FACTORY PDIGGER / PROJECT NOTES
From capture to event
The original five excerpts were captured on a Jetson Orin NX on 13 November 2026 using the installed, identified interpreter. The replacement runtime was not release-qualified at capture time. No production deployment was changed.
One source clock
Removal, two-hand handling and tray replacement retain five seconds of lead-in. The replacement pushing episode uses dry1 01:27.5–01:37.5 from continuous Mac inference beginning at 00:00. The one-hand placement episode was reprocessed on a Mac continuously from the beginning of dry1, through 02:29.4, including five seconds of look-ahead beyond the displayed 02:10–02:24 interval. The overlapping model outputs match the earlier Mac run, so the previously reviewed scene fits are retained. The cut ends after the last placement, before the tray starts moving at 02:24.2. Original presentation timestamps and source frame indices are retained in the episode manifest. The camera video and interactive renderer use the same source timeline, starting at 1×. The camera videos contain 20 frames per second; approximately five model observations per second are interpolated between captured frames. The dry recordings have approximately ten source frames per second.
Measured processing
For the original excerpts, median and p95 are computed from the captured Jetson per-frame total for the displayed interval, excluding lead-in. The replacement pushing excerpt has no NX measurement for its new source window: its figures are explicitly labelled NX reference measurements from dry1 02:10–02:24. They are not timings of the new push or the Mac run. The placement episode retains those established NX measurements unchanged; they do not measure its later Mac history run. The total covers decode, encoder, regions, hands, stones and temporal processing. Stage medians are shown separately. Scene fitting, rendering, encoding, transfer and browser playback are excluded. These numbers are latency per sampled model frame, not a sustained throughput or end-to-end latency result.
The installed scene logic was replayed offline on a Mac against the captured packets and image evidence. The public browser uses the same identified renderer and fitted scene snapshots. For placement, an explicit reviewed occupancy layer keeps verified empty areas empty until a hand placement is visible; detector IDs, native slot events and elapsed time cannot refill them. A false carry associated with a stone that remains in its source location is rejected. Four visible placements are reviewed separately. The first three settled shapes come from later unobscured native fits. The broad final footprint is manually traced because the native tracks fragment or misplace it. Each appears only after visible hand placement. The reviewed carry includes two separate solids at each glove; their relative dimensions and hand-local display identifiers are manually reconstructed. Hidden handoffs and off-screen physical identity remain unresolved. These are corrections for the recorded demonstration, not a new live-model claim. The original native events and geometry remain in the evidence, alongside the occupancy review record. The removal review documents those source corrections. In the pushing review, five top-row stones visibly slide left together while the tray and lower rows stay fixed. Native top-row fits remain frozen, so the displayed movement uses reviewed image anchors and checked intermediate translations. It is a reconstruction of recorded motion, not a prediction of contact forces. The preceding broad placement and persistent lower-row gaps are reviewed separately. A paired video remains available as a fallback when interactive rendering is unavailable. Full file hashes, engine identifiers and capture dates are available in the capture identity record.
Interactive exploration
Scrubbing, frame stepping and event selection move the camera and scene together. Rotate or pan the scene by dragging; use the view buttons or zoom controls to explore it. Arrow keys rotate, Shift plus arrows pan, plus/minus zoom, and Home resets when the scene is focused. On touch screens, use Rotate/Move or two fingers to pan and zoom.
Select a stone in the scene, camera or Stone menu to highlight the same recorded track ID in both views. Camera zoom and pan also move the outlines. A track ID is a recorded model association, not a guaranteed physical identity or grid address. Manually reconstructed objects have negative display IDs and explicit “Reviewed” labels; these are not native model track IDs. Missing tracks are marked as not visible; selection clears when crossing a confirmed tray replacement. The two views do not imply that every model association is correct.
Native observations and interpretations
Each event exposes unchanged supporting scene events, source timestamps, relevant stone geometry and hand state. Plain-language summaries are derived interpretations, reviewed against the footage. An event such as lift, drop, lost or slot-t-release does not by itself confirm removal or a destination. The removal example contains four visible removals. Three additional slots were already empty at the beginning. Reviewed vacancy guards reject later native ID reuse and false refills. Carries are reconstructed only in visibly supported spans; hidden paths, exact obscured departure times and staging destinations remain unresolved. Native hand-pose errors remain in the source evidence. The placement example retains unresolved carried identities because a native track returns to a pre-existing slot. The four destination columns are verified separately from the footage. Its visible lower-edge placement is recorded separately from that rejected association.
Tray addresses
Known locations use a tray ID, one-based row and one-based column. Addresses stay with the tray as it moves; a stone retains its source address while carried. A new tray ID is assigned only after a visibly confirmed replacement. IDs are scoped to the excerpts, not a global factory inventory.
Each grid file defines its orientation and settled reference polygons. The interlocking-stone trays use six alternating rows of six and five slots, with a half-slot stagger. Dry-side rows use explicit mixed-width slots. Missing correspondences and uncertain destinations remain null, and confirmed empty slots are listed separately. A partial native grid is evidence, not a complete physical inventory; unreviewed addresses are not asserted as ground truth.
Privacy & limitations
Faces are blurred in camera frames before annotation and composition. Episode posters, previews, videos and the lower film poster derive from those redacted frames; the upper film thumbnail is the supplied illustration. The scene has no camera texture or face detail and uses generic untextured head geometry derived only from body joints. Public scene files exclude facial landmarks, appearance masks and image payloads. Geometry uses a template-based relative scale rather than calibrated units.