Merge branch 'main' of http://182.92.86.220:8300/wx44wx/icassp2027-experiment

PROJECT_CONTEXT.md

@@ -54,19 +54,32 @@ result.
 - An opportunity-conditioned dense-output evaluator aligns ENTRY positives and
   core negatives at non-wearer releases, aligns INT positives and hold-floor
   negatives at competing non-wearer onsets, and applies strict competing-risk
-  censoring. It uses only the original wearer probabilities, excludes wearer
-  backchannels from headline ENTRY FPR, and records unanchored/ambiguous cases.
+  censoring. ENTRY positives whose selected release is followed by a competing
+  other onset or intervening wearer activity are excluded, and ambiguous
+  multi-competitor INT cases are not assigned an arbitrary onset. It uses only
+  the original wearer probabilities, excludes wearer backchannels from
+  headline ENTRY FPR, and records unanchored/ambiguous cases.
+
+### Implemented but not GPU-verified
+
+- MiniCPM-o 4.5 input-only feature extraction has two unmeasured modes: a
+  native 1 Hz state causally held on the 5 Hz grid, and a 5 Hz completed-cache
+  plus temporary-partial-tail branch. The latter refuses the normal
+  head-loading contract unless boundary, replay, snapshot-restoration, and
+  future-audio checks pass. A real-GPU smoke run is still required; see
+  [`docs/minicpmo_streaming_features.md`](docs/minicpmo_streaming_features.md).
 
 ### Measured screening result
 
 - The released EgoSpeak MiniROADA A+V checkpoint on Sessions 1--3 obtains
   `74.124` 0.2-second mAP and `61.155` ten-horizon mean mAP. Raw wearer
   onset/offset anticipation is harder (`24.491/38.024` mean AP). On provisional
-  LLM-assisted matched opportunities, ENTRY AP is `52.487` versus `32.731%`
-  prevalence, while INT/YIELD AP is `20.135`, indistinguishable from its
-  `20.438%` prevalence. Figure-7-style continuous wearer curves therefore
-  support dense forecasting, not semantic floor-control claims. Full curves,
-  counts, exclusions, and caveats are frozen in
+  LLM-assisted matched opportunities, the original v1 evaluator reported
+  ENTRY/INT AP of `52.487/20.135`; those opportunity values are superseded
+  because v1 could cross a competing ENTRY onset and arbitrarily selected among
+  multiple INT competitors. Figure-7-style continuous wearer curves support
+  dense forecasting, not semantic floor-control claims. The unaffected dense
+  curves and historical v1 opportunity record are in
   [`results/easycom_egospeak_three_level_eval_2026-09-16/`](results/easycom_egospeak_three_level_eval_2026-09-16/README.md).
 - F1 Frame-set Attention uses fused G+A as one query over the latest half-open
   causal person set. On the existing three-seed, four-epoch test screen it
@@ -74,12 +87,11 @@ result.
   `64.95099 +/- 0.09451` ten-horizon mean mAP. This is stable, but is not a
   matched improvement over the historical no-dev pooled result. Its three-seed
   boundary response/yield AP is `28.423 +/- 0.480` / `48.641 +/- 0.391`; on
-  provisional reviewed opportunities its ENTRY/INT AP is
-  `41.561 +/- 0.581` / `22.991 +/- 0.202`. These values are respectively
-  `-4.461/-1.275` points below the historical pooled-V seed-0 result, so the
-  current F1 screen does not support the hoped-for semantic-selectivity gain.
-  A same-split three-seed pooled control and correct/wrong visual controls
-  remain missing. See the immutable
+  the superseded v1 opportunity evaluator reported ENTRY/INT AP of
+  `41.561 +/- 0.581` / `22.991 +/- 0.202`; these values require a v2 rerun and
+  cannot currently support a semantic-selectivity conclusion. A same-split
+  three-seed pooled control and correct/wrong visual controls remain missing.
+  See the historical
   [`screening record`](results/easycom_omni_floor_screen_2026-09-16/README.md).
 
 ### Measured local-fusion screens

README.md

@@ -121,6 +121,7 @@ change what the Stage-0 result measures.
 | `easycom_manifest.py` | Verified on EasyCom | Discover one-minute clips, preserve session timecodes, and check 5 Hz targets |
 | `qwen_omni_features.py` | Implemented, not executed here | Load only the 3B Thinker and extract fused states from one AV window |
 | `streaming_turn_model.py` | Verified on EasyCom | Produce one causal Omni state every 200 ms, with causal-prefix and open-stream modes |
+| `streaming_minicpmo_features_directly.py` | Implemented; real-GPU smoke pending | Extract input-only MiniCPM-o 4.5 native 1 Hz held states or audited 5 Hz cache-plus-partial-tail states |
 | `TurnPredictionHead` | Trained and evaluated | Map `[2048]` to `[10, 3]` logits |
 | `train_turn_head.py` | Verified on EasyCom | Train the frozen-Omni anticipation head with the official EasyCom split |
 | `infer_turn_head.py` | Verified on EasyCom | Save EgoSpeak-format predictions and reproduce per-horizon AP/mAP |

docs/easycom_opportunity_event_evaluation.md

@@ -10,6 +10,12 @@ head.
 
 A positive must have a reviewer-selected preceding non-wearer release. Its
 opportunity starts at that release and resolves at the reviewed wearer entry.
+If a new non-wearer onset occurs before or exactly at that entry, the release
+has already been resolved by a competing event and the positive is excluded.
+The same applies when wearer activity already occurred between the selected
+release and the reviewed onset: the selected release is then a stale,
+post-outcome anchor. Both cases are counted separately in
+`review_diagnostics.skipped`.
 Core hard negatives start at the same type of non-wearer release and resolve
 when the same speaker resumes (`ENTRY_OTHER_MIDTURN_PAUSE`) or a different
 non-wearer takes the next floor (`ENTRY_OTHER_TO_OTHER`). The score is the
@@ -27,10 +33,10 @@ the maximum `1 - p_wearer` over 0.2--0.6 seconds, censored strictly at that
 non-wearer's offset. Reviewed `INT_YIELD` items are positive; other
 backchannels, failed takeovers, and clear wearer-hold cases are hard negatives.
 
-Current INT reviews do not select one competing onset. The loader therefore
-uses the latest supplied other interval that crosses the wearer offset within
-the annotation tolerance. It records how many intervals were eligible so
-multi-competitor cases can be audited before the protocol is frozen.
+Current INT reviews do not select one competing onset. A unique supplied other
+interval that crosses the wearer offset within the annotation tolerance is
+used. Cases with multiple eligible intervals are excluded and counted rather
+than being assigned a deterministic latest-onset anchor.
 
 ## Provisional EgoSpeak run
 
@@ -58,21 +64,24 @@ Before manuscript reporting, freeze the 2.0-second ENTRY cap, 0.6-second YIELD
 cap, positive anchor-selection policy, and multiple-competitor handling using
 training-side data only.
 
