# Frozen 12-task IND and four-task OOD evaluation

## Inputs and entrypoints

IND uses 1,200 rows from `repro/processed/setting_12_ind_test/processed_dataset`: 50 success and 50 failure per training task. OOD uses 400 rows from `repro/processed/setting_4_ood_test/processed_dataset`: 50/50 for each ID listed in `task_registry.json`. Both use the exact saved NPZ arrays and labels.

The two evaluators under `scripts/frozen/` are byte-for-byte snapshots of the actual experiment scripts, including their historical comments. Their inherited top-level docstring still mentions the original two-task experiment; the task tables, wrapper arguments and frozen run manifests define the actual 12-task/4-task evaluation here.

```bash
python /path/to/release/scripts/evaluate.py \
  --robometer-repo /path/to/robometer --split test_ind \
  --checkpoint-root /path/to/setting-a-checkpoints \
  --base-model /path/to/recorded-initialization-checkpoint \
  --output /path/to/new-ind-results --steps 0 100 200 300 400 500 \
  --setting A --batch-size 4 --dry-run
```

Remove `--dry-run` to execute. Use `--split test_ood` for the four-task OOD protocol. `--base-model` means the step-0 Robometer checkpoint, not the Qwen backbone by itself. Settings B/C may use the same test inputs but require their own compatible checkpoint roots.

## Metrics and historical protocol distinctions

- **Progress:** choose up to eight prefix ends from frame index 1 through the last cached frame; each query uses up to eight frames from that prefix. Compare each prefix's final prediction with its corresponding target. Report per-trajectory MSE and Spearman and aggregate the results. MAE may be recomputed from saved prediction/target details; it is not the evaluator's directly emitted progress metric.
- **IND failure detection:** the frozen script predicts failure when a five-point Pearson window is non-increasing (`correlation <= 0`) **and** the final success-head probability is `<= 0.5`. It also reports score-based metrics using `1 - success_probability`.
- **OOD failure detection:** the bundled later script uses a five-point Pearson threshold of `-0.2` **or** final predicted progress `< 0.5`. This differs from the IND rule and must be reported explicitly.
- **Filtering emitted by the frozen evaluator:** task-conditioned prediction traces are ranked separately by Pearson, final progress, delta, slope and late-mean/delta score, excluding the final predicted point. The emitted metric is top-30 correctness, where a correct item must be a successful trajectory of the query task.
- **Later four-task filtering analysis:** the historical final aggregation separately used a four-task candidate pool, top-50, and fixed score ranges (Pearson [-1,1], final [0,1], delta [-1,1]). Its manifest is retained in `protocols/ood_final_filtering_manifest.json`. It is a separate postprocessing protocol, not the top-30 metric emitted by the bundled evaluator. This package does not invent the missing exact aggregation implementation or claim to reproduce its published numbers automatically.

Step 0 is the supplied initialization model. Other candidates must be real `checkpoint-<step>` directories containing `config.json`; missing checkpoints are not fabricated. Use separate output directories for different protocol choices. Workers can receive disjoint step lists via `--steps` and distinct `--worker` values. Resume manifests reject incompatible selections.

## Outputs and validation boundary

Each output contains `run_manifest.json`, progress state, per-step details/metrics and resumable phase files. Compare protocol settings before comparing numbers across historical folders. The reference run records in `protocols/` identify dataset counts and worker step selections; internal source paths are replaced by release-relative paths or external-dependency descriptions.

Release preparation validated data loading, typed video decoding, label alignment and model-input batch construction. It did not run checkpoint inference, train a model, tune thresholds or recompute paper metrics. Checkpoint provenance and the source-specific Robometer modifications remain external dependencies listed in `MODEL_REQUIREMENTS.md`.