EasyCom human-annotated fixed-horizon floor-event evaluation

This protocol evaluates saved EgoSpeak-layout dense predictions against human-confirmed WEARER_ENTRY, INT_YIELD, and hard-negative opportunities. It is separate from the timing-derived proxy in evaluate_floor_state_changes.py.

Scoring contract

At each nominal 0.2--2.0 second horizon:

a positive entry is one corrected 5 Hz behavior-anchor bin, scored by wearer activity probability;
a positive INT-yield is one corrected 5 Hz behavior-anchor bin, scored by one minus wearer activity probability;
a hard negative is one opportunity sample, scored by the maximum erroneous action probability over its annotated half-open region;
a positive timestamp is never expanded into a tolerance window;
an excluded or unreviewed candidate never enters scoring.

The prediction layout is [target_time, class, future_horizon]. Event frames are rasterized from the original 20 Hz annotation with the same frame // 4 rule used by the EasyCom event derivation. A human -0.2/0/+0.2 second boundary correction shifts the resulting target bin; for hard negatives it shifts the whole region.

The reported horizon is the nominal EgoSpeak forecast-bin horizon. Because an acoustic boundary can occur anywhere inside a 200 ms target bin, do not reinterpret it as an exact continuous-time lead without separately reporting the quantization convention.

Provisional LLM-only pipeline smoke

Before human adjudication is complete, unanimous two-pass LLM suggestions can be materialized into the review-file schema for a pipeline smoke test:

python3 materialize_provisional_llm_reviews.py \
  --events reports/easycom_entry_int_session{1,2,3,4}_v14/easycom_eot_int_deepseek_official*.jsonl \
  --output reports/easycom_entry_int_sessions1_4_v14/easycom_eot_int_provisional_llm_reviews.jsonl \
  --acknowledge-provisional-llm-auto-accept

The converter skips pass disagreements, pending API failures, and candidates without an LLM suggestion. It writes JSONL, CSV, and a summary JSON. Every materialized row carries review_source=PROVISIONAL_LLM_AUTO_ACCEPT and human_reviewed=false; the nested human_status=REVIEWED value exists only for review-schema compatibility and must never be described as human review. The discouraged --include-nonunanimous option is available only for debugging the fallback path.

The evaluator rejects these rows by default. A provisional run requires an explicit opt-in and marks its output protocol and semantic scope accordingly:

python3 evaluate_annotated_floor_events.py \
  --reviews reports/easycom_entry_int_sessions1_4_v14/easycom_eot_int_provisional_llm_reviews.jsonl \
  --run egospeak=results/egospeak_miniroada_floor_eval \
  --sessions Session_4 \
  --allow-provisional-llm-reviews \
  --print-event-examples 5 \
  --output results/easycom_egospeak_event_smoke/session4_summary.json

This output is suitable for alignment debugging and preliminary analysis, not for a manuscript claim of human-adjudicated ground truth. Official test Sessions 1--3 still require disjoint training-side calibration unless the separate exploratory in-sample override is used.

Development smoke run

Use only training-side sessions while the annotation and region rules are being audited:

python3 evaluate_annotated_floor_events.py \
  --reviews reports/easycom_event_reviews/reviews.jsonl \
  --run egospeak=/path/to/egospeak/result \
  --sessions Session_12 \
  --matched-recalls 0.7 0.8 0.9 \
  --print-event-examples 3 \
  --output reports/easycom_annotated_floor_events/session12_summary.json

This selects thresholds on the same development events and labels the threshold source as exploratory. It is not a confirmatory test command.

Confirmatory test run

Thresholds must be selected independently for entry and INT at every horizon from training-side out-of-fold predictions:

python3 evaluate_annotated_floor_events.py \
  --reviews reports/easycom_event_reviews/test_reviews.jsonl \
  --run baseline=/path/to/baseline/test_result \
  --run candidate=/path/to/candidate/test_result \
  --sessions Session_3 Session_2 Session_1 \
  --calibration-reviews reports/easycom_event_reviews/oof_train_reviews.jsonl \
  --calibration-run baseline=/path/to/baseline/oof_result \
  --calibration-run candidate=/path/to/candidate/oof_result \
  --calibration-sessions Session_4 Session_5 Session_6 Session_7 Session_8 Session_9 Session_10 Session_11 Session_12 \
  --matched-recalls 0.7 0.8 0.9 \
  --print-event-examples 3 \
  --output results/easycom_annotated_floor_events_YYYY-MM-DD/summary.json

The CLI rejects evaluation involving official Sessions 1--3 unless calibration reviews and calibration sessions are supplied. The --allow-in-sample-thresholds override is exploratory only.

Outputs:

summary.json: protocol metadata, counts, ten horizons per family, AP and operating points;
summary_metrics.csv: one row per run, family, horizon, and requested recall;
summary_events.csv: one scored row per run, annotated opportunity, and horizon for alignment audits and paired analysis.

Before comparing models, verify identical positive, hard-negative, and invalid counts. The headline manuscript operating point is the nominal 0.2 second bin at matched recall 0.8; ten-horizon entry/INT AP and all horizon curves are supporting results.

Required audit before freezing

Print examples and verify the 20 Hz boundary, 5 Hz target row, horizon, and valid mask against the source timeline.
Confirm that every positive occupies one target row.
Confirm that each hard-negative region contributes one max-pooled score, not one negative per frame.
Audit continuous timeline segments without candidate hints to estimate candidate recall.
Freeze the annotation guide, region endpoints, boundary-correction policy, and out-of-fold calibration procedure before inspecting Sessions 1--3.

This evaluator does not emit committed actions or detection latency. Those require a separately frozen online controller with threshold, persistence, refractory period, matching interval, and competing-risk rules.

6.1 KiB Raw Blame History