Literature-scan results (2026-10-14) #
Checks of the mechanisms a literature triage suggested, on a preview with production's flags plus the new flags named per colony. All runs are labeled MEASURED; logs pass through node packages/bench/src/evidence.ts redact.
| Check | Result |
|---|---|
Contest judge (checks/bias_judge.log) |
scripts/contest_bias.ts --contests 20: two synthetic agents propose equal-evidence variants of one change (a one-line arrow function against a three-line function), the order alternating; the judge is asked in both orders. Orders agreed on a side 0 times; in 18 contests the judge saw no clear difference in both orders, in 2 it preferred whichever was shown first; the deterministic tie-break decided 20 of 20. Settle time median 1.9 s, max 2.4 s; 40 calls, $0.0028 ($0.0001 a contest). |
Contest jury (checks/bias_jury.log) |
With FT_CONTEST_JURY naming the colony (three small models from different families, each asked in both orders at once): a majority decided 7 of 20, the tie-break 13; of 60 member seats, 19 were inconsistent (17 positional, 2 late); settle time median 4.4 s, max 45 s; 120 calls, $0.0124 ($0.0006 a contest). Ties fell from 100% to 65% but not by half, settling took longer, and on equal variants a tie may be the right answer: the jury stays off in production. |
Who declared an intent, and the chain gate (checks/intent_chain.log) |
scripts/intent_chain.ts: on the colony FT_INTENT_CHAIN_GATE names, agent A declared intent 1 and took it, declared intent 2 while working (depth 1), agent B took it and its take said A declared it (model, depth, the note); B declared intent 3 (depth 2), which was held: the declare answer said so, GET /escalations listed it with kind chain, a take did not offer it, a person's dismissal released it and C took it with the declarer shown; with the fan-out set to 2, A's third declaration of the hour was held the same way. On the colony the flag does not name, every declaration was taken at once and every take still showed who declared it. 40 checks pass. |