icassp2027-experiment/PROJECT_CONTEXT.md

# AI Project Context

This is the short, authoritative repository map for humans and AI agents.
The long [`README.md`](README.md) is an operational history and command
reference; it is not the experiment scheduler.

## One-sentence objective

Determine whether strictly causal global Omni context plus local audiovisual
floor evidence organized around visible people can improve an invisible
egocentric wearer's **floor-entry** and **floor-yield** behavior in natural
multiparty conversation.

The paper is not simply “replace EgoSpeak's encoders with a newer backbone.”
The decisive test is whether correct person-aligned vision reduces false
wearer entries during other-to-other conversation at matched wearer-onset
recall, without hurting addressed responses or yield behavior.

## Source-of-truth order

| Need | Authoritative file |
|---|---|
| ICASSP submission story, pruning and acceptance evidence bar | [docs/icassp2027_acceptance_storyline_2026-09-10.md](docs/icassp2027_acceptance_storyline_2026-09-10.md) |
| Current paper thesis and claim boundary | [docs/global_local_omni_paper_position_2026-09-07.md](docs/global_local_omni_paper_position_2026-09-07.md) |
| Progressive feature-fusion implementation route | [docs/progressive_feature_fusion_route_2026-09-02.md](docs/progressive_feature_fusion_route_2026-09-02.md) |
| Current hypotheses, experiment gates and stopping rules | [`docs/omni_speaker_floor_experiment_plan_2026-08-30.md`](docs/omni_speaker_floor_experiment_plan_2026-08-30.md) |
| Novelty boundary and claim restrictions | [`docs/multiparty_egocentric_floor_gap_audit_2026-08-30.md`](docs/multiparty_egocentric_floor_gap_audit_2026-08-30.md) |
| Code/data map and current blocker | This file |
| Runnable commands and implementation history | [`README.md`](README.md) and [`docs/README.md`](docs/README.md) |
| Archived measured results | [`results/README.md`](results/README.md) |
| Literature summaries | [`literature_notes/README.md`](literature_notes/README.md) |

When files disagree, prefer the newest scoped document and verify behavior in
code/tests. A planned experiment is not an implemented feature or a measured
result.

## Current state

### Verified assets

- Strictly causal 5 Hz Qwen2.5-Omni global state extraction on EasyCom.
- Prepared-media and direct-original-media extraction paths.
- Frozen head training/inference with EgoSpeak-compatible 10-horizon AP/mAP.
- EasyCom local Qwen visual crops and recent mixed-audio sidecars.
- Encoder-only A/V/AV and `G`, `G+A`, `G+V`, `G+A+V` baselines.
- Ego4D batch and 8-GPU local-feature launchers, with strict frame audits.
- Timing-derived wearer floor-state and event diagnostics.
- Human-annotated fixed-horizon entry/INT evaluator with point positives,
  region-level hard negatives and training-side threshold calibration.
- A separate EgoSpeak-compatible activity-derived event-proxy evaluator that
  pools the 0.6--2.0-s wearer trajectory over equal-width 0.6-s causal decision
  regions. It preserves one score per annotated opportunity and never treats
  the longer annotation evidence interval as a scoring window.
- 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.

### 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
  [`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
  obtained `81.77745 +/- 0.15638` 0.2-second mAP and
  `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
  [`screening record`](results/easycom_omni_floor_screen_2026-09-16/README.md).