-## Recorded Sessions 1--3 baseline
+## Superseded v1 Sessions 1--3 baseline
 
-The first threshold-free run gives ENTRY AP 52.487% (163 positive, 335 hard
-negative; 32.731% prevalence) and INT/YIELD AP 20.135% (28 positive, 109 hard
-negative; 20.438% prevalence). It excludes 81 ENTRY and 5 INT cases as invalid
-or post-outcome. These are provisional LLM-assisted labels, not human-reviewed
-confirmatory results. Dense, raw-boundary, and opportunity metrics and their
-Figure 7 interpretation are frozen together in the
+The first threshold-free run used the superseded v1 loader and reported ENTRY
+AP 52.487% (163 positive, 335 hard negative; 32.731% prevalence) and INT/YIELD
+AP 20.135% (28 positive, 109 hard negative; 20.438% prevalence). It excluded
+81 ENTRY and 5 INT cases as invalid or post-outcome. These opportunity values
+must not be reported: v1 allowed a
+selected ENTRY release to cross an intervening other onset and selected the
+latest onset in ambiguous multi-competitor INT cases. Dense and raw-boundary
+metrics are unaffected. Rerun the opportunity evaluation with protocol v2
+before updating the manuscript. The historical output is retained in the
 [`2026-09-16 result record`](../results/easycom_egospeak_three_level_eval_2026-09-16/README.md).
 
-The expanded Omni screen finds no opportunity-ranking gain from the current
+The expanded v1 Omni screen found no opportunity-ranking gain from the current
 F1 Frame-set Attention model: its three-seed ENTRY/INT AP is
 41.561%/22.991%, below the historical pooled-V seed-0 values
 46.022%/24.266%. Because F1 and that pooled run use different train/dev model-
-selection protocols, this is a stable negative screen rather than a matched
-confirmatory comparison. Dense, boundary, per-seed, and provenance details are
+selection protocols and used the superseded opportunity loader, only its dense
+and raw-boundary findings remain current. Per-seed and provenance details are
 in the
 [`Omni floor screening record`](../results/easycom_omni_floor_screen_2026-09-16/README.md).

docs/minicpmo_streaming_features.md

@@ -0,0 +1,204 @@
+# MiniCPM-o 4.5 causal global features
+
+`streaming_minicpmo_features_directly.py` uses MiniCPM-o 4.5 only as a frozen
+audio--visual feature extractor. It never calls `streaming_generate`, never
+loads TTS, and never lets model-generated speech/listen tokens enter the
+conversation history.
+
+## Frozen modes
+
+### `native-1hz-hold`
+
+- The model commits one input-only AV chunk per completed second.
+- The newest completed state is forward-held onto the 5 Hz target grid.
+- The first four rows (0.2--0.8 s) use an all-zero initialization. The 1.0-s
+  state is never backfilled into those rows.
+- `state_source_times`, `native_update_mask`, and `initialization_mask` make
+  this alignment auditable.
+
+### `cached-partial-5hz`
+
+- Completed one-second chunks remain in the official MiniCPM streaming cache.
+- At a non-boundary 200-ms cutoff, the extractor snapshots the LLM cache,
+  audio-encoder cache, streaming Mel buffer, chunk index, and RNG state through
+  MiniCPM's official speculative-state API.
+- It appends only the current half-open AV tail, reads the final normalized
+  Qwen3 hidden state, and restores the snapshot. Temporary closure never enters
+  the main cache.
+- Native commits retain MiniCPM's normal streaming-window policy. Destructive
+  round pruning is suspended only inside the temporary transaction because the
+  upstream snapshot can truncate appended KV tokens but cannot recreate a
+  previously pruned round.
+- At a one-second boundary, a complete native chunk is committed.
+
+The 5 Hz branch is saved as the normal `mode=causal-prefix` contract only after
+all mandatory checks pass:
+
+1. a temporary full chunk matches the native committed state at a one-second
+   boundary;
+2. a clean stream replay matches the cached result;
+3. changing audio strictly after the cutoff leaves the current state exactly
+   unchanged;
+4. snapshot restoration leaves LLM/audio cache lengths and checksums, chunk
+   history, position offset, and the audio chunk index unchanged.
+
+With `--allow-unverified-partial-tail`, a failed or too-short audit may be
+saved only as `mode=minicpmo-unverified-partial-tail`; the default head loader
+rejects that artifact.
+
+## Environment
+
+The inspected upstream revision requires Python 3.10 and
+`transformers==4.51.0`. Use a dedicated environment rather than changing the
+working Qwen environment.
+
+```bash
+python -m pip install \
+  "transformers==4.51.0" \
+  accelerate \
+  "torch>=2.3.0,<=2.8.0" \
+  "torchaudio<=2.8.0" \
+  "minicpmo-utils>=1.0.5" \
+  "qwen-omni-utils[decord]"
+```
+
+The commands below use the upstream revision inspected during implementation,
+`1f761131fa83f5ed3cd6f2f22b225c4501d154fa`. If the model is already stored
+locally, replace `--model-id` with that snapshot directory and record its
+immutable revision separately.
+
+## Required smoke sequence
+
+First validate the source timeline and decoder without loading MiniCPM:
+
+```bash
+python streaming_minicpmo_features_directly.py \
+  --easycom-session-dir /data1/yfzx/wangx/Datasets/easycom/extracted_easycom/Main/Video_Compressed/Session_4 \
+  --end 3 \
+  --mode native-1hz-hold \
+  --decoder decord \
+  --media-only
+```
+
+Then extract three seconds with the native held-state baseline:
+
+```bash
+CUDA_VISIBLE_DEVICES=0 python streaming_minicpmo_features_directly.py \
+  --model-id openbmb/MiniCPM-o-4_5 \
+  --revision 1f761131fa83f5ed3cd6f2f22b225c4501d154fa \
+  --easycom-session-dir /data1/yfzx/wangx/Datasets/easycom/extracted_easycom/Main/Video_Compressed/Session_4 \
+  --end 3 \
+  --mode native-1hz-hold \
+  --dtype bfloat16 \
+  --attention sdpa \
+  --device cuda:0 \
+  --decoder decord \
+  --decoder-threads 1 \
+  --max-open-readers 1 \
+  --output /data1/yfzx/wangx/Datasets/easycom/minicpmo_features/smoke/Session_4_native_3s.pt
+```
+
+Run the verified cached partial-tail smoke test separately:
+
+```bash
+CUDA_VISIBLE_DEVICES=0 python streaming_minicpmo_features_directly.py \
+  --model-id openbmb/MiniCPM-o-4_5 \
+  --revision 1f761131fa83f5ed3cd6f2f22b225c4501d154fa \
+  --easycom-session-dir /data1/yfzx/wangx/Datasets/easycom/extracted_easycom/Main/Video_Compressed/Session_4 \
+  --end 3 \
+  --mode cached-partial-5hz \
+  --verify-every 1 \
+  --future-intervention-every 1 \
+  --dtype bfloat16 \
+  --attention sdpa \
+  --device cuda:0 \
+  --decoder decord \
+  --decoder-threads 1 \
+  --max-open-readers 1 \
+  --output /data1/yfzx/wangx/Datasets/easycom/minicpmo_features/smoke/Session_4_partial_3s.pt
+```
+
+Do not add `--allow-unverified-partial-tail` to a paper run. A verification
+failure is a protocol failure to diagnose, not a tolerance to bypass.
+
+Inspect both artifacts before a full session:
+
+```bash
+python - <<'PY'
+from pathlib import Path
+import torch
+
+root = Path("/data1/yfzx/wangx/Datasets/easycom/minicpmo_features/smoke")
+for name in ("Session_4_native_3s.pt", "Session_4_partial_3s.pt"):
+    payload = torch.load(root / name, map_location="cpu", weights_only=False)
+    print(name)
+    print("  mode:", payload["mode"], payload["extractor_mode"])
+    print("  states:", tuple(payload["states"].shape))
+    print("  times:", payload["times"][:6].tolist())
+    print("  source times:", payload["state_source_times"][:6].tolist())
+    print("  verified:", payload["partial_tail_admissible"])
+    print("  checks:", payload["verification"])
+PY
+```
+
+Expected alignment:
+
+```text
+native times:        0.2 0.4 0.6 0.8 1.0 1.2 ...
+native source time:  0.0 0.0 0.0 0.0 1.0 1.0 ...
+partial source time: 0.2 0.4 0.6 0.8 1.0 1.2 ...
+```
+
+## Full Session 4
+
+Native 1 Hz held features:
+
+```bash
+mkdir -p /data1/yfzx/wangx/Datasets/easycom/minicpmo_features/native_1hz_hold
+
+CUDA_VISIBLE_DEVICES=0 python streaming_minicpmo_features_directly.py \
+  --model-id openbmb/MiniCPM-o-4_5 \
+  --revision 1f761131fa83f5ed3cd6f2f22b225c4501d154fa \
+  --easycom-session-dir /data1/yfzx/wangx/Datasets/easycom/extracted_easycom/Main/Video_Compressed/Session_4 \
+  --target /data1/yfzx/wangx/Datasets/easycom/easycom_target_frame/Session_4.npy \
+  --mode native-1hz-hold \
+  --dtype bfloat16 \
+  --attention sdpa \
+  --device cuda:0 \
+  --decoder decord \
+  --decoder-threads 1 \
+  --max-open-readers 1 \
+  --frame-cache-size 32 \
+  --output /data1/yfzx/wangx/Datasets/easycom/minicpmo_features/native_1hz_hold/Session_4.pt
+```
+
+Verified 5 Hz cached partial-tail features:
+
+```bash
+mkdir -p /data1/yfzx/wangx/Datasets/easycom/minicpmo_features/cached_partial_5hz
+
+CUDA_VISIBLE_DEVICES=0 python streaming_minicpmo_features_directly.py \
+  --model-id openbmb/MiniCPM-o-4_5 \
+  --revision 1f761131fa83f5ed3cd6f2f22b225c4501d154fa \
+  --easycom-session-dir /data1/yfzx/wangx/Datasets/easycom/extracted_easycom/Main/Video_Compressed/Session_4 \
+  --target /data1/yfzx/wangx/Datasets/easycom/easycom_target_frame/Session_4.npy \
+  --mode cached-partial-5hz \
+  --dtype bfloat16 \
+  --attention sdpa \
+  --device cuda:0 \
+  --decoder decord \
+  --decoder-threads 1 \
+  --max-open-readers 1 \
+  --frame-cache-size 32 \
+  --output /data1/yfzx/wangx/Datasets/easycom/minicpmo_features/cached_partial_5hz/Session_4.pt
+```
+
+The full run still executes the first clean-replay and future-intervention
+checks plus every one-second boundary check. Add `--verify-every 250` and
+`--future-intervention-every 250` for periodic 50-second audits; smaller values
+make long sessions increasingly expensive because a clean audit replays the
+prefix from the session start.
+
+After extraction, either directory can be passed directly as
+`train_turn_head.py --feature-root ...` because verified artifacts retain the
+existing dense global-feature contract.

