docs: restore SEPA as entropy pooling, note surprisal is current proxy

SEPA is conceptually about entropy pooling — the name and theory are
entropy-based. Surprisal (-logprob) is used today because backends only
provide logprobs, not full token distributions. Updated docs and manual
to keep "Selective Entropy Pooling of Attention" while noting that the
implementation currently operates on surprisal values.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Author: teilomillet

docs/sepa.md

@@ -4,14 +4,16 @@ Selective Entropy Pooling of Attention (SEPA) is an adaptive scheduling strategy
 
 ## What SEPA does
 
-In the [advantage pipeline](advantages.md), GTPO weights token advantages by token surprisal (`-logprob`) -- high-surprisal tokens get more credit. SEPA modifies execution-token surprisals *before* GTPO weighting by pulling them toward their mean:
+In the [advantage pipeline](advantages.md), GTPO weights token advantages by per-token uncertainty -- high-uncertainty tokens get more credit. SEPA modifies execution-token uncertainty values *before* GTPO weighting by pulling them toward their mean:
 
 ```
 H_pooled(t) = lambda * mean(H_exec) + (1 - lambda) * H(t)    if execution token
 H_pooled(t) = H(t)                                             if planning token
 ```
 
-When `lambda = 0`, SEPA is off and GTPO sees raw surprisals. When `lambda = 1`, all execution tokens have the same surprisal (the mean), so GTPO's differentiation is concentrated entirely on planning tokens.
+When `lambda = 0`, SEPA is off and GTPO sees raw values. When `lambda = 1`, all execution tokens have the same uncertainty (the mean), so GTPO's differentiation is concentrated entirely on planning tokens.
+
+Today `uncertainty_kind = "surprisal"` (sampled-token `-logprob`) is the only available signal. When backends provide full token distributions, `shannon_entropy` becomes usable and SEPA will pool true entropy values.
 
 ## Why
 

retrain/retrain.man

@@ -278,7 +278,7 @@ CONFIGURATION
                         High-surprisal tokens get more credit.
                         Controlled by gtpo_beta.
         gtpo_hicra      GTPO + HICRA planning token amplification.
-        gtpo_sepa       GTPO + SEPA selective surprisal pooling (default).
+        gtpo_sepa       GTPO + SEPA selective entropy pooling (default).
         gtpo_sepa_amp   GTPO + SEPA with amplification variant.
         gtpo_sepa_amp_c GTPO + SEPA amplification with clamping.
         entropy_mask    Yue et al. surprisal masking. Zeros out advantages
@@ -437,7 +437,7 @@ CONFIGURATION
 
     schedule = "linear"
         "linear" or "auto". Auto adapts based on execution-token
-        surprisal variance decay with linear as fallback floor.
+        entropy variance decay with linear as fallback floor.
 
     delay_steps = 50
         Steps to delay before beginning the SEPA ramp. Must be >= 0.
@@ -861,11 +861,13 @@ ADVANTAGE PIPELINE
         or semantic similarity.
 
     SEPA
-        Selective Surprisal Pooling of Attention.
-        Pools execution-token surprisals toward their mean, reducing
-        surprisal variance on execution tokens while preserving diverse
-        planning behavior. Lambda ramps 0->1 over training, gated by
-        a correct_rate threshold.
+        Selective Entropy Pooling of Attention.
+        Pools execution-token uncertainty values toward their mean,
+        reducing variance on execution tokens while preserving diverse
+        planning behavior. Currently operates on surprisal (the only
+        available signal); will use full entropy when backends provide
+        token distributions. Lambda ramps 0->1 over training, gated
+        by a correct_rate threshold.
 
   Post-Process Hooks
 
@@ -1155,9 +1157,10 @@ GLOSSARY
             weighted credit assignment at the token level.
     HICRA   Hierarchical Credit Assignment. Planning token amplification
             via a learned mask.
-    SEPA    Selective Surprisal Pooling of Attention. Pools execution-token
-            surprisals to reduce variance while preserving planning
-            diversity.
+    SEPA    Selective Entropy Pooling of Attention. Pools execution-token
+            uncertainty values to reduce variance while preserving
+            planning diversity. Uses surprisal today; full entropy
+            when backends provide token distributions.
     LoRA    Low-Rank Adaptation. Parameter-efficient fine-tuning via
             rank-decomposed weight updates.
     SVD     Singular Value Decomposition. Used by LoRA-Squeeze for rank