# Paired NTU60 benchmark examples

Recorded 15 September 2026. Both models evaluated the same 16,487 official NTU60 cross-subject test clips, using released HRNet 2D postures. These are pre-cut action labels, not a test of continuous segmentation.

The compact model is the saved final-epoch LegibleTCN seed-42 checkpoint from the earlier suite. Its per-class correct counts and total accuracy were reproduced before pairing. The earlier landing-page table averages three seeds, while this gallery uses one saved model.

| Reference | Compact model | Measured reference | Both correct | Reference only | Compact only | Both wrong |
|---|---:|---:|---:|---:|---:|---:|
| ST-GCN++ joint stream | 76.12% | 89.29% | 12017 | 2705 | 533 | 1232 |
| PoseC3D joint stream | 76.12% | 93.73% | 12307 | 3147 | 243 | 790 |

PYSKL source is pinned to f2bf3a6b08e2e8dec744692d64efdb187fd6719a. ST-GCN++ uses two actors and ten 100-frame test samples. PoseC3D uses ten 48-frame samples and flipped heatmaps: twenty views. Published model weights, inputs and test views are retained; published scores are not substituted for these measured predictions.

## Example selection

Each group shows three clips sorted by stable sample ID, preferring distinct action labels. Confidence does not influence selection. Group totals cover the complete split. Benchmark labels establish correctness here; the selected animations have not undergone a separate manual label audit.

Each posture preview retains fixed image coordinates and up to 96 original source frames, with the first two actors shown. Preview playback is normalized to three seconds because the pose archive supplies frame indices rather than timestamps. It is not an inference-speed or detection-delay measurement.

## Reusable evidence

- [PoseC3D examples and provenance](posec3d-gallery.json)
- [ST-GCN++ examples and provenance](stgcnpp-gallery.json)
- [All compact-model probabilities, labels and stable IDs](ours-predictions.json.gz)
- [All official PoseC3D output probabilities](posec3d-predictions.pkl)
- [All official ST-GCN++ output probabilities](stgcnpp-predictions.pkl)
- [Gallery checksums](gallery-files.sha256)
- [Reference checkpoint checksums](reference-files.sha256)

Reference pickle rows follow PYSKL dataset.video_infos order. Pairing verified the full ID set, labels, probability sums, dataset hash and official label-map hash. Run-level source hashes are retained in the gallery JSON.

## Numerical execution

PoseC3D runs four videos per GPU batch with cuDNN algorithm autotuning. The first 64 official clips kept identical top-1 labels compared with the original batch-one execution. The largest probability difference was 0.0001862645149230957 (0.01863 percentage points). This failed the initial 0.0001 allclose check; that receipt is preserved. A separate recorded decision accepted a 0.001 absolute probability tolerance with identical prefix labels. This does not establish identical predictions on every clip; the complete run supplies its own measured score. The checkpoint, twenty test views and default TF32 permissions remain unchanged.

- [Original numerical check](posec3d-batch4-validation.json)
- [Recorded execution decision](posec3d-batch4-acceptance.md)
- [Decision and bound source hashes](posec3d-batch4-acceptance.json)
- [Derived batch configuration](posec3d-batch4.py)
- [Execution source checksums](posec3d-batch4-sources.sha256)