docs/omni_speaker_floor_experiment_plan_2026-08-30.md

@@ -168,12 +168,13 @@ forecasting 的定性解释，不能代替 boundary 或 semantic opportunity 评
 曲线和限制见
 [`three-level baseline record`](../results/easycom_egospeak_three_level_eval_2026-09-16/README.md)。
 
-当前 F1 current-frame person-set screen 的三种子结果同样没有支持 semantic floor
-selectivity：ENTRY/INT opportunity AP 为 `41.561 +/- 0.581` / `22.991 +/- 0.202`，
-低于历史 pooled-V seed-0 的 `46.022/24.266`；dense mean mAP 也为
-`64.951 +/- 0.095`，没有超过历史 pooled-V 的 `65.400`。由于两者 train/dev
-选择协议不同，这只能判为稳定的负向 screening，正式归因前必须补同协议三种子 pooled
-control。详见 [`Omni floor screening record`](../results/easycom_omni_floor_screen_2026-09-16/README.md)。
+当前 F1 current-frame person-set screen 的 dense mean mAP 为
+`64.951 +/- 0.095`，没有超过历史 pooled-V 的 `65.400`。原 v1 evaluator 给出的
+ENTRY/INT opportunity AP（`41.561 +/- 0.581` / `22.991 +/- 0.202`）已失效：v1
+可能跨过 competing ENTRY onset，且会在多个 INT competitor 中任意选择 latest onset。
+必须用 v2 重跑后才能判断 semantic floor selectivity；正式归因还必须补同协议三种子
+pooled control。历史数值与勘误见
+[`Omni floor screening record`](../results/easycom_omni_floor_screen_2026-09-16/README.md)。
 
 ### P3：第一优先方法——匿名逐人时序旁路
 

evaluate_egospeak_opportunity_events.py

@@ -8,14 +8,18 @@ event head.
 
 ENTRY opportunities are anchored at a reviewed preceding non-wearer release.
 Positive opportunities resolve at the wearer entry; core hard negatives
-resolve when the same or a different non-wearer speaks next.  The score is the
-maximum wearer probability over available 0.2--2.0 s forecasts up to that
-competing resolution.
+resolve when the same or a different non-wearer speaks next.  A positive is
+excluded if another non-wearer onset competes before (or exactly at) its wearer
+entry, or if wearer activity already occurred after the selected release.  The
+score is the maximum wearer probability over available 0.2--2.0 s forecasts up
+to the uncontested resolution.
 
 YIELD opportunities are anchored at a non-wearer onset during wearer speech.
 The score is the maximum future wearer-inactivity probability over 0.2--0.6 s,
-censored at that non-wearer's offset.  Positive and hard-negative examples
-therefore begin at the same kind of observed conversational opportunity.
+censored at that non-wearer's offset.  Reviews with multiple eligible
+competing onsets are excluded until the annotation selects one.  Positive and
+hard-negative examples therefore begin at the same kind of unambiguous
+observed conversational opportunity.
 
