VALIDATION.md

# Validation Report

## Scope

The current post-v0.3 active-construction candidate validates locally and
against an immutable pull-request revision:

- all 385 canonical Standard rows and candidate direct Standard MP4 paths
- all 128 active Pre-Standard rows and direct Pre-Standard MP4 paths
- persistent record identity and cross-tier uniqueness
- locally audited byte sizes and technical metadata for the 22 new fragments
- container-header technical metadata for every active or candidate MP4
- candidate filename rules, field types, numeric ranges, and stored-orientation
  consistency
- a 27-row, 15-field relationship sidecar covering 22 canonical fragments and
  five noncanonical derivatives across five complete sequences

One 6,587,685,497-byte provenance object is promoted to the canonical root,
and the distinct 496,296,535-byte object previously occupying its corrected
name is moved to its curator-confirmed April 4 path. No video bytes are
transcoded, replaced, downloaded in full, or newly uploaded. The promoted
object's properties were read from its public MP4 header using HTTP range
requests. Sizes and content identities were read from the Hub repository
inventory.

The later Standard ingestion adds 17 creator-verified media objects totaling
9,289,459,921 bytes. Every addition was fully decoded before upload, all 17
local SHA-256 identities are unique, and none matches an existing Hub media
object. The batch is indexed without changing the 29-field schema.

The July 13 tier promotion moves one existing 141,221,810-byte object from
Pre-Standard to Standard without changing its payload, Xet identity, or stable
record ID. All 1,205 video frames and the audio stream decode successfully.
The clip begins after drawing has started and reaches a completed, signed
endpoint. The promotion moves the protocol identifier coherently to
`standard-2025-07-13` and preserves the 485-record total.

The August 2026 Standard ingestion adds three creator-verified media objects
totaling 50,478,909,746 bytes. All 231,383 video frames and every audio stream
decode successfully. The batch includes the first 14 × 17 in Standard record,
retains `Rec01` because `Rec02` is expected, and preserves the
existing 29-field schema and protocol identifier.

The August 22–23 recording-sequence ingestion adds three creator-verified media
objects totaling 51,022,796,325 bytes. All 233,723 video frames and every audio
stream decode. Rec03 produced one end-of-stream non-monotonic DTS warning from
FFmpeg's null output during combined-stream validation; direct inspection found
strictly increasing source packet and decoded-frame timestamps at the tail. The
batch retains literal `Rec01` through `Rec03` sequence suffixes and documents a
creator-confirmed Rec03 change from Pelikan black to unmarked, unknown-brand
black ink at approximately `00:04:59`.

The August 24–27 fragment-sequence ingestion adds 22 creator-confirmed,
unchanged Framelapse source objects totaling 17,389,875,055 bytes and
3,391.699091 seconds of encoded playback. They represent five complete
sequences and are stored as landscape video without audio. Five unchanged
InShot outputs remain nested noncanonical provenance derivatives and are
excluded from active row, byte, duration, orientation, and audio totals. The
relationship sidecar records all 27 media objects without changing the
29-field video schema, `standard-2025-07-13`, or historical rows, and without
introducing a global `work_id`.

## Result

**LOCAL STRUCTURAL AND IMMUTABLE-PR PASS — zero validation errors**

| Check | Result |
| --- | ---: |
| Active indexed rows | 513 |
| Canonical Standard rows | 385 |
| Pre-Standard rows | 128 |
| Metadata columns per index | 29 |
| Unique persistent `record_id` values | 513 |
| Complete encoded-video metadata with remote Xet binding | 513/513 |
| Stored-orientation conflicts | 0 |
| Duplicate local SHA-256 values among the 27 new media objects | 0 |
| New media objects remotely bound by SHA-256, size, and Xet identity | 27/27 |
| Fragment-sequence relationship rows | 27 |
| Relationship fields | 15 |
| Canonical fragments / noncanonical derivatives | 22 / 5 |
| Owner-confirmed complete sequences | 5 |
| Base-to-metadata-revision diff | 39 expected paths; 0 deletions |
| Inventory-binding regression tests | 5/5 |

**REMOTE PR INVENTORY AND RELATIONSHIP VALIDATION — PASS**

The sidecar's `source_hub_revision` pins the immutable metadata-bearing PR
commit `8a21015210c19936d3b18175ae3f4061ae1e5d8d`. At that revision, validation
confirmed 385/385 direct Standard paths, 128/128 direct Pre-Standard paths,
all 27 relationship paths, ordinary SHA-256 values, byte sizes, Xet identities,
and all 22 canonical `record_id` joins. The remote root `metadata.csv` receipt
matches the local 208,651-byte file at SHA-256
`0de0b9111effa84cd41b13f93e6627b9b94f33f648dd29ded5d4eb8a85c5bfad`.
The relationship validator reported 27 rows, 22 canonical fragments, five
noncanonical derivatives, five complete sets, and zero errors.

The media-only PR head `bb4160045cae21a71bf360278549398dc9d2a028`
contained exactly the 27 planned additions relative to the pinned base, with
zero deletions or changed existing paths. The metadata-bearing revision has
exactly the expected 39-path cumulative diff and zero deletions. Viewer
settlement and merge are separate gates and are not asserted by this PASS.

Both Hub builder configurations explicitly declare the same 29 features. This
prevents per-CSV inference from representing an integer-only canonical
`encoded_frame_rate` column as `int64` while representing historical
fractional rates as `float64` in the Pre-Standard configuration.

