icassp2027-experiment/docs/easycom_event_review_tool.md

# EasyCom wearer-entry/INT pre-label and review pipeline

The workflow has three explicit stages. Speech activity timestamps are
read-only throughout; reviewed events never modify the dense 5 Hz activity
ground truth.

```text
EasyCom participant speech intervals
  -> all wearer boundaries + independent rule regions + structured factual evidence
  -> DeepSeek semantic pre-label (optional but recommended)
  -> human Accept / Relabel / Exclude
```

## 1. Generate candidates

Develop the rules on training-side Sessions 4--12:

```bash
python3 prepare_easycom_eot_int_review.py \
  --dataset-root /path/to/easycom \
  --sessions Session_4 Session_5 Session_6 Session_7 Session_8 Session_9 Session_10 Session_11 Session_12 \
  --output-dir reports/easycom_floor_event_candidates
```

The generator emits two independent files:

- `easycom_eot_int_candidates.jsonl` (legacy-compatible filename) contains every
  raw wearer `0->1` boundary (`WEARER_ONSET`, `WEARER_ENTRY_CANDIDATE`), every
  raw wearer `1->0` boundary (`WEARER_OFFSET`, `YIELD_CANDIDATE`), and hard-negative
  regions that need no wearer transition:
  `SAME_SPEAKER_PAUSE`, `OTHER_TO_OTHER`, `OTHER_BACKCHANNEL`, and
  `NON_TAKEOVER_OVERLAP`;
- `easycom_eot_int_controls.jsonl` is retained for file-format compatibility
  and legacy inputs. The current boundary protocol does not deterministically
  demote wearer onsets or offsets into this file.

No short-gap merge or takeover heuristic controls wearer-boundary admission.
For each raw onset, `temporal_evidence` stores the previous wearer offset,
silence duration, every other-speaker interval intersecting that gap, interval
coverage, whether it continues across the onset, transcript, and
`Target_of_Speech`. DeepSeek and the human reviewer decide whether the onset is
a genuine entry, within-turn resumption, backchannel, or excluded non-event.

Entry construction searches backward for nearby non-wearer releases as optional
context. Every option carries factual flags for intervening wearer activity,
intervening other onsets, and activity at the wearer onset; the generator does
not declare one a valid conversational cause. Each wearer onset produces exactly
one primary row. When present, `opportunity_options` contains the only release
anchors that DeepSeek or a human may optionally select; neither may invent a
timestamp.

All wearer onsets and offsets carry the rule-stage value `candidate`; timing
alone is not a semantic positive label. For each offset, the JSON records the
wearer activity-run start, nearby other intervals, pre-existing and new-other
flags, activity immediately before/after the boundary, strict and configurable
50-ms-tolerant crossing, overlap duration, and subsequent speech onsets.
`wearer_start < other.start < wearer_offset < other.end` is evidence rather than
an INT prerequisite. Hard-negative scans are rule hints and still require AI
and human review. The JSONL retains candidate source, original event time,
participant-level activity, raw durations/gaps/overlap evidence, transcript,
and `Target_of_Speech` when present. `Target_of_Speech` is annotation assistance
only and cannot be a deployable model input.

All heuristic durations are configurable rather than frozen paper rules:

```text
--relation-window-seconds
--same-speaker-pause-max-seconds
--other-handover-max-seconds
--short-vocalization-max-seconds
--continuation-min-seconds
--int-offset-tolerance-seconds
```

Use `--max-candidates 1` for a one-event smoke test. `summary.json` records the
exact configuration and candidate counts. `easycom_session_timelines.json`
stores the read-only participant speech lanes used by the full-session web
timeline; it is not sent to DeepSeek.

## 2. Add DeepSeek pre-labels

Prompt v14 first reconstructs the exact participant-resolved chronology and
then uses transcript content and `Target_of_Speech` to determine conversational
function. It treats the EasyCom transcript marker `[L]` as laughter-only evidence,
not as a lexical floor take. Wearer laughter during another participant's
retained turn is directed toward `ENTRY_WEARER_BACKCHANNEL`; non-wearer laughter
overlapping a retained wearer turn is directed toward `INT_OTHER_BACKCHANNEL`.
Ambiguous laughter falls back to `EXCLUDE`, and `[L]` alone is neither evidence
of a floor take nor evidence of a visible smile. Preceding release options are
optional entry metadata. No entry/within-turn or INT/takeover semantic decision
is made before the prompt. A positive `WEARER_ENTRY` may optionally select one
supplied release; entries from silence or during other activity use `null`.

`needs_human_review` is not requested from DeepSeek: every pre-label enters the
human queue by protocol, so the adjudicator assigns that boolean locally. This
prevents a string-valued model field from invalidating and retrying an otherwise
valid batch.

First inspect exactly what would be uploaded:

```bash
python3 adjudicate_easycom_eot_int_candidates.py \
  --input reports/easycom_floor_event_candidates/easycom_eot_int_candidates.jsonl \
  --output reports/easycom_floor_event_candidates/easycom_eot_int_deepseek.jsonl \
  --model deepseek-v4-flash \
  --max-events 1 \
  --dry-run
```

After setting `DEEPSEEK_API_KEY` and obtaining authorization to upload the
candidate excerpts:

```bash
python3 adjudicate_easycom_eot_int_candidates.py \
  --input reports/easycom_floor_event_candidates/easycom_eot_int_candidates.jsonl \
  --output reports/easycom_floor_event_candidates/easycom_eot_int_deepseek.jsonl \
  --model deepseek-v4-flash \
  --allow-external-transcript-upload
```

