embodied-olympics/release1-media

Embodied Olympics: match media pack

Match data for the team's video, rendering and live work. Results are provisional and unpublished: do not share. Access is granted by the project.

How to pull

Teammates: once the project has given you access, run pip install -U huggingface_hub and hf auth login once, then:

hf download embodied-olympics/release1-media --repo-type dataset --include "manifest.*" --include "README.md" --include "tools/*" --include "matches/tennis_g1/*" --include "assets/tennis_g1/*" --local-dir eo-media

Change tennis_g1 to the sport you want (both places; matches/* and assets/* take everything). To pick by entrant, board or media, use the tool that comes with it:

python eo-media/tools/pull.py --list --sport tennis_g1
python eo-media/tools/pull.py --sport tennis_g1 --entrant ENTRANT_ID --only video

manifest.csv / manifest.json list every match; tools/draw_frame.py draws any frame with any camera from a match's state log (each match folder's README has the command). The pack is updated as matches finish.

What a match folder holds

matches/<sport>/<board>/<fixture>/: match.json (entrants, score, status, media facts), final.json (official record), replay.mp4 and replay.fixedcam.mp4 where a video exists, spectator_states.pkl where a state log exists, README.md, SHA256SUMS. assets/<sport>/ holds the textures and meshes a state log's scene names by path; draw_frame.py finds them there (pull them with the match: pull.py does).

Only cut what the manifest lists under matches. Rows under excluded (void or replaced) have no folder and must not be used, and a status of provisional means the result may still change.

Analysis layer (for the team's analysis of the official run)

Status: provisional until the official run ends. Top level: INDEX.md (every sport, board and fixture: entrants, result, validity, what replaced it, date, folder), index.csv and index.json (the same rows for scripts), and sports/<sport>/retests/ (growth_raw.json: the re-test results per re-tested program snapshot; capability.json: the capability curve data). Each valid match folder has analysis/:

file	what	how to read it
analysis/analysis.md, analysis.json	the platform's post-match analysis: score course, program version timeline, switches, findings	read the .md first; the .json has the same per seat, plus heuristics
analysis/verdict.json	the judge's verdict: validity, checks, faults, ranking, termination	valid, termination, faults
analysis/evidence.json	the sealed evidence list (digests of the files the verdict rests on)	match a file to its digest
analysis/seats/<seat>/seat.json	entrant, model, harness, transcript files, every deployed program version with its fingerprint (snapshot, controller / program sha, deploy rows with control steps, content digest, tar sha256), notes, files held back by the scan	start here for a seat
analysis/seats/<seat>/transcript/*.zst	the agent's own transcript files as its harness wrote them (session streams / logs; encrypted reasoning blobs removed)	zstd -d FILE; mostly JSON lines; seat.json names the harness
analysis/seats/<seat>/think_calls.jsonl.zst, think_usage.json	every model call through the metering proxy: time, tokens, tool calls, a short excerpt; and the totals	tokens over time, calls per phase
analysis/seats/<seat>/programs/vNN-snapNNN-<sha>.tar.zst	one tar per deployed program version: program and config files as in the seat (app/output/...), the weights the program names	tar --zstd -xf FILE; diff two versions to see what changed
analysis/seats/<seat>/notes/	the seat's own markdown notes from its output folder	as written by the agent
analysis/SHA256SUMS, analysis/README.md	checksums; what this match's files are	cd analysis && sha256sum -c SHA256SUMS

Pull the analysis layer with python eo-media/tools/pull.py --sport tennis_g1 --analysis (--analysis-light leaves out transcripts and program tars). Seeds, credentials, platform files, whole workspaces and practice logs are never in this repo; every file passed a secret and seed scan before upload, and a file that did not is left out and named in its seat.json.