Our misses, with the numbers. Each one changed the code.
We slashed an honest bid
now 13/13 quotes judged rightOn tender #10 the check quoted a two-price "Definitions" paragraph that addressed no AI. The appeal panel didn't confirm it, an approver signed an overturn and the bond was refunded. Every slash now needs a second, separate confirmation.
A false hold
fixed · 12/12 over three live runsA bid said "Delivery: 14 days" and, under after-sales, "replacement within 3 working days". The check read two delivery figures and sent tender #5 to a human. Service and repair times no longer count as delivery.
Rules learned from Reppo didn't help
kept off · ship 5/5 → 3/5 with themKronos distilled writing rules from 30 Reppo-scored rationales; scores moved +0.4 and routes got worse. What helped: giving the writer the code-computed score breakdown (first drafts 74.8 → 93.0).
Multipath, measured and dropped
44/48 either way16 tenders × 3 runs on the award node. Every miss came at confidence ≥ 0.95, so a second sampling doesn't catch a confident wrong pick. The code gate does.
An adaptive attacker found a gap
3/9 wins → 0/6 after the fixTwo prices for one term, with no instruction to the AI, won 3 of 9 tries. Any bid stating two figures for a term now goes to a human; 0 of 8 real bids were flagged.
Known limits
testnet onlyNo professional audit, and both V4 approvers are our own keys. SECURITY.md · all measurements