For SiliconFlow, keep the same API-key environment variable and select its
OpenAI-compatible endpoint explicitly:

```bash
python3 adjudicate_easycom_eot_int_candidates.py \
  --input reports/easycom_entry_int_sessions1_3_v14/easycom_eot_int_candidates.jsonl \
  --output reports/easycom_entry_int_sessions1_3_v14/easycom_eot_int_deepseek.jsonl \
  --base-url https://api.siliconflow.cn/v1 \
  --model deepseek-ai/DeepSeek-V4-Flash \
  --allow-external-transcript-upload
```

The default request batch remains 8. If a returned batch violates the strict
label schema, the adjudicator isolates the bad record by recursively splitting
8 to 4, 2, and 1. Each successful leaf batch is atomically recorded in
`OUTPUT.checkpoint.json`; rerunning the identical command resumes from that
checkpoint and does not request completed passes again. A singleton that still
fails after `--max-retries` is written to `OUTPUT_STEM_pending.jsonl` for manual
relabeling, while the remaining candidates continue. Provider/network failure
does not create a semantic pending item: the command stops with its checkpoint
intact and the identical command resumes later. Use `--checkpoint` or
`--pending-output` to override the derived paths. `--overwrite` intentionally
starts a fresh run instead of resuming.

DeepSeek sees candidate source, structured timing, per-speaker activity, and
transcript together. It returns one of `POSITIVE`, `HARD_NEGATIVE`, or
`EXCLUDE`; positives additionally receive
`WEARER_ENTRY` or `INT_YIELD`, and hard negatives receive a named subtype. It
cannot modify timestamps. The output preserves the original rule candidate,
the selected preceding-release UID when applicable, and the complete AI
decision/provenance. `EXCLUDE` is the sole non-scoring outcome, but it is never
an untyped bucket. Every AI and human exclusion stores one closed-vocabulary
`exclude_reason` and the derived `exclude_reason_group`. `OUT_OF_SCOPE_*` means
the annotations are valid but the behavior is outside the scored target;
`INSUFFICIENT_EVIDENCE` means the supplied valid material cannot resolve
conversational function; and `DATA_*` identifies unusable annotation,
recording, context, artifact, or timestamp quality. Model-pass disagreement is
recorded as `INSUFFICIENT_EVIDENCE`, never as data corruption.

Hard negatives use behavior-specific subtypes. ENTRY distinguishes a same-other
mid-turn pause, a different-other transition, and a wearer backchannel that
does not take the floor. INT first distinguishes an other backchannel, then a
non-backchannel overlap after which the wearer demonstrably holds or resumes
the floor, and finally another interpretable overlap that fails to become a
sustained takeover. `OTHER_HARD_NEGATIVE` is reserved for valid false-positive
opportunities outside these named behaviors; it is never an uncertainty
fallback. Prompt v15 gives adjacency-pair completion priority over duration:
a short but complete answer to an addressed question remains a positive wearer
entry, and floor entry does not require the wearer to retain the next turn.
Later resumption by the questioner or another speaker therefore does not turn
the completed answer into a wearer backchannel. `ENTRY_WEARER_BACKCHANNEL` is
reserved for feedback that does not answer a question, accept or reject a
proposal, provide requested information, or otherwise complete an adjacency
pair.

During live adjudication, the CLI displays a dependency-free progress bar over
all candidate/pass work, including the current pass, batch, percentage, and
elapsed time. Redirected logs emit periodic progress lines instead of carriage
return animation.

Every current `WEARER_ONSET` and `WEARER_OFFSET` row is sent for semantic
adjudication regardless of its timing pattern. Explicit control rows from older
protocol files remain excluded unless controls are intentionally reviewed.

## 3. Human review

Run the web workspace on the AI-enriched JSONL:

```bash
python3 review_floor_events.py \
  --dataset-root /path/to/easycom \
  --events reports/easycom_floor_event_candidates/easycom_eot_int_deepseek.jsonl \
  --output reports/easycom_event_reviews/reviews.jsonl
```

Open `http://127.0.0.1:8765`. The page shows a session candidate overview,
participant activity near the event, video/audio, transcript, structured timing,
and the AI verdict/confidence/reason. A decision guide first states the behavioral
question, then presents the exact AI verdict/subtype with a plain-language
meaning; the Accept control repeats that exact label so its effect is explicit.
The common path is one operation:

- **Accept** keeps the AI verdict/type and optional preceding-release metadata;
- **Relabel** expands verdict and subtype controls; a positive entry may
  optionally select one of the supplied release anchors;
- **Exclude** requires a concrete out-of-scope or data-quality reason.

The current primary-review UI deliberately has no second-review action. Older
saved `second_review` records remain readable for backward compatibility.

Boundary correction is hidden under an optional disclosure and is limited to
one 200 ms bin. The original timestamp is always retained. Each save atomically
rewrites JSONL and CSV while preserving the rule candidate, AI decision, human
decision, relabel flag, exclusion state, confidence, and provenance. The current
UI does not ask for a reviewer name; the export retains an empty `reviewer` field
only for compatibility with older review files.

The primary review queue filters only explicit legacy control rows. To audit
those controls intentionally,
pass `--events .../easycom_eot_int_controls.jsonl --include-controls` and use a
separate review output file. The old `--include-offset-controls` spelling
remains an alias.

Protocol wording for the paper appendix:

> Every observed wearer speech onset and offset is retained as a semantic review
> candidate. Deterministic construction records only participant-resolved timing,
> transcript, and target-of-speech facts; it does not infer within-turn
> resumption, floor entry, interruption, or takeover.

## Freeze discipline

Develop the guide and resolve disagreements only on Sessions 4--12. Freeze
categories and thresholds before reviewing official test Sessions 1--3. Use two
independent human review files plus a separate adjudication output for paper-
quality labels. DeepSeek is pre-label metadata, not an independent reviewer.