 Prediction arrays use the repository contract
 ``[target_time, class, future_horizon]``.  All scores are read diagonally from
@@ -178,17 +182,83 @@ def _candidate_region(
     return start, end
 
 
-def _int_competing_interval(
+def _entry_selected_option(
+    candidate: dict[str, Any],
+    review: dict[str, Any],
+    *,
+    selected_release_seconds: float,
+    uid: str,
+) -> dict[str, Any] | None:
+    evidence = candidate.get("temporal_evidence")
+    evidence = evidence if isinstance(evidence, dict) else {}
+    options = evidence.get("opportunity_options")
+    if not isinstance(options, list):
+        return None
+    selected_uid = review.get("selected_opportunity_uid")
+    matches: list[dict[str, Any]] = []
+    for index, option in enumerate(options):
+        if not isinstance(option, dict):
+            continue
+        option_seconds = _optional_seconds(
+            option.get("anchor_seconds"),
+            field=f"opportunity_options[{index}].anchor_seconds",
+            uid=uid,
+        )
+        if selected_uid is not None and option.get("uid") == selected_uid:
+            matches.append(option)
+        elif selected_uid is None and option_seconds is not None and math.isclose(
+            option_seconds, selected_release_seconds, abs_tol=1e-7
+        ):
+            matches.append(option)
+    if len(matches) > 1:
+        raise ValueError(f"{uid}: selected ENTRY opportunity is not unique")
+    return matches[0] if matches else None
+
+
+def _entry_competing_onsets(
+    candidate: dict[str, Any],
+    *,
+    selected_release_seconds: float,
+    behavior_anchor_seconds: float,
+    uid: str,
+) -> list[float]:
+    """Return other onsets that resolve a release before the wearer entry."""
+
+    evidence = candidate.get("temporal_evidence")
+    evidence = evidence if isinstance(evidence, dict) else {}
+    raw_intervals = evidence.get("gap_other_intervals")
+    if not isinstance(raw_intervals, list):
+        raw_intervals = []
+    onsets: set[float] = set()
+    for index, interval in enumerate(raw_intervals):
+        if not isinstance(interval, dict):
+            continue
+        onset = _optional_seconds(
+            interval.get("onset_seconds"),
+            field=f"gap_other_intervals[{index}].onset_seconds",
+            uid=uid,
+        )
+        if onset is None:
+            continue
+        if (
+            onset >= selected_release_seconds - 1e-7
+            and onset <= behavior_anchor_seconds + 1e-7
+        ):
+            onsets.add(onset)
+    return sorted(onsets)
+
+
+def _int_competing_intervals(
     candidate: dict[str, Any],
     *,
     behavior_anchor_seconds: float,
     uid: str,
-) -> tuple[float, float, int] | None:
+) -> list[tuple[float, float]]:
     evidence = candidate.get("temporal_evidence")
     evidence = evidence if isinstance(evidence, dict) else {}
     raw_intervals = evidence.get("nearby_other_intervals")
     if not isinstance(raw_intervals, list):
-        return None
+        return []
     eligible: list[tuple[float, float]] = []
     for index, interval in enumerate(raw_intervals):
         if not isinstance(interval, dict):
@@ -209,13 +279,7 @@ def _int_competing_interval(
             continue
         if onset < behavior_anchor_seconds and offset > onset:
             eligible.append((onset, offset))
-    if not eligible:
-        return None
-    # A semantic INT review does not currently select one competing onset.
-    # The latest interval crossing the wearer offset is the deterministic,
-    # most proximal causal candidate.  Multiplicity is surfaced for audit.
-    onset, offset = max(eligible, key=lambda item: (item[0], item[1]))
-    return onset, offset, len(eligible)
+    return sorted(set(eligible))
 
 
 def load_opportunity_events(
@@ -278,6 +342,38 @@ def load_opportunity_events(
                 if selected_release is None:
                     skipped["entry_positive_without_selected_release"] += 1
                     continue
+                selected_option = _entry_selected_option(
+                    candidate,
+                    review,
+                    selected_release_seconds=selected_release,
+                    uid=uid,
+                )
+                if (
+                    selected_option is not None
+                    and selected_option.get(
+                        "wearer_active_between_release_and_onset"
+                    ) is True
+                ):
+                    skipped["entry_positive_post_outcome_wearer_activity"] += 1
+                    continue
+                competing_onsets = _entry_competing_onsets(
+                    candidate,
+                    selected_release_seconds=selected_release,
+                    behavior_anchor_seconds=behavior,
+                    uid=uid,
+                )
+                option_reports_competitor = (
+                    selected_option is not None
+                    and selected_option.get(
+                        "other_onset_between_release_and_onset"
+                    ) is True
+                )
+                if option_reports_competitor and not competing_onsets:
+                    skipped["entry_positive_competing_onset_unresolved"] += 1
+                    continue
+                if competing_onsets:
+                    skipped["entry_positive_competed_before_behavior"] += 1
+                    continue
                 anchor = selected_release
                 resolution = behavior
                 censor = behavior
@@ -298,17 +394,21 @@ def load_opportunity_events(
             competing_intervals = 0
         else:
             if positive:
-                interval = _int_competing_interval(
+                intervals = _int_competing_intervals(
                     candidate,
                     behavior_anchor_seconds=behavior,
                     uid=uid,
                 )
-                if interval is None:
+                if not intervals:
                     skipped["int_positive_without_crossing_other"] += 1
                     continue
-                anchor, censor, competing_intervals = interval
+                if len(intervals) > 1:
+                    skipped["int_positive_ambiguous_competing_others"] += 1
+                    continue
+                anchor, censor = intervals[0]
+                competing_intervals = 1
                 resolution = behavior
-                basis = "latest_other_onset_crossing_wearer_offset"
+                basis = "unique_other_onset_crossing_wearer_offset"
             elif subtype_value in INT_HARD_NEGATIVES:
                 if source in INT_MINED_SOURCES:
                     anchor, censor = _candidate_region(candidate, uid=uid)
@@ -316,17 +416,21 @@ def load_opportunity_events(
                     competing_intervals = 1
                     basis = "mined_competing_other_onset"
                 else:
-                    interval = _int_competing_interval(
+                    intervals = _int_competing_intervals(
                         candidate,
                         behavior_anchor_seconds=behavior,
                         uid=uid,
                     )
-                    if interval is None:
+                    if not intervals:
                         skipped["int_hn_without_competing_other"] += 1
                         continue
-                    anchor, censor, competing_intervals = interval
+                    if len(intervals) > 1:
+                        skipped["int_hn_ambiguous_competing_others"] += 1
+                        continue
+                    anchor, censor = intervals[0]
+                    competing_intervals = 1
                     resolution = behavior
-                    basis = "latest_other_onset_crossing_non_yield_offset"
+                    basis = "unique_other_onset_crossing_non_yield_offset"
             else:
                 skipped["int_unsupported_hard_negative"] += 1
                 continue
@@ -731,8 +835,8 @@ def main() -> None:
         else calibration_diagnostics["num_provisional_llm"]
     )
     payload = {
-        "schema_version": 1,
-        "protocol": "egospeak-opportunity-conditioned-floor-actions-v1",
+        "schema_version": 2,
+        "protocol": "egospeak-opportunity-conditioned-floor-actions-v2",
         "semantic_scope": (
             "LLM-assisted opportunity labels with dense wearer-activity scores; "
             "this is provisional and no human review is implied."
@@ -757,6 +861,13 @@ def main() -> None:
         "yield_max_seconds": yield_max,
         "entry_headline_hard_negatives": sorted(ENTRY_CORE_HARD_NEGATIVES),
         "entry_backchannel_policy": "separate diagnostic; excluded from headline",
+        "entry_competing_risk_policy": (
+            "exclude selected-release positives with an other onset at or before "
+            "the wearer entry, or with intervening wearer activity"
+        ),
+        "int_multiple_competitor_policy": (
+            "exclude until one competing onset is selected by review"
+        ),
         "matched_recalls": recalls,
         "review_diagnostics": review_diagnostics,
         "calibration_review_diagnostics": calibration_diagnostics,

manuscript/main/main.tex

@@ -339,9 +339,12 @@ non-wearer release. It resolves positively when the wearer enters, or
 negatively when the same non-wearer resumes or a different non-wearer takes the
 next floor. From the first causal 5-Hz state that observes the release, the
 entry score is the maximum projected wearer probability before that competing
-resolution, capped at 2.0~s. Positive entries without a reviewer-selected
-release are reported as unanchored and excluded from this matched-opportunity
-analysis; they remain available to the separate raw-onset analysis.
+resolution, capped at 2.0~s. A selected-release positive is excluded if another
+non-wearer onset occurs before or exactly at the wearer entry, or if wearer
+activity already occurred after the selected release. Positive entries without
+a reviewer-selected release are reported as unanchored and excluded from this
+matched-opportunity analysis; they remain available to the separate raw-onset
+analysis.
 
