Epistemic PE / Active Inference routing：区分可学习误差与噪声，再驱动 β_t / curiosity

Blind spot (现状盲点)

DM-1 处理 PE distribution，OA-9 处理 PE faithfulness，但还缺少一个关键区分：哪些 PE 是可学习的 epistemic error，哪些只是 aleatoric noise。如果不分清，系统会把不可学噪声当作 curiosity / credit / needs 的来源，导致 noisy-TV 式探索或关系焦虑误报。

Adoptable suggestions (可落地动作)

1. 1.在 prediction-error spec 中把 Curiosity-Critic 的 "cumulative PE improvement" 作为 PE readout 候选：只有可降低的 PE 才进入 curiosity / credit 主路径。PROPOSED

   Not a runnable A/B candidate — evaluated by the path above, not ablation.

2. 2.在 temporal-abstraction spec 中加入 ODAR / active inference 路由：β_t 切换同时考虑 epistemic gain 与 pragmatic cost。PROPOSED

   Not a runnable A/B candidate — evaluated by the path above, not ablation.

3. 3.第一阶段只做 SHADOW 对照：epistemic-filtered PE vs raw PE 在 held-out dialogue / social cognition probes 上的校准差异。PROPOSED

   Not a runnable A/B candidate — evaluated by the path above, not ablation.

Traceability

No plugins / runs linked yet. Scaffold a suggestion to start.

Expected benefit (预期收益)

- 避免不可学噪声驱动 controller / memory / vitals。 - 给 curiosity 与 intrinsic motivation 一个 R-PE 兼容定义。 - 让 regime / β_t 切换更接近 free-energy / active inference 的统一形式。

Cited paper (引用论文)

**Curiosity-Critic**（2026）、**ODAR**（2026）、**WorldLLM**（2025）、**Predictive Coding ≈ Backprop**。详见 [`research/arxiv-survey-2026-05.md`](../../research/arxiv-survey-2026-05.md) §2 与 [`research/probe/11_vz_implications.md`](../../research/probe/11_vz_implications.md) R-PE。 ---