Click the photo to zoom · Esc to close
The data
What a licensed case actually contains – shown, not described: the clip with its annotation drawn on the picture, the schema behind it, and the tree that lands in your bucket.
Proof of method · sample case
Ninety seconds of implant placement with every layer drawn on the picture at once: the phase, the instrument list of the keyframe on screen, the action triplets, the anatomy in view, the annotator's reasoning, the clinical context, and the surgeon's own words underneath. Nothing here is a mock-up – each line is the stored value for the keyframe you are looking at.
A single case, played with its annotation rendered over the picture. The header carries the phase, the black strip the instrument list of the keyframe on screen, and the panel underneath the remaining layers – down to the surgeon's own sentence at that second. Where a layer holds nothing for a frame, the clip prints a dash instead of filler.
The bar under the clip is the case's phase timeline, width proportional to annotated keyframes – hover any segment. Where a layer is empty for a frame the clip shows a dash rather than filler: coverage is reported, not padded.
Annotation schema v2
From verbatim surgeon speech to clinical context – each layer answers a different question, and every label records where the knowledge came from: seen in the image, stated by the surgeon, or both. Card figures are from the released sample case; across the line so far – with the next cases always in annotation – the layers count 4,574 quotes, 14,187 instrument labels and 5,859 action triplets over 3,392 keyframes.
Plus three auxiliary fields: anatomy_visible – 21-term closed anatomy
vocabulary, 1,446 labeled spans, including safety-critical structures (mental nerve, mental
foramen); no_surgical_field – 60 frames flagged and reported separately, never
hidden in the coverage number; uncertainty – what the annotator could not resolve,
quoted verbatim instead of silently corrected.
What you receive
Recording runs year-round and cases come off the line continuously. Each one lands in the same tree – the picture, the sound, every annotated frame as a still, six layers of annotation, the blind control pass, the transcript, the reading guide and the compliance file, with a checksum for all of it. What changes between cases is length and procedure. The structure does not.
Every camera angle of the case in one H.264 file with a burnt-in timer that all timestamps in the package refer to. No audio track inside the video – the sound ships separately so picture and speech can be processed independently.
The intraoperative narration, aligned to the master's timer, lossless and compressed. This is the track every L0 quote is traceable back to, which is why it travels uncompressed as well.
Full-resolution JPEGs in per-camera folders, plus a keyframe index and a filename map, so a frame id resolves to a file and a timestamp without guesswork. Longer cases simply carry more.
annotations_merged.json carries L0–L5 for every keyframe, and each label records its confidence and whether it came from the image, the surgeon, or both. Batch files are kept beside the merge: a batch records what the annotator was told at the time.
The same frames annotated a second time without the audio, plus the comparison report. Every quality number we publish is computed from these files, so you can recompute them instead of taking our word for it.
De-identified narration on the master's clock. Clinical terms are corrected against the operating surgeon's own review, never accepted from speech recognition as-is.
Closed vocabularies for instruments and anatomy, the phase timeline, the dated correction log, and the dataset card with the timebase note. Including the list of things this case cannot support – stated up front rather than discovered later.
The per-case compliance summary: IRB protocol, the consent basis, the de-identification method and the frame-by-frame audit behind it. It ships inside the package, not on request.