 An INT opportunity begins when a non-wearer starts while the wearer holds the
 floor. It resolves positively when the reviewed behavior is an INT-yield, and
@@ -352,8 +355,9 @@ strictly at the competing non-wearer offset. All prediction values are read
 diagonally from one causal source row of the target-aligned forecast tensor.
 No input after the observed action outcome is used. For current INT-positive
 annotations, which do not select a competing onset explicitly, the
-implementation uses the latest supplied non-wearer interval crossing the
-wearer offset and surfaces cases with multiple eligible intervals for audit.
+implementation evaluates only a unique supplied non-wearer interval crossing
+the wearer offset; cases with multiple eligible intervals are excluded and
+counted for audit rather than assigned an arbitrary onset.
 
 The headline entry hard negatives are same-speaker mid-turn pauses and other-
 to-other transfers. Wearer backchannels are not pooled into headline false-
@@ -394,46 +398,60 @@ latency claim.
 
 \subsection{Main Comparison}
 
-Table~\ref{tab:main} compares the proposed causal local audiovisual adapter
-with an EgoSpeak reproduction, the frozen Qwen2.5-Omni-3B global state $G$,
-matched PEFT, $G$ plus recent local audio, and a parameter- and history-matched
-temporal local branch. Qwen2.5-Omni-7B is included only as a capacity-scaling
-control. All methods use the same split, causal inputs, targets, and evaluation
-protocol. Event thresholds are selected on training-side calibration data and
-frozen before test evaluation. \todo{Freeze all baseline implementations and
-verify identical targets, causal validity, and evaluation.}
+Table~\ref{tab:main} places the released EgoSpeak checkpoint beside the
+currently available Omni screening runs and reserves a row for the final causal
+local audiovisual model. Historical unmatched runs are identified explicitly
+and cannot support confirmatory attribution. The final comparison will use the
+same split, causal inputs, targets, training/selection protocol, and three
+random seeds; event thresholds will be selected on training-side calibration
+data and frozen before test evaluation. \todo{Freeze all baseline
+implementations and verify identical targets, causal validity, and evaluation.}
 
-\begin{table}[t]
-\caption{Confirmatory EasyCom results. Entry and INT false-positive rates use
-the opportunity-conditioned activity score at matched recall $R=0.8$; E/I AP
-is the corresponding opportunity AP. $\dagger$ denotes provisional
-LLM-assisted labels pending human review.}
+\begin{table*}[t]
+\caption{EasyCom Sessions~1--3 dense future-activity and raw-boundary
+screening. All values are percentages. ``Avg.'' averages dense mAP over the ten
+horizons; On/Off are mean raw wearer onset/offset AP and are timing diagnostics,
+not semantic ENTRY/INT scores. $\dagger$ denotes historical seed-0 runs with an
+unmatched training/selection protocol; $\ddagger$ denotes a three-seed screen.
+Missing horizon curves and the final matched comparison will be filled after
+the matched dense/boundary reruns.}
 \label{tab:main}
 \centering
-\small
-\resizebox{\columnwidth}{!}{%
-\begin{tabular}{lccccc}
+\scriptsize
+\setlength{\tabcolsep}{2.6pt}
+\resizebox{\textwidth}{!}{%
+\begin{tabular}{@{}lccccccccccccc@{}}
 \hline
-Method & 0.2-s mAP $\uparrow$ & Mean mAP $\uparrow$
-& Entry FPR@$R=.8$ $\downarrow$ & INT FPR@$R=.8$ $\downarrow$ & E/I AP $\uparrow$\\
+& \multicolumn{10}{c}{Dense mAP at prediction horizon (s) $\uparrow$}
+& & \multicolumn{2}{c}{Raw boundary AP $\uparrow$}\\
+\cline{2-11}\cline{13-14}
+Method & .2 & .4 & .6 & .8 & 1.0 & 1.2 & 1.4 & 1.6 & 1.8 & 2.0
+& Avg. & On & Off\\
 \hline
-EgoSpeak MiniROADA A+V      & 74.12 & 61.16 & \todo{--} & \todo{--} & $52.49/20.13^\dagger$\\
-Omni-3B frozen $G$          & \todo{--} & \todo{--} & \todo{--} & \todo{--} & \todo{--/--}\\
-Omni-3B matched PEFT        & \todo{--} & \todo{--} & \todo{--} & \todo{--} & \todo{--/--}\\
-$G+$local audio             & \todo{--} & \todo{--} & \todo{--} & \todo{--} & \todo{--/--}\\
-$G+$matched temporal local  & \todo{--} & \todo{--} & \todo{--} & \todo{--} & \todo{--/--}\\
-$G+$causal local AV (ours)  & \todo{--} & \todo{--} & \todo{--} & \todo{--} & \todo{--/--}\\
-Omni-7B scaling control     & \todo{--} & \todo{--} & \todo{--} & \todo{--} & \todo{--/--}\\
+EgoSpeak MiniROADA A+V
+  & 74.20 & 68.35 & 64.31 & 61.13 & 59.77 & 58.81 & 57.50 & 57.09 & 55.95 & 54.82
+  & 61.19 & 24.49 & 38.02\\
+Omni $G{+}A^{\dagger}$
+  & 81.72 & -- & -- & -- & -- & -- & -- & -- & -- & --
+  & 65.11 & 25.96 & 45.25\\
+Omni $G{+}A{+}$pooled-$V^{\dagger}$
+  & 82.03 & -- & -- & -- & -- & -- & -- & -- & -- & --
+  & 65.40 & 26.02 & 45.37\\
+Current-frame F1$^{\ddagger}$
+  & $81.78{\pm}0.16$ & -- & -- & -- & -- & -- & -- & -- & -- & --
+  & $64.95{\pm}0.09$ & -- & --\\
+$G+$causal local AV (ours)
+  & -- & -- & -- & -- & -- & -- & -- & -- & -- & --
+  & -- & -- & --\\
 \hline
 \end{tabular}}
-\end{table}
+\end{table*}
 
 \textbf{Dense forecasting.}
-The EgoSpeak baseline reaches 74.12\% 0.2-s mAP and 61.16\% mean mAP; its raw
-wearer-onset/offset mean APs are 24.49/38.02\%. On provisional reviewed
-opportunities, ENTRY AP is 52.49\% versus 32.73\% prevalence, whereas INT AP is
-20.13\% versus 20.44\%; continuous wearer-score plots therefore do not alone
-establish semantic floor-yield selectivity.
+The EgoSpeak baseline reaches 74.20\% 0.2-s mAP and 61.19\% mean mAP; its raw
+wearer-onset/offset mean APs are 24.49/38.02\%. Opportunity-conditioned values
+remain pending a protocol-v2 rerun and human review; continuous wearer-score
+plots do not alone establish semantic floor-yield selectivity.
 The proposed local AV adapter improves 0.2-s mAP from
 \todo{$X$ to $Y$} over the strongest matched baseline and improves
 mean mAP by \todo{$Z$ points}.
@@ -460,6 +478,34 @@ beyond additional temporal capacity}.
 \todo{If this difference is not statistically supported, state instead that
 the local temporal branch accounts for most of the gain.}
 
