scripts: train_specialist, distillation for every arena plus an RLCR stage rewarding right and honest answers

[BANADDA]

Step 8 of the full system plan, part two. Stacked on #120.

scripts/train_decider.py becomes scripts/train_specialist.py (renamed, not copied), and gains an RLCR stage:

• Distillation as before: calibrated decider training on answer probabilities for choice arenas (KL plus Brier), and fine tuning on frontier teacher completions for every other arena.
• RLCR (--rlcr-steps N): for each task it draws --rlcr-samples answers and scores each with the arena's metric. The reward is the score minus --rlcr-weight times the squared gap between the answer's own confidence (exp of its mean token logprob) and whether it was right. Advantages are normalised within each task's group and the policy is pushed towards the better answers.
• Reward ordering: right and sure 1.00, right and unsure 0.64, wrong and unsure −0.01, wrong and sure −0.81. Confident mistakes are what it trains out.
• Choice arenas skip RLCR, because the Brier rule they already train on rewards honest confidence directly.
• Export still keeps the output layer at Q8, and the temperature fold (fold_temperature.py) puts calibration into the file.

References in docs/system_miner.md and pyproject.toml move to the new name.

🤖 Generated with Claude Code