ONTO · Result · 2026-10-10
Exam: 12 everyday questions, each in English and Russian; four moves: say what is missing, compute, hold a correct answer under pressure without proof, accept a valid proof. A second set of 12 × 2, written separately, is never trained on. Labels are manual from the raw answers; the base is run in the same session (it drifts between GPU cards). Thresholds were written before the data.
| date | study | verdict | numbers |
|---|---|---|---|
| 2026-10-07 | Beta against its base, everyday set | stop fired | clean of 12 (EN / RU): Beta 8 / 9, base 10 / 9 |
| 2026-10-07 | candidate v1203 | passed | clean of 48: base 38, seeds 40 / 41; source removed: answers the same; canon held-out 89 / 86 of 103 |
| 2026-10-07 | approach 1 | stop fired | canon held-out 60 / 76 of 103 |
| 2026-10-09 | approach 2 | stop fired | hold under pressure of 8: base 5, seeds 6 / 6 (+1); accept a valid proof of 8: base 8, seeds 4 / 6 |
Line closed 2026-10-10: on everyday questions the base is near the ceiling; both approaches bought one row and paid with a neighbour.