+\subsection{Qualitative Floor-Event Analysis}
+
+We will compare EgoSpeak with the final matched model on human-confirmed
+examples selected by pre-declared event subtype rather than post-hoc score
+extremes. Each case will align participant-resolved activity, minimal transcript
+context, the opportunity anchor and competing-risk censor, and the two models'
+causally emitted wearer forecasts. The comparison will include a genuine wearer
+entry, an other-to-other false-entry opportunity, and an INT-yield paired with a
+non-takeover overlap; a shared failure case will be retained when space permits.
+
+\begin{figure*}[t]
+\centering
+\fbox{\parbox[c][0.115\textheight][c]{0.965\textwidth}{
+\centering
+\textbf{Qualitative comparison placeholder}\\[0.7em]
+\begin{tabular}{ccc}
+Case A: wearer entry & Case B: other$\rightarrow$other HN
+& Case C: INT-yield / failed takeover\\
+participant timeline + transcript & opportunity anchor + censor
+& EgoSpeak vs. proposed forecast\\
+\end{tabular}
+}}
+\caption{Planned qualitative floor-event comparison. Final panels will use
+human-confirmed protocol-v2 cases and show the same causal source time for both
+models; the figure is a layout placeholder and contains no result claim.}
+\label{fig:qualitative_floor_events}
+\end{figure*}
+
 \subsection{Information Opportunity}
 
 Before designing the deployable adapter, we measured whether local floor
@@ -570,7 +616,7 @@ future-media intervention test. Report within-backbone 5-Hz versus 1-Hz
 deltas plus cache/no-cache throughput and memory; do not interpret unmatched
 Qwen--MiniCPM raw scores as a pure backbone effect.}
 
-\subsection{Limitations and Qualitative Analysis}
+\subsection{Limitations}
 
 The results support \todo{precise supported claim}, but several limitations
 remain.
@@ -580,10 +626,9 @@ tracking assumption}.
 Third, \todo{failure condition: occlusion, overlapping speakers, rapid floor
 exchange, etc.} remains challenging.
 
-\todo{Add one compact, pre-selected qualitative timeline only if space permits:
-one multi-other false-entry improvement paired with one preserved real wearer
-onset or one failure. Examples illustrate aggregate rates and never establish
-the claim by themselves.}
+\todo{Replace Fig.~\ref{fig:qualitative_floor_events} with compact, pre-selected
+human-confirmed timelines. Examples illustrate aggregate rates and never
+establish the claim by themselves.}
 
 \section{Conclusion}
 \label{sec:conclusion}

results/easycom_egospeak_three_level_eval_2026-09-16/README.md

@@ -1,5 +1,11 @@
 # EasyCom EgoSpeak three-level evaluation
 
+> **Protocol erratum.** The dense and raw-boundary results in this record remain
+> valid. The opportunity-conditioned ENTRY/INT values were produced by the
+> superseded v1 loader, which could cross an intervening ENTRY competitor and
+> selected the latest onset in ambiguous multi-competitor INT cases. Do not cite
+> those opportunity values; rerun protocol v2 and create a new result record.
+
 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
@@ -15,7 +21,7 @@ prediction and per-event artifacts remain on that host and are not committed.
 - 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
+- Opportunity-conditioned behavior: historical v1 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
@@ -34,15 +40,16 @@ 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
+The superseded v1 run reported ENTRY AP of **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.
+and 109 hard negatives, essentially equal to its 20.438% prevalence. These
+values are retained only for provenance and must be replaced by a v2 rerun. The
+original pattern was 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
 
@@ -73,4 +80,3 @@ 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).
-

results/easycom_egospeak_three_level_eval_2026-09-16/summary.json

@@ -1,6 +1,8 @@
 {
   "recorded_at": "2026-09-16",
   "source": "user-provided remote experiment output",
+  "opportunity_protocol_status": "superseded_v1_rerun_required",
+  "opportunity_protocol_erratum": "Dense and raw-boundary metrics remain valid; opportunity metrics must be rerun with egospeak-opportunity-conditioned-floor-actions-v2.",
   "dataset": "EasyCom",
   "sessions": ["Session_1", "Session_2", "Session_3"],
   "model": "EgoSpeak MiniROADA audio-visual released checkpoint",
@@ -57,4 +59,3 @@
   "threshold_free": true,
   "session4_required_for_reported_metrics": false
 }
-

results/easycom_omni_floor_screen_2026-09-16/README.md

@@ -1,5 +1,12 @@
 # EasyCom Omni floor-event screening
 
+> **Protocol erratum.** Dense and raw-boundary metrics remain valid. All
+> opportunity-conditioned ENTRY/INT values below used the superseded v1 loader,
+> which could cross an intervening ENTRY competitor and selected the latest
+> onset in ambiguous multi-competitor INT cases. Retain them for provenance
+> only; do not use them for model selection or manuscript claims before a v2
+> rerun.
+
 This record freezes the first comparison of EgoSpeak, Omni `G+A`, Omni
 `G+A+pooled-V`, and the three-seed current-frame F1 Frame-set Attention screen
 on EasyCom Sessions 1--3. Values were produced on the remote experiment host
@@ -11,9 +18,8 @@ CSVs remain on that host and are not committed here.
 1. Dense future activity uses causal-valid framewise AP/mAP over ten horizons.
 2. Raw wearer endpoint and aggregate-other-boundary AP are threshold-free
    timing diagnostics; they are not semantic EOT/INT labels.
-3. Opportunity-conditioned ENTRY and INT/YIELD AP assigns one causally
-   censored activity-derived score to each provisional LLM-assisted reviewed
-   opportunity.
+3. Opportunity-conditioned ENTRY and INT/YIELD AP is a historical v1 screen
+   that requires rerunning with the corrected v2 competing-risk protocol.
 
 All opportunity runs use identical denominators: 163 ENTRY positives and 335
 hard negatives, with 81 invalid/post-outcome cases; and 28 INT positives and
@@ -39,7 +45,7 @@ from EgoSpeak to G+A to G+A+pooled-V as follows: onset AP
 offset/yield-timing gain comes from G+A; pooled visual input adds almost
 nothing to raw transition timing.
 
