026-10-14)

Round-11 code on a preview with production's flags, and read-only checks of the live service. All runs are labeled MEASURED; logs pass through node packages/bench/src/evidence.ts redact.

Check Result
First red commit on a long history (checks/bisect40.log, checks/bisect100.log) scripts/bisect.ts --landings 40: the 40 landings since green landed in 111.4 s, the search found the breaking landing in 6 test runs (432 ms), the fewest possible (ceil(log2 41) = 6). --landings 100 --break-at 7: 100 landings in 198.8 s, found in 7 runs (375 ms), the fewest possible. The "off" halves of both checks report a search too, because the preview had FT_BISECT on for every colony, as production does; they are not failures of the search.
Tours (checks/tour_first-red.log, checks/tour_order-backend.log, checks/tour_p25_first-red.log) Against the live service: the first-red tour's 14 numbers and the order-backend tour's 13 agree with /certifications, /review, the snapshot and /metrics (hono's 10 unchanged); on the preview, the first-red tour on a 40-landing colony agrees on all 12.
Signed-notes card (checks/signed_notes.log) On the preview with a test key: the signed check passes (a landing intact and signed by the key, a stock clone verifying it, a forged rewrite reading bad signature), and the card reads "Signed by ", the key's since-date, "newest landing fc6c417: note intact and signed" and the stock-git line. The colony's first creation answered 500 twice while a 40-landing run was landing on the fresh preview; the third try, after that run, created it at once.
Try panel (checks/tryit.log) /#try on the preview: four scenes render and the default merge scene prints the trunk writer's decision ("Lands as a line-by-line merge: trunk moved src/limits.ts …"); no horizontal scroll.
Spend per intent (checks/p25-hono2.log, checks/intent_spend.log) With FT_INTENT_SPEND on and quoted acceptance off on the preview: a 20-agent Gemma Hono run, 958 → 0 in 113.4 s, 283 landed, 0 returned, $0.046 as the colony priced it. scripts/intent_spend_check.rb: 25 s after the run the sum over intents was $0.043 (the colony reads 16 journals per pass, 15 s apart); minutes later $0.046, a gap of 0; a sampled intent shows $0.0001 in one agent call and nothing in review.
Quoted acceptance and absence lines (checks/p25-hono.log) With quoted acceptance on, the first Hono demo on the preview hung: its prerequisite intent carries "No other option, file or formatting changes", the second opinion could quote no changed line for an absence, the line counted unclear and the change waited for a person. Fixed: a line stating an absence that the model says is met counts on the model's word, recorded as such; with the fix and the flag on (checks/p25-hono3.log) the same demo ran through, 958 → 0 in 118.2 s, 283 landed, 0 returned.

In this folder