Add EgoSpeak dense event proxy evaluation

manuscript/main/main.tex

@@ -328,32 +328,47 @@ configuration, trainable parameter counts, and inference latency.}
 We first retain dense future-activity evaluation for direct comparison with
 EgoSpeak, reporting AP at the 0.2--2.0~s horizons and their mean mAP.
 
-Behavioral evaluation uses fixed-horizon event classification. For both
-wearer-entry and INT-yield, we align the dense forecast to each adjudicated
-item at each of the ten nominal horizons. A positive remains a single corrected
-5-Hz behavior-anchor bin: its entry score is wearer activity probability and
-its INT-yield score is one minus wearer activity probability. We do not expand
-the positive timestamp into a tolerance window. A hard-negative opportunity is
-one sample whose score is the maximum erroneous-action probability over its
-labeled half-open region at the same fixed horizon. Thus, each positive event
-or hard-negative opportunity contributes exactly one score per horizon.
+For dense-only models, behavioral evaluation begins with an explicitly
+activity-derived event proxy rather than treating wearer voice activity as a
+semantic event detector. Following the prediction-region principle used by
+VAP~\cite{ekstedt2022vap}, every adjudicated item contributes one score from a
+fixed 0.6-s (three-frame) causal decision region. A positive uses the three
+source frames immediately preceding its corrected behavior anchor. An
+independently mined hard negative uses the first three source frames after its
+opportunity begins; a raw wearer boundary relabeled as a hard negative instead
+uses the same pre-boundary region as a positive. Items without a complete
+three-frame region are excluded and counted. The longer interval supplied to
+the annotator is evidence context only and is never a model-scoring window.
+
+At each decision frame, we follow the target-aligned forecast tensor
+diagonally to recover the wearer probabilities emitted at that same causal
+source time. We average the 0.6--2.0-s horizons to obtain a sustained wearer-
+activity score. This mean is the entry proxy, and one minus the mean is the
+INT-yield proxy. Max-pooling the three decision-frame scores gives exactly one
+score per positive event or hard-negative opportunity. This construction can
+capture sustained activity versus a brief vocalization, but it cannot by
+itself establish semantic floor intent; in particular, wearer backchannels and
+within-turn pauses remain deliberately difficult negatives. A later linear
+event probe, if used, is fit only on training-side annotations and compared at
+matched capacity across representations, following the separated downstream-
+probe logic of MuVAP~\cite{qi2026muvap}.
 
 \textbf{False-positive scoring rule.}
-A false positive is a hard-negative opportunity whose region-level score
+A false positive is a hard-negative opportunity whose fixed-region proxy score
 exceeds the frozen threshold, regardless of whether the wearer subsequently
-exhibits an onset or offset. Region max-pooling means that multiple high-scoring
-frames still count as one false-positive opportunity. Thus, false-positive
+exhibits an onset or offset. Max-pooling means that multiple high-scoring
+decision frames still count as one false-positive opportunity. Thus, false-positive
 evaluation includes conversational situations in which the correct wearer
 behavior is to remain silent or retain the floor.
 
-For every horizon we report event AP and the recall--false-positive trade-off.
-The headline behavioral operating point uses the first, nominal 0.2-s forecast
-bin at matched recall $R=0.8$; thresholds are selected independently for entry
-and INT from training-side out-of-fold predictions. We additionally report the
-complete 0.2--2.0-s curves and the other-to-other hard-negative subset to test
+We report threshold-free proxy AP for entry and INT. The headline behavioral
+operating point additionally reports false-positive rate at matched recall
+$R=0.8$; thresholds are selected independently for entry and INT from training-
+side out-of-fold predictions. The other-to-other hard-negative subset tests
 whether local visual evidence specifically reduces erroneous wearer action
-scores during multiparty floor transfers. \todo{Freeze the hard-negative
-region endpoints and verify identical event denominators across methods.}
+scores during multiparty floor transfers. \todo{Freeze the 0.6-s decision
+window and 0.6--2.0-s aggregation range using training-side data, run window
+sensitivity analysis, and verify identical event denominators across methods.}
 
 A deployment-style committed-event analysis with one-to-one matching and
 signed latency is secondary and is omitted unless a threshold, persistence,
@@ -378,8 +393,8 @@ verify identical targets, causal validity, and evaluation.}
 
 \begin{table}[t]
 \caption{Confirmatory EasyCom results. Entry and INT false-positive rates use
-the nominal 0.2-s forecast bin at matched recall $R=0.8$; E/I AP averages each
-event family over all ten fixed horizons.}
+the fixed-region activity-derived proxy at matched recall $R=0.8$; E/I AP is
+the corresponding event-balanced proxy AP.}
 \label{tab:main}
 \centering
 \small
@@ -411,7 +426,7 @@ evidence primarily improves \todo{the interpretation supported by the data}.
 \textbf{Annotated behavior.}
 At matched wearer-entry recall $R=0.8$, the proposed adapter reduces entry FPR
 from \todo{$X$} to \todo{$Y$}; at matched INT-yield recall, it changes INT FPR
-from \todo{$A$} to \todo{$B$}. Its ten-horizon mean entry/INT event APs are
+from \todo{$A$} to \todo{$B$}. Its entry/INT activity-proxy APs are
 \todo{$E/I$}. These recall-matched opportunity error rates test whether lower
 scores on confusing floor situations are obtained without suppressing
 legitimate entry or yield events.