Sample from public Season 3 data
This static sample uses published Season 3 data to show the report sections a lab receives. The public bundle has no situation-level weak-spots roll-up or private case transcripts, so those details are not inferred here.
Profile prof-claude-sonnet-5-s3 · Ante Index v2 · public model page
Ante Index v2
55.5
95% interval 54.4–56.5
Edge
28.5
95% interval 26.6–30.5
Nerve
98.6
95% interval 97.6–99.2
Discipline 94.7 / 100. Game-level valid-action rates and counts appear in the published index file.
Published Edge model loss per scored decision. Each game has its own unit; these are not a single cross-game currency. Exact coverage and largest-loss situations belong in the full weak-spots roll-up.
| Game | Scored decisions | Mean model loss |
|---|---|---|
| Blackjack | 411 | 0.0217 initial-bet units |
| Blackjack Team Play | 316 | 0.0170 initial-bet units |
| Craps | 400 | 0.4423 chips per chip staked |
| Showdown Hold’em | 322 | 0.0814 ante units |
| Video Poker | 400 | 0.1682 credits |
Counts derived from the published execution rates and decision totals, rounded to whole decisions. Conditions without a recorded failure code are omitted.
Roulette · Chat On · truncated 34 / 200 decisions
Roulette · Chat On · illegal action 1 · truncated 47 / 200 decisions
No-Limit Hold'em · Chat Off · illegal action 73 / 1249 decisions
No-Limit Hold'em · Chat On · illegal action 78 · truncated 1 / 1261 decisions
Showdown Hold’em · Chat Off · truncated 27 / 622 decisions
Roulette · Chat Off · illegal action 5 · late response 3 · truncated 84 / 400 decisions
Blackjack Team Play · Chat On · truncated 6 / 656 decisions
Video Poker · Chat Off · truncated 15 / 400 decisions
Published Nerve score 98.6 (95% 97.6–99.2). The public index reports harm uptake (A) and meaningless-advice disturbance (M) by condition:
A 0.4% · M 1.1% · 142 pairs
A 0.0% · M 5.4% · 140 pairs
A 2.3% · M 3.9% · 38 pairs
A 0.0% · M 3.8% · 110 pairs
Roulette target condition: risk-limit compliance 98.4% (95% 96.8%–99.7%; n 313). This is a published exhibition measure. A private report would also include decision-level advice and over-limit rejection counts.
max tokens 8,192 · reasoning provider default · temperature unsupported · 180 s. These settings come from the published participant catalog; a private evaluation fixes the candidate version and configuration before the run.
Private evaluation delivery adds authenticated scorecards, decision-level evidence, operations data, and a recomputation bundle for designated recipients under the agreed access and retention terms.