### Measured local-fusion screens

- Two current-sidecar screening controls: `audio-person-frame-set` adds a
  zero-gated audio-to-person read before the F1 query, and
  `gated-person-state-adapter` adds a zero-gated person residual after the
  frozen Thinker readout. They are explicitly recorded as F5/F6 screening,
  not full history-conditioned F5 or an internal Thinker F6 adapter.
- `track-gru-residual` builds strictly causal histories from anonymous provided
  track IDs, applies one shared GRU per track, lets G+A query current track
  states, and adds a zero-gated correction to an exactly initialized pooled
  `G+A+V` checkpoint. The matched Session-12 three-seed screen was null
  (`+0.01306 +/- 0.02368` percentage points in ten-horizon mean mAP). A seed-0
  Sessions-1--3 audit reproduced pooled predictions to within `-0.00030`
  points. Saved residual gates were only `-0.0015` to `+0.0108`; because a zero
  scalar gate also blocks the initial branch gradient, this is recorded as an
  optimization failure rather than evidence that track continuity is useless.
  See the immutable
  [`result record`](results/easycom_track_gru_residual_4ep_2026-09-07/README.md)
  and [`implementation note`](docs/easycom_track_gru_residual.md).
- `causal-person-adapter-v2` is now implemented but unmeasured. It preserves an
  exact matched pooled `G+A+V` base, reuses the same causal anonymous track
  histories, replaces the blocking zero scalar with a sample-dependent vector
  gate, and zero-initializes only the final residual-logit projection. Its
  isolated runbook is
  [`docs/easycom_causal_person_adapter_v2.md`](docs/easycom_causal_person_adapter_v2.md).
  This is an optimization-repair/incremental screen, not evidence that track
  identity matters.
- `causal-person-adapter-egospeak-gru` is an unmeasured controlled ablation of
  V2. It changes only the per-track visual temporal stem to EgoSpeak
  MiniROADA's `Linear-LN-ReLU-Dropout-GRU-ReLU` ordering while retaining V2's
  pooled base, G+A person query, set attention, dynamic gate and zero-output
  residual. It is an EgoSpeak-style per-track adaptation, not a reproduction of
  EgoSpeak's global 5 Hz RGB+audio GRU. See its
  [`runbook`](docs/easycom_egospeak_style_person_gru.md).
- `highrate-person-av-attention` is now implemented but unmeasured. A separate
  extractor runs the published, frozen MuVAP frontend at 25 Hz: mixed audio is
  encoded once as a shared CPC stream, while each face track contributes a
  causal visual state plus one AV active-speaker score. At each 5 Hz right edge
  `G+U` queries that variable person set. The first screen loads and freezes
  the matched trained global-Omni + traditional-audio base and learns a
  zero-initialized person-logit residual. This is global Omni context plus
  traditional high-rate audio/person perception;
  it is not the full MuVAP forecasting model and does not yet include online
  detection. See the
  [`runbook`](docs/easycom_highrate_person_av.md).
- A MuVAP-only Stage-3 baseline is now implemented as a clean-room paper
  reconstruction. It freezes the released 25 Hz mono role-VAP and causal ASD
  backbones, lets `Z_VAP` query the variable `Z_ASD` person set at each frame,
  applies a one-layer causal temporal Transformer, and trains only that fusion
  plus the EgoSpeak 10-horizon head. It deliberately excludes Omni and is not
  described as upstream Stage-3 code because that entry point remains
  unpublished. See the
  [`runbook`](docs/easycom_muvap_stage3_baseline.md).
- `egospeak-omni-zero-residual` is implemented but unmeasured. It strictly
  loads and freezes the released MiniROADA A+V/GRU/anticipation path, reuses
  existing causal 5 Hz Omni readouts, adds an explicit ten-bin partial-tail
  phase embedding, and learns only a zero-output residual over the original
  logits. The launcher verifies exact initial recovery and exports fused and
  matched frozen-EgoSpeak predictions together. Existing Omni states already
  combine completed history with the current partial tail; separating those
  states remains a future extraction ablation. See the
  [`runbook`](docs/easycom_egospeak_omni_residual.md).

### Measured oracle-information screen

- A seed-0, four-epoch Session-12 development screen supplied transcript-derived
  current/past speech activity only through the explicit analysis-only oracle
  contract. `aggregate-visible` improved over the matched local-audio baseline
  by `+6.96649` points at 0.2 seconds and `+2.91047` points in ten-horizon mean
  mAP. Visibility-only and 1-second-delayed controls stayed near audio.
- Person identity continuity is not supported: `aggregate-visible` exceeded
  `per-person`, while `per-person` and `shuffled-id` were effectively tied.
  The result motivates a learned causal local active-speaker/floor-evidence
  branch, but does not justify a track-continuity claim. See the immutable
  [`result record`](results/easycom_oracle_information_4ep_2026-09-07/README.md).

