Record EgoSpeak three-level evaluation

results/easycom_egospeak_three_level_eval_2026-09-16/README.md

@@ -0,0 +1,76 @@
+# EasyCom EgoSpeak three-level evaluation
+
+This record freezes the first Sessions 1--3 evaluation of the released
+EgoSpeak MiniROADA audio-visual checkpoint under three deliberately separate
+questions. The values below were produced on the remote experiment host; the
+aggregate console output was copied into this record on 2026-09-16. Generated
+prediction and per-event artifacts remain on that host and are not committed.
+
+## Protocol and split
+
+- Test split: EasyCom Sessions 1--3.
+- Model: released EgoSpeak MiniROADA audio-visual checkpoint.
+- Dense forecasting: framewise AP for background, wearer, and aggregate
+  others at ten horizons from 0.2 to 2.0 seconds.
+- Raw boundary anticipation: threshold-free AP for wearer `OFF->ON` and
+  `ON->OFF` changes. These are timing-derived diagnostics, not semantic EOT or
+  INT claims.
+- Opportunity-conditioned behavior: one causally censored activity-derived
+  score per reviewed ENTRY or INT/YIELD opportunity. Labels are provisional
+  two-pass LLM-assisted reviews and have not been human reviewed.
+- No test-set threshold is fitted for any AP reported here. Session 4 is needed
+  only for threshold-dependent operating points, which are intentionally not
+  summarized as headline findings in this record.
+
+## Findings
+
+Dense forecasting reaches **74.124% mAP at 0.2 s** and **61.155% mean mAP**
+over ten horizons. The wearer class alone reaches 63.750% AP at 0.2 s and
+53.493% averaged over horizons. Restricting evaluation to causally valid rows
+changes mean mAP by only +0.038 percentage points (61.193% versus 61.155%).
+
+Raw wearer onset and offset anticipation obtain **24.491%** and **38.024%**
+mean AP, respectively. These lower values do not contradict the dense result:
+dense AP rewards correct future state over sustained regions, whereas a raw
+boundary event is counted once and has many harder non-boundary comparisons.
+
+On the provisional opportunity labels, ENTRY AP is **52.487%** for 163
+positives and 335 hard negatives, versus 32.731% positive prevalence
+(+19.756 points; 1.60x prevalence). INT/YIELD AP is **20.135%** for 28 positives
+and 109 hard negatives, essentially equal to its 20.438% prevalence. This is
+consistent with the output contract: wearer probability can rank many entry
+opportunities, but `1-p(wearer)` does not encode whether an interruption should
+cause a yield. The ENTRY and INT loaders additionally exclude 81 and 5
+invalid/post-outcome cases; these exclusions require an audit before any final
+paper claim.
+
+## Relation to EgoSpeak Figure 7
+
+Figure 7 of EgoSpeak plots a continuous target/wearer-speaking confidence over
+ground-truth speaking regions. It therefore corresponds most directly to the
+dense **wearer AP**, not to raw onset/offset AP or semantic ENTRY/YIELD AP. A
+visually convincing dense curve can coexist with weak boundary or semantic
+selectivity. The figure caption does not identify one forecast horizon, and it
+uses the paper's Transformer variants rather than proving the aggregate result
+of this MiniROADA run.
+
+For our eventual qualitative figure, the continuous wearer curve can be kept,
+but it should also show the causal opportunity anchor, event boundary,
+competing-risk censor, and the single score assigned to each positive or hard
+negative.
+
+## Interpretation limits
+
+- The three metric layers answer different questions and their similar numeric
+  values must not be equated.
+- Aggregate `others` activity cannot identify a particular speaker, addressee,
+  semantic EOT, or interruption intent.
+- The opportunity result is provisional until label/exclusion audit and human
+  review; it is useful for baseline diagnosis, not a final confirmatory claim.
+- This is one released checkpoint, not a multi-seed uncertainty estimate.
+
+Machine-readable curves and counts are in [`summary.json`](summary.json) and
+the compact comparison is in [`metrics.csv`](metrics.csv). The opportunity
+protocol and reproduction command are documented in
+[`docs/easycom_opportunity_event_evaluation.md`](../../docs/easycom_opportunity_event_evaluation.md).
+