# Changelog ## Unreleased — August 29–30 complete fragment-sequence ingestion - Added 39 creator-confirmed, unchanged canonical Framelapse source fragments totaling 17,470,306,160 bytes and 3,403.665326 seconds of encoded playback. - Preserved five complete source-recording sequences as ordered `RecNN` sets containing 9, 5, 4, 15, and 6 canonical fragments. - Increased active membership to 443 Standard plus 128 Pre-Standard records, or 571 total, without changing the 29-field schema or `standard-2025-07-13` protocol ID. - Added a separate 39-row, 15-field dated relationship sidecar containing only canonical fragments; no concatenated or transcoded derivative was added. - Left `original_filename` blank because, at ingest-audit time, the source files bore canonical archive names and no pre-rename device-name mapping was independently retained. - Verified 39/39 local SHA-256 values and complete packet-stream demuxes, plus 117/117 decoded start/middle/end frame samples. The literal container tag `com.android.capture.fps=5.000000` is retained without inferring real drawing time. - Bound the 39 media objects at immutable media-only PR revision `b7ac091a06917eb712a4493dbb763e751c2bf459`; the relationship sidecar pins metadata-bearing revision `7293edccff883b4c604d710addbf26c85910cfcc`. - Retained media-object-level `record_id` identity. This update does not add a global `work_id`, backfill historical rows, create a v3 release, or otherwise bump the dataset version. ## Unreleased — canonical tool-token and timing clarification - Corrected five active Standard paths whose tool tokens contradicted their metadata: three `BallpoinPen` typos, one split `Ballpoint.Pen` token, and one split `Fountain.Pen` token. - Retained every existing `record_id`, media payload, Xet identity, byte size, and technical metadata. This is a server-side path normalization, not a re-encoding or new recording. - Added validation requiring each declared Standard tool to appear as an exact, case-sensitive filename token. Tool and medium lists remain independent; no positional pairing is inferred. - Left revision-pinned historical relationship sidecars unchanged because their earlier paths remain correct evidence at their declared revisions. - Clarified that 0.2 seconds is the Standard protocol's nominal capture setting, not a confirmed per-record measurement or calibrated real-time basis. ## Unreleased — verification and documentation consolidation - Clarified that each active metadata row represents one published video object, not necessarily one drawing, session, or complete process. - Kept documented fragment-sequence and physical-reference relationships in separate sidecars without adding a global `work_id` or changing either 29-field active video index. - Replaced batch-by-batch README narration and stale Viewer migration counts with current, revision-bound summaries; added a provenance navigation index for the nine documented fragment sequences. - Corrected nine physical-reference `hash_state` values that claimed paired hashes despite blank paired filename/hash fields. The correction reports ledger availability only and does not reinterpret any photograph or owner evidence. - Added a dedicated physical-reference validator and regression tests; fragment validation now checks canonical byte sizes and states whether a pinned remote inventory was supplied. - Removed the unsupported `video-classification` task facet while retaining `feature-extraction` and `annotations_creators: expert-generated`. - This consolidation does not alter media, active video-index rows, record identities, or acquisition-protocol identifiers. ## Unreleased — August 7, 2025 date correction - Corrected creator-confirmed record `tla-c043b4f55b583c4e` from August 7, 2026 to August 7, 2025 in both its canonical path and metadata date. - Retained the unchanged 86,415,894-byte media payload, stable `record_id`, Xet identity, technical metadata, collection membership, and 23:06 time. - This is a descriptive date correction, not a new recording, re-encoding, or change to the `standard-2025-07-13` acquisition protocol boundary. ## Unreleased — August 28–29 complete fragment-sequence ingestion - Added 19 creator-confirmed, unchanged canonical Framelapse source fragments totaling 10,661,404,849 bytes and 2,089.745536 seconds of encoded playback. - Preserved four complete drawing sequences as ordered `RecNN` sets containing 10, 4, 2, and 3 canonical fragments. - Increased active membership to 404 Standard plus 128 Pre-Standard records, or 532 total, without changing the 29-field schema or `standard-2025-07-13` protocol ID. - Added a separate 19-row, 15-field dated relationship sidecar containing only canonical fragments; no concatenated or transcoded