Owned talk models, scored on frozen floors. Updated from the ledger. (champion 336M (B.20, 2026-08-24))
Each point is the model that was PROMOTED that day, scored on the same floors. B.10 predates the floors, so it has no points — it was chosen on held-out loss, which B.13 later showed ranks models backwards. Between B.13 and B.15 the champion's clean fell 0.96 to 0.92 while input-sensitivity rose 0.90 to 0.98: the corpus grew and its MIX changed with it, and this is the only view in which that shows up at all.
| model | params | held-out loss | status |
|---|---|---|---|
| 46M-eot-merged | 46.48M | 2.173 | |
| 336M | 336.39M | 3.184 | live |
B.20 — identical corpus (214,743 pairs) and identical budget; only the parameter count changes. Loss and the floors disagree, which is the second sweep running where selecting on loss would ship the wrong model. ood is 0.00 for every model ever trained here — none signals confusion on out-of-distribution input, and doubling the corpus did not change it.
The trust curve is not shown, because it has not been measured. The self-improvement gate writes a trust-curve log as it promotes and rolls back checkpoints; no such log exists yet, so there is nothing to plot. An earlier version of this page displayed a hardcoded point — 4/19 correct, honesty 100% — that came from a literal in the build script rather than from any run. It is removed. A measurement that has not been taken is shown as absent, not as a number.