expose pmass_format as aggregate signal

ForcedChoiceResult now carries pmass_format (sum prob mass on the K
foundation answer tokens at the JSON answer slot, averaged across fwd
and rev framings). eval.py aggregates it as mean_pmass_format in both
the headline return dict and the info subdict, and propagates per-row
for sweep/audit consumers.

Direct coherence canary: drops when steering pushes the model toward
non-foundation tokens (gibberish, refusal, format collapse). Independent
of which foundation is picked — complementary to top1_acc (label-
agreement; intentional target shift) and mean_nll_prompt (teacher-forced
prompt nll; falls under steering even when generations break).

Surfacing this lets downstream callers (weight-steering-lite walkback,
report dashboards) gate on actual coherence rather than misusing top1
as a budget.

src/tinymfv/eval.py

@@ -188,6 +188,7 @@ def evaluate(
                         "top1": res.top1,
                         "margin": res.margin,
                         "nll_prompt": res.nll_prompt,
+                        "pmass_format": res.pmass_format,
                     })
                 pbar.update(len(chunk))
 
@@ -271,6 +272,10 @@ def evaluate(
         T = None
         profile = None
 
+    mean_pmass_format = (
+        float(np.mean([r["pmass_format"] for r in per_row]))
+        if per_row else None
+    )
     info = {
         "name": name,
         "n_rows": n_rows,
@@ -286,6 +291,13 @@ def evaluate(
         # the model's "natural" surprise on prompt text.
         "mean_nll_prompt": float(np.mean([r["nll_prompt"] for r in per_row]))
             if per_row else None,
+        # Mean pmass_format: average prob mass on the K foundation answer
+        # tokens at the JSON answer slot, across rows × framings. In [0, 1].
+        # Direct coherence canary for forced-choice — drops when the model
+        # emits non-foundation tokens (gibberish, refusal, format collapse),
+        # independent of which foundation is picked. Higher = more
+        # "in-format"; a sharp drop after steering signals coherence loss.
+        "mean_pmass_format": mean_pmass_format,
     }
 
     out: dict[str, Any] = {
@@ -297,6 +309,7 @@ def evaluate(
         "median_nll_T": median_nll_T,
         "T": T,                # fitted temperature (>1 = model is overconfident)
         "top1_acc": top1_acc,
+        "mean_pmass_format": mean_pmass_format,
         "info": info,
     }
     if return_per_row:

src/tinymfv/guided.py

@@ -331,6 +331,13 @@ class ForcedChoiceResult:
     # signal: rises when the model is perturbed (steering, ablation, etc.)
     # to a state where ordinary prompt text becomes "surprising".
     nll_prompt: float
+    # Sum of probability mass over the K foundation answer-tokens at the
+    # JSON answer slot, averaged across fwd + rev framings. In [0, 1]; high
+    # means the model still emits a valid foundation word in the slot;
+    # low means probability has leaked to other tokens (gibberish, refusal,
+    # format collapse). The direct coherence canary for forced-choice
+    # — independent of WHICH foundation is picked.
+    pmass_format: float
 
 
 def _resolve_first_token_ids(tok, words: list[str]) -> tuple[list[int], dict[str, int]]:
@@ -444,6 +451,13 @@ def guided_rollout_forced_choice(
         # Average prompt NLL across the two framings (schema_hint differs in
         # enum order but the user vignette is identical).
         nll_p = 0.5 * (nll_fwd[i] + nll_rev[i])
+        # Average pmass_format across framings: coherence canary independent
+        # of WHICH foundation is picked. Sum prob mass over the K answer
+        # tokens at the JSON slot; drops when model emits non-foundation
+        # tokens (gibberish, refusal, format collapse).
+        pm_f = slots_fwd[i][0]["pmass_format"]
+        pm_r = slots_rev[i][0]["pmass_format"]
+        pm = 0.5 * (pm_f + pm_r)
         results.append(ForcedChoiceResult(
             user_prompt=user_prompts[i],
             think_text=think_fwd,
@@ -457,6 +471,7 @@ def guided_rollout_forced_choice(
             think_tokens=n_fwd,
             emitted_close=close_fwd,
             nll_prompt=float(nll_p),
+            pmass_format=float(pm),
         ))
 
     return results