data/validation/README.md

# validation — the 2026-09-09 calibration epoch

Everything recorded on 2026-09-09 lives here, kept out of `data/motherboard`
and `data/pushT` because it does not share those trees' calibration epoch.
Same rig, same processing pipeline; a separate folder so nothing silently
inherits the wrong extrinsics. `calibration/` holds the epoch solved that
morning at 07:41 (`calib_points_20260909_074139.json`).

9 episodes, 25.2 minutes, two tasks.

| episode | task | frames | duration | segments | bad frames | source recording |
|---|---|---|---|---|---|---|
| episode_000 | motherboard | 6661 | 222.0 s | 1 | 0 | `motherboard/2026-09-09/episode_000` |
| episode_001 | motherboard | 6755 | 225.2 s | 1 | 0 | `motherboard/2026-09-09/episode_001` |
| episode_002 | motherboard | 5647 | 188.2 s | 1 | 0 | `motherboard/2026-09-09/episode_002` |
| episode_003 | pushT | 13484 | 449.5 s | 62 | 1147 | `pushT/2026-09-09/episode_000` |
| episode_004 | motherboard | 1995 | 66.5 s | 1 | 0 | `motherboard/2026-09-09/episode_000_seg00` |
| episode_005 | motherboard | 1870 | 62.3 s | 1 | 0 | `motherboard/2026-09-09/episode_001_seg00` |
| episode_006 | motherboard | 1465 | 48.8 s | 1 | 0 | `motherboard/2026-09-09/episode_001_seg01` |
| episode_007 | motherboard | 4489 | 149.6 s | 1 | 0 | `motherboard/2026-09-09/episode_001_seg02` |
| episode_008 | motherboard | 2942 | 98.1 s | 1 | 0 | `motherboard/2026-09-09/episode_001_seg03` |

## Episode numbers do not identify a recording

The recorder reuses episode numbers within a date. `motherboard/2026-09-09/episode_000.h5`
was written **four times** that day, each write replacing the last.
`episode_000`–`episode_002` here are the 04:00 session; `episode_004`–`episode_008`
are the 17:20/17:41 session — different recordings that happened to be given
the same numbers.

So the arrivals were renumbered onto the end of the folder rather than merged
by name, and every row of `episodes.jsonl` carries `source_recording`, naming
the file it actually came from. `task` is likewise per-episode now:
`segments.json` and `bad_frames.json` used to carry one task at the top, and
this folder holds two.

`episode_003` was previously published as `data/pushT_2026-09-09/episode_000`
and is the same data, renumbered. That path has been removed.

## Two cutting policies

| episodes | floor | how they were published |
|---|---|---|
| 000–003 | 16 frames | whole episodes; `segments.json` indexes the clean spans inside them |
| 004–008 | 900 frames (30 s) | already cut — each file **is** one clean span, so it carries no defects |

For 000–003 you must honour `segments.json`: those videos contain flagged
frames. For 004–008 you need not — the flagged frames are the gaps *between*
episodes, which is why their coverage of the recording is not contiguous.

The withdrawn `pushT_2026-09-09` folder shipped a segment index claiming
`n_segments: 79` over an empty list, while its own episode index said 62. Recomputing the clean spans from
`bad_frames.json` at the documented 16-frame floor reproduces 62 spans and
1147 bad frames exactly, so the merged file carries the rebuilt list.

## Format

`meta/<date>/<episode>.parquet`, `videos/<date>/<episode>/*.mp4` — three
RealSense views (`view_left`, `view_middle`, `view_right`), two GelSights
(`tactile_left`, `tactile_right`) and the two Arducam wrist cameras
(`wrist_left`, `wrist_right`). Poses are Z-up.

`duration_s` throughout is `n_frames / 30`, the nominal write tick. The
RealSense streams actually run at 29.80 Hz and the GelSights reach about
17.8 Hz against that tick, so roughly 40 % of tactile frames repeat the
previous one — `tactile_*_is_new` marks which are real.

Two gaps worth knowing:

* `depth/` covers `episode_000`–`episode_003` only. The 17:20/17:41 session
  was published without depth.
* `previews/` covers every episode except `episode_003`: the preview panel is
  rendered from the source recording, and that H5 has been deleted.