metadata/README.md

Recording metadata

This release provides recorded controls, cameras, observed world state, static map geometry, and frame correspondence data for studying paired videos and rendering alternative proxy styles. It covers 21 recording sessions. The training subset contains 5,000 selected paired clips; metadata retains the wider sessions and temporal context.

Files and extraction

Path	Contents
episodes/episode_###.tar.gz	Controls, cameras, world observations, timing, available Parquet tables, and proxy frame/scene metadata
maps/scene_001.tar.gz	Static placements, spatial databases, collision meshes, materials, and water
source_index/global_index.sqlite3	Portable frame, segment, source, and cross-view correspondence index
renderer/proxy_source_5_5_1.zip	Minimal renderer source package
renderer/environment/requirements.lock	Recorded Python dependency versions
integrity/*.members.jsonl.gz	SHA-256 and byte size of every archive member
catalog.json	Artifact sizes, hashes, and archive counts
schema.json	Main fields and timing conventions

Extract recording and map archives under source/:

mkdir -p source
tar -xzf metadata/episodes/episode_001.tar.gz -C source
tar -xzf metadata/maps/scene_001.tar.gz -C source

This creates source/recordings/episode_001/ and source/worlds/scene_001/. Database paths are relative to source/. Paths inside recording JSON are relative to the release root. Original long-form RGB/proxy videos and audio are omitted; their source identities and timing remain in the index, while selected videos are in the training archives.

Some preserved recording manifests still name historical validation reports or derived directory indices. Those diagnostic files are excluded from this public release. Use the released SQLite index, member integrity manifests, and JSONL observations for inspection and replay. Source video paths in the index are provenance references and are not expected to resolve to long-form media in this release.

Controls, cameras, and world state

recordings/<episode>/.spool/agents/<agent>/client.jsonl is the authoritative client stream. Each row contains episode, agent, sample, frame, and clock identifiers, plus:

• action: movement/look axes; jump, sprint, crouch, interaction, vehicle entry/exit, primary/secondary controls, and available keyboard/mouse input. Button states distinguish pressed, released, and held.
• camera: rendered camera position, rotation, field of view, projection rays, clip planes, resolution, and frame counter.
• agent_state, subject_observation, and scene_observation: transforms, motion, appearance, coarse body observations, and nearby entities with availability/completeness flags.

recordings/<episode>/.spool/world_state.jsonl contains server observations: sample sequence, monotonic time, environment, agents, vehicles, pedestrians, objects, task state, relations, and lifecycle events. Preserve nulls and availability flags. Missing data does not imply zero or an absent entity.

Where available, agents/<agent>/*.parquet and server/*.parquet provide fixed identifier columns and a JSON payload_json column. Parquet is not present for every session; authoritative JSONL streams cover all 21.

Timing and joins

Nanosecond fields use integer nanoseconds; _ms fields use milliseconds. Capture and simulation monotonic times remain available for joins. In-game environment time is retained. Collection wall-clock fields are omitted.

Processed clips preserve 124 frames at 24 FPS, nominally 5.1667 seconds. Per-clip alignments connect left/right output frames to numeric frame_id and source segment. In SQLite:

• frames.frame_id identifies an indexed source frame.
• source_id plus rgb_decode_index identifies an original source-video frame.
• proxy_frame_index selects the matching RGB/proxy frame within a segment.
• world_capture_tick and client sample identifiers connect frames to observations.
• cross_view records the nearest usable other-view frame in the same session and clock domain, with signed difference and a 75 ms threshold.

Cross-view correspondence is a clock estimate, not independent visual synchronization ground truth. Equal frame numbers do not imply matched views. status=estimated_alignment marks usable indexed frames; other statuses retain missing-data evidence.

Map and proxy geometry

The map includes placements, static spatial SQLite data, heightmap, collision databases/indices, compressed triangle meshes, materials, and water geometry. GTA asset names and relative game-internal archive paths are retained as resource identifiers. Mesh records contain nine little-endian float32 values, one uint16 material, and one uint16 flags field, totaling 40 bytes.

GTA V and its game assets are third-party material. Their presence and resource names provide attribution and scientific context; this metadata documentation does not grant or invent a license for those assets.

Per-segment scene_*.json files record static scene selection and geometry policy. frames.jsonl, source_frames.jsonl, source_alignment.jsonl, compiler_manifest.json, and appearance records retain rendering/alignment details. Underlying world/version hashes are scientific identifiers; use the release member integrity manifests to verify rewritten public files.

Public identifiers and exclusions

Sessions use episode_001 through episode_021; the map uses scene_001. Server clock domains use consistent clock_### identifiers shared with training alignments. Other client/clock identifiers use stable pseudonyms. Numeric simulation, entity, frame, segment, and sample identifiers remain available for joins.

Public metadata removes collection wall-clock fields, workstation paths, machine/client UUID strings, authentication values, historical recovery manifests, incomplete sidecars, runtime logs, and operational documentation. This curated representation is not a byte-identical recording-directory export. The private identity map is not distributed. Archives are checked for known private identifiers and field patterns used by this pipeline; this is not a claim that all potentially identifying visual content has been removed.

Scope and limitations

The observations and renderer support alternative proxy colors, geometry conventions, shading, and camera visualizations. Arbitrary new player actions require a simulator or learned dynamics model; recorded trajectories do not establish counterfactual world evolution.

The map is an exported installed scene and does not recover every recording's dynamic interior/map-layer state. World completeness depends on available observations. Human proxies primarily encode root motion/coarse orientation; missing articulation is not reconstructed. Source limitations include some log-pile geometry, roof/foot occlusion, barriers, and vegetation detail. Vehicle occupant smoothing changes display transforms while retaining observations.

See docs/RENDERING.md and the release examples for frame inspection and reconstruction.