# Artifacts

The `artifacts/` directory contains small frozen CSV/JSON files used for Level 1 reproduction. These are not full result trees and do not include videos.

## `artifacts/original_n5/`

### `main_results_aggregate.csv`

Frozen aggregate table for the original N=5 generative benchmark.

Key columns:

- `model`
- `resolution`
- `n`
- `mad_mean`
- `mad_std`
- `sec_per_frame_mean`
- `sec_per_frame_std`
- `vram_mean`
- `vram_std`
- `mad_step_1` through `mad_step_8`

Consumed by:

- `scripts/make_release_artifacts.py`

## `artifacts/expanded_n30/`

### `extended_metrics_aggregate.csv`

Frozen aggregate MAD/SSIM rows for N=30 methods.

Key columns:

- `method_type`
- `method`
- `resolution`
- `n`
- `mad_mean`
- `mad_std`
- `mad_ci95_*`
- `ssim_mean`
- `ssim_std`
- `ssim_ci95_*`

### `extended_metrics_per_clip.csv`

Per-clip rollout MAD/SSIM rows for N=30 methods.

Key columns:

- `method_type`
- `method`
- `resolution`
- `clip_id`
- `logical_clip_id`
- `rollout_mad`
- `rollout_ssim`
- `runtime_sec_cpu`

Path-bearing provenance fields are sanitized for the release.

### `extended_metrics_per_step.csv`

Per-step MAD/SSIM rows for N=30 methods.

Key columns:

- `method_type`
- `method`
- `resolution`
- `clip_id`
- `prediction_step`
- `mad`
- `ssim`

Consumed by:

- `scripts/make_release_artifacts.py`

### `paired_baseline_comparisons.csv`

Paired clip-level differences between LTX-2B and each baseline.

Key columns:

- `model`
- `baseline`
- `resolution`
- `metric`
- `n`
- `mean_difference`
- `std_difference`
- `ci95_*`
- `model_wins`
- `clip_ids`

### `cpu_baseline_runtime_aggregate.csv`

Aggregate CPU baseline timing breakdown.

Key columns include method, resolution, clip count, OpenCV thread count, decode timing, and per-component algorithm timings.

### `qualitative_candidate_selection.csv` / `.json`

Deterministic numerical candidate-selection provenance for N=30 qualitative review.

Key fields:

- `selection_category`
- `clip_id`
- `LTX_MAD`
- `Persistence_MAD`
- `History_Farneback_MAD`
- pairwise MAD differences
- SSIM values
- `reason_selected`

### `qualitative_selection.json`

Final compact qualitative-selection provenance used by the N=30 artifact script.

## Regenerating Level 1 Outputs

```bash
python scripts/make_release_artifacts.py \
  --artifacts-dir artifacts \
  --out-dir outputs/release_level1
```