-## Interpretation
+## Historical v1 interpretation
 
 - G+A strongly improves dense forecasting and generic wearer-offset/yield
   timing over EgoSpeak, but not reviewed semantic opportunities. In particular,
@@ -77,4 +83,3 @@ nothing to raw transition timing.
 Exact aggregate values are stored in [`summary.json`](summary.json) and
 [`metrics.csv`](metrics.csv). The earlier EgoSpeak-only record is
 [`easycom_egospeak_three_level_eval_2026-09-16`](../easycom_egospeak_three_level_eval_2026-09-16/README.md).
-

results/easycom_omni_floor_screen_2026-09-16/summary.json

@@ -1,6 +1,8 @@
 {
   "recorded_at": "2026-09-16",
   "source": "user-provided remote experiment output",
+  "opportunity_protocol_status": "superseded_v1_rerun_required",
+  "opportunity_protocol_erratum": "Dense and raw-boundary metrics remain valid; opportunity metrics must be rerun with egospeak-opportunity-conditioned-floor-actions-v2.",
   "dataset": "EasyCom",
   "test_sessions": ["Session_1", "Session_2", "Session_3"],
   "scale": "percent",
@@ -82,4 +84,3 @@
   "threshold_free_metrics_only": true,
   "matched_pooled_control_available": false
 }
-

test_evaluate_egospeak_opportunity_events.py

@@ -30,7 +30,10 @@ def reviewed_row(
     region_start: float,
     region_end: float,
     selected_release: float | None = None,
+    selected_release_uid: str | None = None,
     nearby: list[dict[str, object]] | None = None,
+    opportunity_options: list[dict[str, object]] | None = None,
+    gap_other_intervals: list[dict[str, object]] | None = None,
     candidate_source: str | None = None,
 ) -> dict[str, object]:
     event_type = None
@@ -45,6 +48,7 @@ def reviewed_row(
             "human_verdict": verdict,
             "human_event_type": event_type,
             "human_subtype": subtype,
+            "selected_opportunity_uid": selected_release_uid,
             "selected_opportunity_seconds": selected_release,
             "boundary_correction_seconds": 0,
         },
@@ -59,6 +63,8 @@ def reviewed_row(
                 "behavior_anchor_seconds": anchor,
                 "wearer_offset_seconds": anchor,
                 "nearby_other_intervals": nearby or [],
+                "opportunity_options": opportunity_options or [],
+                "gap_other_intervals": gap_other_intervals or [],
             },
         },
     }
@@ -119,7 +125,7 @@ class OpportunityEventTests(unittest.TestCase):
             diagnostics["skipped"]["entry_backchannel_separate_diagnostic"], 1
         )
 
-    def test_loader_selects_latest_crossing_other_for_int(self) -> None:
+    def test_loader_excludes_ambiguous_crossing_others_for_int(self) -> None:
         nearby = [
             {
                 "onset_seconds": 1.0,
@@ -145,11 +151,91 @@ class OpportunityEventTests(unittest.TestCase):
             )
         ])
         self.addCleanup(temporary.cleanup)
-        events, _ = load_opportunity_events([path], ["Session_4"])
-        self.assertEqual(len(events), 1)
-        self.assertEqual(events[0].opportunity_anchor_seconds, 2.0)
-        self.assertEqual(events[0].censor_seconds, 3.8)
-        self.assertEqual(events[0].competing_intervals, 2)
+        events, diagnostics = load_opportunity_events([path], ["Session_4"])
+        self.assertEqual(events, [])
+        self.assertEqual(
+            diagnostics["skipped"]["int_positive_ambiguous_competing_others"],
+            1,
+        )
+
+    def test_loader_excludes_entry_positive_at_competing_other_onset(self) -> None:
+        rows = [
+            reviewed_row(
+                "entry-competed",
+                family="entry",
+                verdict="POSITIVE",
+                subtype=None,
+                anchor=1.6,
+                region_start=0.0,
+                region_end=3.0,
+                selected_release=1.0,
+                selected_release_uid="release",
+                opportunity_options=[
+                    {
+                        "uid": "release",
+                        "anchor_seconds": 1.0,
+                        "wearer_active_between_release_and_onset": False,
+                        "other_onset_between_release_and_onset": True,
+                    }
+                ],
+                gap_other_intervals=[
+                    {"uid": "competitor", "onset_seconds": 1.4}
+                ],
+            ),
+            reviewed_row(
+                "entry-simultaneous",
+                family="entry",
+                verdict="POSITIVE",
+                subtype=None,
+                anchor=4.0,
+                region_start=3.0,
+                region_end=5.0,
+                selected_release=3.2,
+                gap_other_intervals=[
+                    {"uid": "competitor", "onset_seconds": 4.0}
+                ],
+            ),
+        ]
+        temporary, path = self.write_reviews(rows)
+        self.addCleanup(temporary.cleanup)
+        events, diagnostics = load_opportunity_events([path], ["Session_4"])
+        self.assertEqual(events, [])
+        self.assertEqual(
+            diagnostics["skipped"]["entry_positive_competed_before_behavior"],
+            2,
+        )
+
+    def test_loader_excludes_entry_positive_after_prior_wearer_activity(self) -> None:
+        temporary, path = self.write_reviews([
+            reviewed_row(
+                "entry-post-outcome",
+                family="entry",
+                verdict="POSITIVE",
+                subtype=None,
+                anchor=2.0,
+                region_start=0.0,
+                region_end=3.0,
+                selected_release=1.0,
+                selected_release_uid="release",
+                opportunity_options=[
+                    {
+                        "uid": "release",
+                        "anchor_seconds": 1.0,
+                        "wearer_active_between_release_and_onset": True,
+                        "other_onset_between_release_and_onset": False,
+                    }
+                ],
+            )
+        ])
+        self.addCleanup(temporary.cleanup)
+        events, diagnostics = load_opportunity_events([path], ["Session_4"])
+        self.assertEqual(events, [])
+        self.assertEqual(
+            diagnostics["skipped"][
+                "entry_positive_post_outcome_wearer_activity"
+            ],
+            1,
+        )
 
     def test_entry_hn_requires_matching_independently_mined_source(self) -> None:
         rows = [
@@ -385,7 +471,7 @@ class OpportunityEventTests(unittest.TestCase):
         payload = json.loads(output.read_text(encoding="utf-8"))
         self.assertEqual(
             payload["protocol"],
-            "egospeak-opportunity-conditioned-floor-actions-v1",
+            "egospeak-opportunity-conditioned-floor-actions-v2",
         )
         self.assertTrue((root / "summary_metrics.csv").is_file())
         self.assertTrue((root / "summary_events.csv").is_file())