derivative was added. - Left `original_filename` blank because, at ingest-audit time, the source files bore canonical archive names and no pre-rename device-name mapping was independently retained. - Verified 19/19 local SHA-256 values and complete packet-stream demuxes, plus 57/57 decoded start/middle/end frame samples. The literal container tag `com.android.capture.fps=5.000000` is retained without inferring real drawing time. - Bound the 19 media objects at immutable media-only PR revision `0c421b555a2558887b65b0b25380b35c642fa7e1`; the final relationship sidecar pins metadata-bearing revision `429541ae7f93c869056df956873f3441d580914c`. - Retained media-object-level `record_id` identity. This update does not add a global `work_id`, backfill historical rows, create a v3 release, or otherwise bump the dataset version. ## Unreleased — August 24–27 complete fragment-sequence ingestion - Added 22 creator-confirmed, unchanged canonical Framelapse source fragments totaling 17,389,875,055 bytes and 3,391.699091 seconds of encoded playback. - Preserved five complete drawing sequences as ordered `RecNN` sets containing 2, 2, 6, 2, and 10 canonical fragments. - Increased active membership to 385 Standard plus 128 Pre-Standard records, or 513 total, without changing the 29-field schema or `standard-2025-07-13` protocol ID. - Added a separate 27-row, 15-field dated relationship sidecar covering the 22 canonical fragments and five unchanged noncanonical InShot derivatives. - Stored the five derivatives under the nested provenance `InShot/` directory; they remain outside both active metadata indexes, Viewer configurations, and active byte and duration totals. - Retained media-object-level `record_id` identity. This update does not add a global `work_id`, backfill historical rows, create a v3 release, or otherwise bump the dataset version. - Verified the 27 planned media additions, with no deletion or changed existing path, at immutable media-only PR revision `bb4160045cae21a71bf360278549398dc9d2a028`. - Verified all 27 relationship paths, SHA-256 values, byte sizes, Xet identities, canonical metadata joins, and exact root metadata bytes against immutable metadata-bearing PR revision `8a21015210c19936d3b18175ae3f4061ae1e5d8d`. This validation does not assert Viewer settlement or merge status. ## Unreleased — June 2026 recording-sequence filename correction - Corrected the canonical paths of 15 existing 30 × 40 in recordings dated June 18–30, 2026, to preserve the creator-confirmed continuous filename sequence `Rec01` through `Rec15`. - Retained every media payload, stable `record_id`, Xet identity, byte size, technical field, and membership count; this is a filename correction, not a media addition or re-encoding. - Clarified that documented `RecNN` order may continue across finish dates and does not reset merely because the date changes. - The suffixes document recording order only; they do not by themselves assert separate drawings, sessions, or completion states. ## Unreleased — August 22–23 recording-sequence ingestion - Added three creator-verified Standard recordings dated August 22–23, 2026, totaling 51,022,796,325 bytes and 3,895.390997 seconds of encoded playback. - Increased active membership to 363 Standard plus 128 Pre-Standard records without changing the 29-field schema or `standard-2025-07-13` protocol ID. - Retained literal `Rec01` through `Rec03` filename suffixes for the continuing 14 × 17 in sequence; the suffixes do not independently claim an endpoint. - Verified direct visual continuity and drawing progression across all three recordings; integrity validation covers every complete video and audio stream. - Documented the creator-confirmed Rec03 ink change at approximately `00:04:59`: Pelikan black before the visible refill handling and unmarked, unknown-brand black ink afterward. No replacement brand is inferred. - Rec03 produced one end-of-stream non-monotonic DTS warning from FFmpeg's null output during combined-stream validation. Source packet and decoded-frame timestamps at the tail remained strictly increasing, and both streams decoded. ## Unreleased — August 2026 Standard ingestion - Added three creator-verified Standard recordings dated August 17–19, 2026, totaling 50,478,909,746 bytes and 3,856.391997 seconds of encoded playback. - Increased active membership to 360 Standard plus 128 Pre-Standard records without changing the 29-field schema or `standard-2025-07-13` protocol ID. - Fully decoded all 231,383 video frames and every audio stream; no source corruption, dropped frames, duplicated frames, or broken endpoints were found. - Added the first 14 × 17 in Standard record and documented its Strathmore 500 Series Premium Bristol plate surface from the creator-supplied