PN — Calibration metrics (ECE over 10 bins, MCE, Brier) are derived from…
Calibration metrics (ECE over 10 bins, MCE, Brier) are derived from per-token logprobs via the Ollama logprobs API (requires logprobs:true and Ollama >= 0.12.11); class confidence uses a numerically stable softmax over first-token candidates.
Source: puma docs/metrics_reference.md · Traceability: corpus unit
PUMA-029· Confidence: repo-backed