Notes to build mental models for how things work.
How Claude Opus 5.5 and Inkling 2 Small diagnose and edit Qwen3.5-0.8B’s weights to stop it looping on math: direct logit attribution, difference of means, token directions, and every edit tried so far.
Agreement minus luck, divided by the room above luck. Why a judge is gated on kappa but not picked on it, from the Legal Research Bench calibration study.
Why a mismatched inference engine pushes the trainer toward itself, why ratios cannot fix it cleanly, and the one-line change that does: subtract the engine’s distribution instead of the trainer’s.
Every RL-for-LLM weighting rule, from REINFORCE to DPPO, measured on a toy where the true gradient is exact: bias, variance, the cap, and why a trust region is needed.