packaging. - Retained `Rec01` on the 14 × 17 in filename because `Rec02` is expected; the suffix distinguishes recording sequence only. - Applied media integrity, identity, rights, and traceable-metadata criteria without filtering by perceived artistic quality. ## Unreleased — July 13 tier promotion - Promoted `tla-3f49a8917fe22676` from Pre-Standard to Standard after the creator confirmed that clean media integrity is the controlling membership criterion. - Verified all 1,205 video frames and the audio stream decode successfully. The shortened clip begins mid-process and reaches a completed, signed endpoint without corruption or a broken ending. - Preserved the 141,221,810-byte media payload, Xet identity, stable record ID, and the 485-record active total while relocating its repository path. - Moved the documented protocol boundary coherently to `standard-2025-07-13` across the canonical index. - Updated active membership to 357 Standard plus 128 Pre-Standard records without changing the 29-field schema. ## Unreleased — Standard ingestion ### Membership and evidence policy - Added 17 creator-verified Standard records totaling 9,289,459,921 bytes and 1,071.199 seconds of encoded playback. - Increased active membership to 356 Standard plus 129 Pre-Standard records. - Applied technical integrity, identity, rights, and traceable-metadata criteria without filtering records by perceived artistic quality or degree of completion. - Preserved the existing 29-field schema and separate `canonical/archive` and `pre_standard/archive` configurations. ### Qualified media conditions - Documented two shortened derivatives as active Standard representatives without describing them as full-duration capture masters. - Recorded pronounced lighting variation for the 2026-01-11 record. - Retained one non-monotonic duplicate video-timestamp warning for each of the 2026-01-11 and 2026-02-08 records; both files decode completely. - Corrected the publication path `2025-09.27...` to canonical date syntax `2025-09-27...` without altering the local source file. ## Post-v0.3 — March/April provenance correction ### Record identity and physical metadata - Promoted historical `2026.03.04(214.0).mp4` from the unresolved provenance subtree as the true `2026-03-04 14:21` BallpointPen / Ink / Paper / 9x12 Standard record. Its 6,587,685,497-byte payload receives a new stable record ID based on its former archival path. - Corrected the distinct 496,296,535-byte object formerly occupying that path to `2026-04-04 16:49` after curator verification against the physical work. Its existing stable record ID remains attached to the same media object. - Increased active membership to 339 Standard plus 129 Pre-Standard records; no video bytes were added to repository storage. - Preserved the immutable `v0.3.0` tag at its original 467-record commit. ### Media payload - Moved both existing Hub objects server-side without transcoding, replacing, downloading in full, or uploading video bytes. - Removed the resolved `Unresolved_Series_9x12/2026.03.04(214.0).mp4` path only after preserving that same object at its corrected canonical path. - No audio stream was modified. ## v0.3.0 — Release candidate ### Metadata contract - Corrected record `tla-a6bd9380e0256aff` from an unknown finish time to the curator-confirmed `2025-09-06 13:46` while retaining the confirmed 12x9 physical dimensions. - Linked historical `2025.11.15(116.1).mp4` to that record after curator confirmation that it depicts the same physical drawing and near-duplicate video content. The files have distinct byte sizes, SHA-256 values, and Xet identities, so the relationship is not described as blob identity. - Added persistent `record_id` values to all 129 active Pre-Standard records. - Added exact Hub Xet file IDs and byte sizes for every Pre-Standard file. - Added playback duration, stored dimensions/orientation, reported frame rate, codec, pixel format, container, and audio-stream presence for every Pre-Standard file. - Normalized Pre-Standard dimension casing and removed duplicate values from semicolon lists without inventing missing historical facts. - Retained the 338 Standard / 129 Pre-Standard media membership unchanged. ### Viewer configuration - Replaced the combined `default/train` surface with separate `canonical/archive` and `pre_standard/archive` configurations. - Made `canonical` the default configuration. - Declared the full feature contract inside each builder configuration so the generated Parquet schemas remain identical even when a source CSV column happens to contain only integer-looking values. - Declared stored dimensions and byte sizes as integers and `has_audio` as a boolean; retained floating-point frame rate for historical encodings. - Continued