## Audited statistics

- Total indexed playback duration: 53.590 hours
- Canonical playback duration: 43.994 hours
- Pre-Standard playback duration: 9.596 hours
- Median active-record playback duration: 310.172 seconds
- Total indexed video size: 2,020,171,359,199 bytes (2.020 TB decimal)
- Stored orientation: 479 portrait, 34 landscape
- Audio streams: 485 present, 28 absent
- Unknown finish dates: 9
- Unknown finish times: 62

## Outstanding warnings

1. Fifteen canonical 30×40 records use `support = Paper`, while
   `Archive_Specifications.md` describes the 30×40 support as Illustration
   Board. The curator must determine whether the rows or the specification are
   wrong; this release candidate does not manufacture a correction.
2. Three creator-confirmed Standard records are shortened derivatives rather
   than full-duration masters. Their start/end limitations remain explicit in
   `notes`; none is represented as a complete capture lineage. The promoted
   July 13 record starts mid-process and reaches a completed, signed endpoint.
3. The 2026-01-11 and 2026-02-08 additions each produced one non-monotonic
   duplicate video-timestamp warning during validation but decoded completely.
   Rec03 from 2026-08-23 produced one end-of-stream non-monotonic DTS warning
   from FFmpeg's null output during combined-stream validation, while source
   packet and decoded-frame timestamps at the tail remained strictly increasing.
   The 2026-01-11 recording also has pronounced lighting variation.
4. Five Pre-Standard records retain `time = 00:00` while their finish date is
   unknown. Confirm whether those are documented midnight times or historical
   placeholders before normalizing them.
5. Two Pre-Standard records retain incomplete material fields because the
   filename evidence is insufficient or ambiguous.
6. Most masters retain an audio stream. Stream presence does not establish
   meaningful signal, and rights/privacy review remains necessary.
7. Encoded frame rate is playback rate and cannot be used as original
   drawing-time sampling.
8. The 2012–2016 layer has no public video payload at this snapshot.
9. Separate Hub configurations prevent tier mixing, but no leakage-safe
   train/validation/test partition has been defined. The dated
   `fragment_set_label` is not a global `work_id`, so no work-level split is
   currently possible.

## Reproduction

Validate both active indexes:

```bash
python tools/validate_metadata.py metadata.csv \
  --pre-standard-metadata Pre_Standard_Time_Lapses/metadata.csv
```

For the strongest repository-inventory check, first augment the pinned Hub API
JSON with the Xet file IDs returned by the same revision's resolve headers:

```bash
curl --fail --location \
  --output repo_api_blobs.json \
  "https://huggingface.co/api/datasets/maxwellinked/time-lapse-artifacts/revision/<40-hex-commit>?blobs=true"

python tools/capture_hub_inventory.py \
  --repo-json repo_api_blobs.json \
  --repo-id maxwellinked/time-lapse-artifacts \
  --revision "<40-hex-commit>" \
  --output repo_inventory_xet.json
```

The helper also fetches the small root `metadata.csv` at that immutable
revision and records its exact SHA-256, byte size, and `X-Repo-Commit` receipt
under `validationFiles`. The relationship validator rejects a local metadata
file whose bytes do not match that receipt.

Then validate paths, Xet identities, and exact byte sizes together:

```bash
python tools/validate_metadata.py metadata.csv \
  --pre-standard-metadata Pre_Standard_Time_Lapses/metadata.csv \
  --repo-inventory-json repo_inventory_xet.json
```

Validate the dated relationship sidecar against the immutable metadata-bearing
revision recorded in `source_hub_revision`. Confirm all 27 unique paths, Xet
identities, and sizes; all 22 canonical `record_id` joins; contiguous `RecNN`
positions; five complete ordered source sets; derivative source lists limited
to their own set; and exclusion of the five nested derivatives from both active
indexes and Viewer configurations.

```bash
python tools/validate_fragment_relationships.py \
  --relationships Standard_Time_Lapses/Provenance/2026-08-27/fragment_sequence_relationships.csv \
  --metadata metadata.csv \
  --repo-inventory-json repo_inventory_xet.json
```

The inventory-binding regression suite is self-contained and performs no Hub
writes:

```bash
python tools/test_inventory_binding.py
```

A plain JSON list of repository paths is also accepted through the legacy
`--repo-files-json` spelling, but that checks membership only; it cannot
independently verify hashes or byte sizes.

Use `--viewer-json` only with a Viewer snapshot produced from the same
repository revision and configuration under review.

The Pre-Standard technical enrichment is reproducible from a pinned Hub API
response containing `siblings` with LFS metadata and an FFmpeg executable:

```bash
python tools/enrich_prestandard.py \
  --metadata Pre_Standard_Time_Lapses/metadata.csv \
  --repo-json repo_inventory_xet.json \
  --revision "<revision-under-review>" \
  --ffmpeg /path/to/ffmpeg \
  --cache prestandard_probe_cache.json \
  --output Pre_Standard_Time_Lapses/metadata.csv
```

The revision is intentionally explicit: reproducing against a moving branch
would weaken both identity and technical-metadata claims.

The canonical v0.2 index remains reproducible from its pinned 335-row source
plus the three audited additions:

```bash
python tools/repair_metadata.py source_metadata.csv metadata.csv \
  --additions tools/metadata_additions_v0.2.csv
```