### Current blocker

`evaluate_floor_state_changes.py` now implements and unit-tests the required
competing-risk attribution: an other-offset response window ends before the
next other onset, and an other-onset yield window ends before the next other
offset. It also supports disjoint training-side threshold calibration, rejects
official-test calibration on Sessions 1--3, reports false-entry/false-stop
counts, and prints auditable timelines.

P0 is still not frozen. The remaining blocker is a real-data audit: run the v2
evaluator on saved calibration/test predictions, inspect event timelines and
censoring/invalid counts, independently review sampled events, and freeze the
scenario definitions. These values remain timing-derived behavior proxies,
not semantic EOT/INT or addressee ground truth. See the
[`runbook`](docs/easycom_competing_risk_floor_evaluation.md).

Three explanation-oriented diagnostics are implemented but unmeasured: a
separately identified post-boundary reactive evaluator, a direct
competing-risk response/yield probe over the same frozen Omni/audio/oracle
inputs, and a privileged direct-to-wearer `Target_of_Speech` oracle with an
equal-width zero-channel control. They diagnose timing, objective and
addressee-information limitations respectively; none changes or freezes P0.
See [`docs/easycom_floor_event_diagnostics.md`](docs/easycom_floor_event_diagnostics.md).

### Next experiment gates

The user has temporarily chosen to run the dense-AP model screen while P0 is
unfinished. This does not relax the paper claim: event numbers remain proxy
diagnostics until the v2 attribution, calibration and scenario audit is frozen
on real data.

1. **P0:** run and audit competing-risk v2, then freeze event attribution,
   calibration, splits and scenario definitions.
2. **P1:** confirm the aggregate/per-person oracle screen over seeds 1--2,
   then train a causal visible-person active-speaker/floor-evidence estimator;
   compare correct, wrong-person and time-shifted visual inputs.
3. **P2:** re-evaluate `G`, `G+A`, `G+A+pooled-V` under the frozen event protocol.
4. **P3:** only if learned person-organized evidence beats aggregate/no-ID
   controls, repair the matched anonymous-track-GRU residual optimization by
   replacing the scalar zero gate with a zero-initialized residual output
   layer, and verify nonzero branch gradients on Session 12. Add broken-ID,
   wrong-time and remove-person controls only if that corrected screen improves.
5. **P4:** attempt person-memory adapter/Thinker LoRA only if P1/P3 pass.

The detailed architecture, controls and stopping gates for P3 are frozen in
[the progressive feature-fusion route](docs/progressive_feature_fusion_route_2026-09-02.md).

## Stable data contracts

### Time and labels

- Sampling rate: 5 Hz (`STEP_SECONDS = 0.2`).
- State row `j` uses media no later than `(j+1)/5` seconds.
- Future horizons: 10, corresponding to 0.2, 0.4, ..., 2.0 seconds.
- Class order: `background`, `camera_wearer_speaking`,
  `other_people_speaking`.
- EasyCom targets are dense `[T, 3]`; predictions are
  `[target_time, class, future_horizon]`.
- EasyCom split: train Session 4--12; test Session 1--3.
- Locked final head protocol: exactly 4 epochs on all Sessions 4--12, no
  development checkpoint selection, followed by one evaluation on Sessions
  1--3. Historical Session-12 screens remain development diagnostics and are
  not the final training protocol.

### Global feature sidecar

Produced by `streaming_turn_model.py` or
`streaming_turn_model_directly.py` as a `.pt` dictionary. Required training
fields include:

- `states`: `[T, 2048]` or multi-layer `[T, L, 2048]`;
- `times`: `[0.2, 0.4, ...]`;
- `label_indices`: dense zero-origin indices into the target timeline;
- extraction `mode`, configuration and source metadata.

Never silently accept non-zero origins, missing label indices or mismatched
time grids.