to exclude the nested provenance subtree through direct-file globs. ### Validation and documentation - Extended validation across both active tiers, including real calendar-date parsing, positive numeric checks, hash format, boolean values, filename/date agreement for canonical files, cross-tier identity uniqueness, and complete technical metadata. - Added direct comparison of all indexed content hashes and byte sizes against a pinned Hub API repository inventory. - Corrected three recently added canonical rows whose `hub_xet_hash` values contained the LFS-compatible SHA-256 instead of the Xet file ID. Their local SHA-256 evidence remains documented in `notes`. - Documented 50.06 total playback hours and full active-index technical coverage. - Preserved unresolved semantic issues as explicit warnings: the canonical 30×40 Paper/IllustrationBoard conflict, five unknown-date midnight values, and two incomplete Pre-Standard material descriptions. - Added `tools/enrich_prestandard.py` so the technical enrichment can be repeated against a pinned Hub inventory and revision without intentionally transferring full video payloads. - Added `tools/capture_hub_inventory.py` to capture the Hub Xet file IDs used by `hub_xet_hash`; these are distinct from the LFS-compatible SHA-256 values exposed by the ordinary repository inventory. ### Media payload - Renamed the canonical object server-side to `2025-09-06.1346.BallpointPen.Ink.Paper.12x9.mp4` without changing its video bytes. - Removed the now-resolved near-duplicate `2025.11.15(116.1).mp4` from the unresolved provenance payload; its path, relationship, differing object identities, and removal remain documented in metadata and repository history. - No video was downloaded in full, transcoded, replaced, or newly uploaded. - No audio stream was modified. ## v0.2.0 — Release candidate Base repository revision: `339e6149de130cbe2d001f7a0e5247872bc788ec` ### Metadata contract - Added persistent `record_id` values for all 338 canonical records. - Added `collection_tier`, `is_canonical`, and `acquisition_protocol_version`. - Normalized canonical multi-value `tool` and `medium` fields from pipes to semicolons. - Defined tool and medium fields as independent ordered unique lists rather than unsupported positional pairs. - Preserved every canonical file path, content hash, technical property, and provenance value from the pinned source index. - Removed one pre-standard metadata row whose corresponding media is not distributed because the required ownership and provenance threshold was not met. No identifying local path or content hash is published for the excluded holding. - Added complete canonical rows for three previously unindexed Standard files. Their local inbox SHA-256 values exactly match the current Hub LFS identities; stream metadata was extracted from those matching bytes. ### Documentation - Distinguished archive scope from currently published video coverage. - Documented that the default Hub `train` split is not an official experiment split. - Distinguished capture orientation from stored encoded orientation. - Added intended-use, unsupported-use, bias, privacy, incidental-rights, audio, access, citation, and reproducibility guidance. - Documented missing temporal and physical calibration required for rigorous motor-control measurement. ### Validation - Added `tools/repair_metadata.py` for the reproducible v0.1-to-v0.2 transform. - Added `tools/validate_metadata.py` for identity, schema, technical-field, orientation, duplicate, and Viewer-path checks. - Retained the 465-row pinned Dataset Viewer snapshot as a historical base audit input. Final revised membership is validated against the PR repository tree because the Viewer has not yet rebuilt from the amended indexes. - Reconciled the repository tree with both active indexes: 338 indexed canonical records and 129 published pre-standard records, both with one-to-one row/file coverage. - Added `tools/metadata_additions_v0.2.csv` so the 338-row index regenerates deterministically from the pinned 335-row source plus 3 audited additions. ### Media payload - No video was downloaded, moved, deleted, renamed, transcoded, or replaced. - No audio stream was modified. - No historical relationship was inferred or promoted. ### Deferred - Formal Hub configurations separating `canonical` and `pre_standard` remain pending a Viewer-build test because canonical media and provenance holdings share a directory subtree. - Technical ingestion of 129 pre-standard records remains incomplete. - Public ingestion of the earlier 2012–2016 archive—primarily uploaded 1080p drawing videos, with some livestream-derived material—remains pending. - No lightweight proxy-video release is included in this metadata-only change.