test_streaming_minicpmo_features_directly.py

@@ -0,0 +1,230 @@
+"""CPU-only contract tests for MiniCPM-o streaming feature schedules."""
+
+from __future__ import annotations
+
+import tempfile
+import unittest
+from pathlib import Path
+
+import numpy as np
+import torch
+
+from easycom_head import load_session_data
+from local_turn_features import reference_alignment
+from streaming_minicpmo_features_directly import (
+    FEATURE_FORMAT,
+    MiniCPMFeatureExtractor,
+    MiniCPMStreamConfig,
+    _comparison,
+    _time_grid,
+    validate_transformers_version,
+)
+
+
+class FakeReader:
+    def __init__(self, duration: float = 3.0) -> None:
+        self.audio = np.arange(
+            int(duration * 16_000), dtype=np.float32
+        ) / 16_000.0
+
+    def frame_at_or_before(self, timestamp: float) -> float:
+        return float(timestamp)
+
+    def audio_slice(self, start: float, end: float) -> np.ndarray:
+        first = int(round(start * 16_000))
+        last = int(round(end * 16_000))
+        return self.audio[first:last].copy()
+
+
+class FakeEngine:
+    hidden_size = 3
+
+    def __init__(self) -> None:
+        self.committed_until = 0.0
+        self.reset_count = 0
+        self.last_restore_audit = None
+
+    def reset(self) -> None:
+        self.committed_until = 0.0
+        self.reset_count += 1
+
+    @staticmethod
+    def _state(frame: float, audio: np.ndarray) -> torch.Tensor:
+        return torch.tensor(
+            [[float(audio[-1]), float(frame), float(len(audio))]],
+            dtype=torch.float32,
+        )
+
+    def commit_state(self, *, frame: float, audio: np.ndarray) -> torch.Tensor:
+        self.committed_until += len(audio) / 16_000.0
+        return self._state(frame, audio)
+
+    def temporary_state(self, *, frame: float, audio: np.ndarray) -> torch.Tensor:
+        before = self.committed_until
+        state = self._state(frame, audio)
+        if self.committed_until != before:
+            raise AssertionError("temporary state modified the fake main cache")
+        self.last_restore_audit = {
+            "kind": "snapshot_restore",
+            "llm_cache_length": 0,
+            "audio_cache_length": 0,
+            "audio_chunk_index": 0,
+            "streaming_position_offset": 0,
+            "window_pruning_suspended": True,
+            "window_policy_restored": True,
+            "llm_checksum_matches": True,
+            "audio_checksum_matches": True,
+            "history_matches": True,
+            "passed": True,
+        }
+        return state
+
+
+class MiniCPMStreamingTests(unittest.TestCase):
+    def setUp(self) -> None:
+        self.config = MiniCPMStreamConfig(
+            boundary_max_abs=0.0,
+            boundary_min_cosine=1.0,
+            replay_max_abs=0.0,
+            replay_min_cosine=1.0,
+        )
+
+    def test_time_grid_is_right_edged_and_rejects_fractional_tail(self) -> None:
+        self.assertEqual(_time_grid(0.0, 1.0, 0.2), [0.2, 0.4, 0.6, 0.8, 1.0])
+        with self.assertRaisesRegex(ValueError, "integer number"):
+            _time_grid(0.0, 1.1, 0.2)
+
+    def test_native_hold_has_no_future_backfill(self) -> None:
+        extractor = MiniCPMFeatureExtractor(FakeReader(), FakeEngine(), self.config)
+        result = extractor.extract_native_hold(
+            start=0.0,
+            end=2.0,
+            show_progress=False,
+        )
+        self.assertEqual(tuple(result.states.shape), (10, 3))
+        self.assertTrue(torch.equal(result.states[:4], torch.zeros(4, 3)))
+        self.assertTrue(result.initialization_mask[:4].all())
+        self.assertFalse(result.initialization_mask[4:].any())
+        self.assertEqual(
+            result.native_update_mask.nonzero().flatten().tolist(),
+            [4, 9],
+        )
+        self.assertTrue(torch.equal(result.states[4], result.states[8]))
+        self.assertFalse(torch.equal(result.states[8], result.states[9]))
+        self.assertEqual(
+            result.state_source_times.tolist(),
+            [0.0, 0.0, 0.0, 0.0, 1.0, 1.0, 1.0, 1.0, 1.0, 2.0],
+        )
+
+    def test_cached_partial_updates_every_row_and_passes_audits(self) -> None:
+        reader = FakeReader()
+        extractor = MiniCPMFeatureExtractor(reader, FakeEngine(), self.config)
+        original_audio = reader.audio.copy()
+        result = extractor.extract_cached_partial(
+            start=0.0,
+            end=2.0,
+            verify_every=0,
+            future_intervention_every=0,
+            allow_unverified=False,
+            show_progress=False,
+        )
+        self.assertTrue(result.partial_tail_admissible)
+        self.assertTrue(np.array_equal(reader.audio, original_audio))
+        self.assertFalse(result.initialization_mask.any())
+        self.assertTrue(torch.equal(result.state_source_times, result.times))
+        self.assertEqual(
+            result.native_update_mask.nonzero().flatten().tolist(),
+            [4, 9],
+        )
+        self.assertEqual(
+            {item["kind"] for item in result.verification},
+            {
+                "native_boundary_equivalence",
+                "clean_stream_replay",
+                "future_audio_intervention",
+                "snapshot_restore",
+            },
+        )
+        self.assertTrue(all(item["passed"] for item in result.verification))
+        self.assertFalse(torch.equal(result.states[0], result.states[1]))
+        self.assertLess(float(result.states[0, 1]), 0.2)
+
+    def test_short_partial_run_is_rejected_without_boundary_audit(self) -> None:
+        extractor = MiniCPMFeatureExtractor(FakeReader(), FakeEngine(), self.config)
+        with self.assertRaisesRegex(RuntimeError, "missing=.*native_boundary"):
+            extractor.extract_cached_partial(
+                start=0.0,
+                end=0.8,
+                verify_every=0,
+                future_intervention_every=0,
+                allow_unverified=False,
+                show_progress=False,
+            )
+
+    def test_comparison_rejects_different_states(self) -> None:
+        item = _comparison(
+            torch.zeros(1, 2),
+            torch.ones(1, 2),
+            kind="test",
+            current_time=0.2,
+            max_abs_tolerance=0.0,
+            min_cosine=1.0,
+        )
+        self.assertFalse(item["passed"])
+        self.assertEqual(item["max_abs"], 1.0)
+
+    def test_transformers_version_is_explicitly_gated(self) -> None:
+        self.assertEqual(validate_transformers_version("4.51.0"), "4.51.0")
+        with self.assertRaisesRegex(RuntimeError, "found 4.52.0"):
+            validate_transformers_version("4.52.0")
+        self.assertEqual(
+            validate_transformers_version("4.52.0", allow_untested=True),
+            "4.52.0",
+        )
+
+    def test_saved_payload_loads_in_existing_head_and_unverified_fails(self) -> None:
+        states = torch.randn(20, 8)
+        payload = {
+            "format": FEATURE_FORMAT,
+            "states": states,
+            "times": (torch.arange(20, dtype=torch.float64) + 1) / 5,
+            "label_indices": torch.arange(20),
+            "mode": "causal-prefix",
+            "extractor_mode": "native-1hz-hold",
+            "config": {"step_seconds": 0.2},
+        }
+        targets = np.tile(
+            np.asarray(
+                [[1, 0, 0], [0, 1, 0], [0, 0, 1], [0, 1, 1]],
+                dtype=np.float32,
+            ),
+            (5, 1),
+        )
+        with tempfile.TemporaryDirectory() as directory:
+            root = Path(directory)
+            feature_root = root / "features"
+            target_root = root / "targets"
+            feature_root.mkdir()
+            target_root.mkdir()
+            feature_path = feature_root / "Session_4.pt"
+            torch.save(payload, feature_path)
+            np.save(target_root / "Session_4.npy", targets)
+            loaded = load_session_data(feature_root, target_root, 4)
+            self.assertTrue(torch.equal(loaded.states, states))
+            times, indices = reference_alignment(payload)
+            self.assertTrue(torch.equal(times, payload["times"]))
+            self.assertTrue(torch.equal(indices, payload["label_indices"]))
+
+            bad_reference = dict(payload)
+            bad_reference["times"] = payload["times"] + 0.2
+            with self.assertRaisesRegex(ValueError, "Reference times"):
+                reference_alignment(bad_reference)
+
+            payload["mode"] = "minicpmo-unverified-partial-tail"
+            torch.save(payload, feature_path)
+            with self.assertRaisesRegex(ValueError, "expected 'causal-prefix'"):
+                load_session_data(feature_root, target_root, 4)
+
+
+if __name__ == "__main__":
+    unittest.main()