### Local feature sidecar

Produced by `extract_turn_local_features.py`, validated by
`local_turn_features.py`, and stored as `qwen-local-pooled-v1`. It contains
native 2048D visual/audio vectors, validity masks, times, label indices,
source metadata and the exact extraction config. A partial smoke sidecar must
not be reused as a full-session feature file.

### Inference result

`infer_turn_head.py` writes:

```text
RESULT_DIR/
  metrics.json
  pred/<sample>.npy       # [T, 3, 10]
  valid/<sample>.npy      # [T, 10]
  transition_pred/        # only when the auxiliary head exists
```

`evaluate_floor_state_changes.py` consumes `pred/` and `valid/`; it does not
run a GPU model.

## Canonical pipelines

### Direct global features (preferred semantic path)

```text
original EasyCom/Ego4D media
  -> streaming_turn_model_directly.py
  -> global feature .pt
  -> train_turn_head.py
  -> checkpoint .pth + .json
  -> infer_turn_head.py
  -> pred/ + valid/ + metrics.json
  -> evaluate_floor_state_changes.py
```

This path samples source video causally in Python and avoids requiring a
re-encoded fixed-FPS MP4.

### Prepared-media compatibility path

```text
prepare_easycom_session.py / prepare_ego4d_2fps.py
  -> strict 2 FPS media + metadata
  -> streaming_turn_model.py
  -> the same training/inference pipeline
```

Keep it for reproducibility, cache-replay verification and comparisons with
already extracted features.

### Local evidence path

```text
raw media + person boxes + existing global reference .pt
  -> extract_turn_local_features.py
     (or extract_ego4d_local_multigpu.py supervisor)
  -> local .pt sidecar
  -> train_turn_head.py --local-feature-root ...
  -> infer_turn_head.py --local-feature-root ...
```

The current local visual feature is pooled. It does not by itself prove
speaker identity, active-speaker alignment or addressee understanding.

## Root-code map

Root scripts remain flat because remote L20 commands call them directly.
Do not reorganize them physically without compatibility wrappers.

### Shared libraries

| File | Responsibility |
|---|---|
| `easycom_head.py` | Dataset constants, feature loading, heads, losses and AP metrics |
| `local_turn_features.py` | Local-sidecar contract, alignment, crops and pooling |
| `omni_cached_av.py` | Cached AV structures used by Omni extraction/LoRA paths |

### Global feature extraction

| File | Responsibility |
|---|---|
| `streaming_turn_model_directly.py` | Preferred direct raw-media causal extraction |
| `streaming_turn_model.py` | Prepared-media causal-prefix/open-stream extraction and replay checks |
| `qwen_omni_features.py` | Single-window feature inspection |
| `cache_omni_av_pieces.py` | Build reusable AV pieces for LoRA experiments |
| `extract_ego4d_batch.py` | Persistent-model Ego4D global batch extraction |
| `analyze_block_phase.py` | Streaming block-phase diagnostics |
| `profile_streaming_bottleneck_gpu7.py` | GPU extraction profiling |

### Data preparation and audit

| File | Responsibility |
|---|---|
| `easycom_manifest.py` | EasyCom clip inventory and dense-timeline validation |
| `prepare_easycom_session.py` | Optional strict-grid EasyCom media preparation |
| `prepare_ego4d_2fps.py` | Optional strict-grid Ego4D media preparation |
| `prepare_encoder_only_features.py` | Convert local sidecars into encoder-only references |
| `audit_egospeak_features.py` | Audit released EgoSpeak feature layout |
| `derive_easycom_interaction_events.py` | Deterministic transcript/timing proxies |
| `adjudicate_easycom_interaction_events.py` | Optional ambiguous-event adjudication |
| `audit_easycom_floor_events.py` | Count EOT/takeover/backchannel candidates |
| `prepare_easycom_eot_int_review.py` | Retain every raw wearer onset and offset as an unresolved semantic candidate; attach nearby participant boundaries, activity relations, transcripts, and target-of-speech facts without entry/takeover gating; also scan independent hard-negative regions; semantic labels remain AI/human review outputs |
| `adjudicate_easycom_eot_int_candidates.py` | Add optional DeepSeek transcript/timing suggestions; never human ground truth |
| `prepare_ego4d_av_eot_int_review.py` | Build the equivalent unresolved wearer onset/offset and hard-negative review queue from official Ego4D AV voice, transcription and annotation-only `is_at_me` fields without changing source boundaries |
| `adjudicate_ego4d_av_eot_int_candidates.py` | Apply the checkpointed strict LLM pre-label client with an Ego4D AV-specific evidence policy; every suggestion still requires human review |
| `review_floor_events.py` / `review_ego4d_av_events.py` | Serve the local EasyCom or Ego4D AV video/transcript workspace and atomically persist human event reviews |
| `prepare_easycom_oracle_features.py` | Build explicit past-only oracle person-activity sidecars and analysis-only addressee controls |

