Mixed sim + animal training + deviation analysis

[hanhou]

Put real animals and synthetic agents in one embedding space.

Scope

• Quantify where/how animal behavior deviates from its best-fit Q-learning agent (residual structure = what classical RL misses)
• XAI framing: synthetic anchors calibrate the meaning of embedding axes

[hanhou]

🔴 Not started — but half the machinery already exists

The synthetic side of this issue is built and validated. 04-gru-vs-disrnn-embedding-recovery puts Q-learning-family agents (QL variants, CompareToThreshold, LossCounting) into a learned embedding space with ground-truth params recovered at 97.5–100% — exactly the "synthetic anchors calibrate the meaning of embedding axes" XAI framing in this issue's scope.

What's missing is the mixing: training real mice and synthetic agents in one embedding space, then quantifying where animal behavior deviates from its best-fit Q-learning agent (the residual structure = what classical RL misses).

Blocked on #20: "deviation from its best-fit Q-learning agent" requires per-mouse RL fits. A simple per-mouse Bari fit exists (r8) — where the GRU already beats it by +0.0136 on 100% of held-out mice, which is the deviation signal in aggregate — but the interpretable per-subject posterior params need the hierarchical Bayes fit.

Worth flagging: this may be cheaper than its P2 label suggests, since the synthetic generator, the recovery analysis, and a per-mouse RL baseline are all already committed.