Processing pipeline | Docs

Every session goes through the same stages. Each one is keyed by its inputs, so retries and reprocessing are safe and an unchanged input is never processed twice.

Stage	What happens	What you see
Verify	Whole-file checksum checked; original stored unchanged	`status: uploaded`
Validate	Schema, channel rates, missing data, calibration lookup	`issues` with `severity: error` stop here
Protect	Faces and other personal details blurred in image channels	Nothing; originals stay in storage
Align	Clock offsets estimated and applied; streams placed on one timeline	`clock_offset_estimated` issues
Register	Poses registered into the site's world frame	`registration_residual_m` in session metadata
Reconstruct	Gaussian splats, meshes and collision geometry; moving objects separated	Scene and object assets
Label	Entities segmented and tracked; interactions detected	Entities and events
Segment	Recording cut at interactions, with margins	Segments become searchable
Check	Lossless encoding verified on read-back; quality scored on held-out frames	`quality` on each segment

Lossless storage

Reconstructions are stored with a lossless codec and verified after encoding: an asset is published only if decoding it reproduces exactly the data that was encoded. What you retrieve is therefore exactly what was processed.

Quality scores

Score	Meaning
`psnr`	Peak signal-to-noise ratio of renders against held-out camera frames, in dB
`coverage`	Fraction of the segment's extent seen by at least three frames
`registration_residual_m`	Mean pose error after registration to the site frame

Use quality_min when searching to keep low-quality segments out of training.

Reprocessing

When the pipeline improves, sessions can be reprocessed. New segment versions are created next to the old ones; slices keep pointing at the versions they were frozen with. Each segment records the pipeline_version that produced it.