### Local features

| File | Responsibility |
|---|---|
| `extract_turn_local_features.py` | Qwen person-crop visual and recent-audio sidecars |
| `extract_ego4d_local_multigpu.py` | Foreground supervisor for disjoint multi-GPU workers |

### Training, inference and evaluation

| File | Responsibility |
|---|---|
| `train_turn_head.py` | Frozen-backbone head training, optional local fusion/auxiliary losses |
| `infer_turn_head.py` | Save EgoSpeak-compatible predictions and AP/mAP |
| `evaluate_floor_state_changes.py` | CPU floor-state/event proxy evaluation |
| `train_easycom_floor_event_probe.py` | Analysis-only direct response/yield probe with shared competing-risk labels |
| `summarize_easycom_floor_event_probes.py` | Compare direct event-probe conditions under one evaluation contract |
| `summarize_local_baselines.py` | Aggregate local baseline `metrics.json` files |
| `run_easycom_track_gru_residual.sh` | Matched pooled and anonymous track-GRU training/inference launcher |
| `run_easycom_oracle_information.sh` | Prepare and screen aggregate/per-person/shuffled/delayed oracle conditions on Session 12 |
| `run_easycom_floor_event_probes.sh` | Prepare, train and compare aggregate/addressee direct event probes |
| `run_easycom_post_boundary_eval.sh` | Compare existing dense runs after the aggregate-other boundary is observed |
| `train_omni_turn_lora.py` | Thinker LoRA training from cached AV |
| `eval_omni_turn_lora.py` | Thinker LoRA evaluation |

### Launchers and tests

- `run_*.sh` and `run_*.py` are reproducible launchers, not libraries.
- `smoke_test.py` is the model/environment smoke check.
- `test_*.py` are CPU-oriented regression tests colocated at root for current
  import and command compatibility.

## Terminology restrictions

- **Global (`G`)**: a Thinker readout that may mix audio, video and language
  context; do not describe it as audio-free.
- **Local audio (`A`)**: recent mixed audio encoded outside the Thinker.
- **Local visual (`V`)**: person-box visual features; current pooling discards
  stable person structure.
- **Encoder-only**: Qwen modality encoder output without Thinker states.
- **Speaker grounding**: evidence that the correct visible person is linked to
  the correct speech/time. AV improvement alone is insufficient.
- **Addressee grounding**: evidence about whom speech targets; aggregate
  `others` labels cannot supervise it.
- **Floor entry/yield**: observed wearer OFF→ON / ON→OFF behavior under a
  defined causal event window; not automatically semantic response or
  interruption intent.

## Environment boundaries

- The open-stream replay implementation is pinned to
  `transformers==4.52.3` because it uses version-specific internals.
- The local encoder path has been tested with 4.52.3 and 4.57.6; sidecars from
  one experiment must share the same recorded contract/version.
- Remote paths differ (`/data/...` and `/data1/...`). Commands should expose
  dataset/model roots as variables rather than hard-code one machine.
- Large models, datasets, checkpoints and extracted features live outside Git.
  The explicitly audited 72 MiB MuVAP ASD-CPC release and 7 MiB CPC frontend
  are a deliberate exception: they are vendored with their exact source and
  provenance under `third_party/MuVAP` for reproducible extraction.