docs/DETAILS.md

Release details — RoboDojo seed 0

1. Selection

The release is the complete seed 0 that was scored, chosen exactly as the merge that produced the official summary chose it. Each evaluation round ran one layout for every task configuration that uses it — all 54 configurations (the 42 tasks, the 12 Generalization tasks twice, standard and _random) on layouts 0–24, the 30 others on layouts 25–49, and the refill layouts where needed — and a round's job list records every attempt. Per round and configuration the merge keeps the last attempt of each job and, among them, ranks a scored episode (success / fail) above anything else. The episode's official verdict is the evaluator's own _result.json entry for that layout.

That gives 2,116 chosen episodes: 2,100 scored ones — the table — and 16 the evaluator marked unstable — unstable/. Per task: 50 rows; per Generalization task 25 standard + 25 randomized; layouts 0–55, of which 15 are refill layouts (imitate_sorting_sequence 6, deposit_coin 5, fill_egg_holder 3, general_pickup 1).

Not released (90 rows in results/seed0/EXCLUDE_s0.csv): development round (spare layouts) 56, infrastructure retry 33, duplicate 1. An infrastructure retry is an attempt that ended in an infrastructure failure — a provider stream error, a simulator fault, out of memory at start-up — or did not finish, and was run again; the re-run is the released episode. The development rounds ran spare layouts before the evaluation and are not part of the seed.

2. What a row is built from

All from the session folder of the episode and the evaluator's result folder; nothing else is read.

column	source
outcome, score	the evaluator's _result.json entry for the layout (identical to the job record on every row)
official	episode_metrics.json: official_take_action_cnt, action_budget, ended_by
total_wall_s, timing	episode.json started_wall; episode_metrics.json wall_duration_s (total_wall_s) and control_time_s (sim_clock_s); tool_timing.jsonl (sim_s, the sum of its duration_s; agent_s is the rest). The job records' own wall_s equals wall_duration_s except on the 54 episodes of layout 0, where it also counts the simulator's boot; wall_duration_s is used on every row
tokens	the last token_count event of the codex session log, total_token_usage; from that log only these four totals and the images of §4 are used
model, effort, backend	the frozen model settings in episode.json
agent_messages, traces, tools	agent_events.jsonl (the codex event stream) and preload/ (the launch observation)
instruction, prompt_text	goal/INSTRUCTION.md, PROMPT.md
frame_first, frame_last, video_preview, videos	the evaluator's recordings

traces[]: step 0 from preload/launch_message.md and the three preload images; then, in file order, one element per completed agent_message, per mcp_tool_call (at its completion; a call that never completed — the episode ended while it ran — at its start, with ok NULL), per error item and per stream error / turn.failed event. 69,986 tool calls, 1,189 of them still running when their episode ended.

3. Sanitization

• Absolute home-directory paths are rewritten to <HOME>/… and the codex home to <CODEX_HOME>/…, in every text of the table, the event streams and the documents.
• Dropped: the codex thread_id (2,106 values, one per event stream). No account, lease, slot or session-hash field reaches any published file: the job records, episode_metrics.json's account block and the codex session log are not published, and only the four token totals and image parts are taken from the latter.
• The scan: every text byte that is uploaded — all string values of the three tables (nested ones included), the event streams, the documents, the CSV manifests, the views' JSON members, the card and these docs — is searched for absolute home-directory paths, local user, host and directory names, e-mail addresses, JWT / API-key / Hub-token shapes, the OpenAI creator user and account id fields, account and session hashes, and OpenAI user ids. Hits: 0 over 2,433 text files and the three tables.

4. Encoding

• Recordings: the evaluator's files.
• frame_first / frame_last: the first and last decoded frame of the head recording, JPEG quality 85, 640×480.
• video_preview: ffmpeg -vf scale=480:-2 -fps_mode passthrough -c:v libx264 -preset medium -crf 30 -pix_fmt yuv420p -an -movflags +faststart -map_metadata -1 (ffmpeg 7.0.2): 480×360, every frame of the source kept at 25 fps — a frame is an official action. Frame counts equal the source's on all 2,100 rows.
• Views: the PNG images the model was sent, decoded from base64. Images that the event log truncated were taken from the record of the model input.
• Tool interface: the packaged policy (XPolicyLab/policy/AGP/agp_runtime, the RoboDojo submission) imported read-only with mcp 1.30.0 and pydantic 2.13.5 (its pinned versions), its MCP server built with no robot attached and the switches the episodes recorded, and list_tools() written out. It is the interface of the episodes of layouts 4–55; layouts 0–3 ran an earlier build of the policy, whose prompts — like every episode's — are in documents/prompts/.
• Parquet: pyarrow, explicit schema with the huggingface features block, zstd, one row per row group, statistics on the scalar columns only.

5. Checks

check	result
2,100 rows with unique episode_id, 16 unstable rows; RELEASE lists all 2,116; every job record of every round (2,186) and every session folder of the released runs (2,135) is released or listed in EXCLUDE, never both	pass
the official summarize script, run on a result tree rebuilt from the table (one _result.json per configuration, entries as the merge writes them): Avg SR 56.13, Avg Score 63.63 and the five dimensions; the tasks table equals the 42 per-task cells and the 24 Generalization half cells	pass
any 100 consecutive rows of episodes/ under 300,000,000 bytes uncompressed (largest window 57,794,807, largest row 1,754,449)	pass
every videos[].url and trace_url resolves to a file; the 108,987 view keys of the tables, of the event streams, of views/metadata.csv and of the tars are one set, each key in exactly one tar	pass
the scan of §3: zero hits	pass
no folder holds more than 10,000 files	pass
total_wall_s equals timing.total_wall_s, agent_s + sim_s = total_wall_s and agent_s ≥ 0 on every row, and total_wall_s, sim_s and sim_clock_s match the session files (2,100 / 2,100 / 2,100 / 2,100 rows carry total_wall_s / agent_s / sim_s / sim_clock_s); the tasks medians and totals match the rows	pass
load_dataset reads both configs from the card's YAML and with the parquet builder (episodes 2,100, tasks 42, unstable 16) and decodes an Image and a Video cell	pass

6. Size

folder	files	size
(root)	2	18 KiB
docs	2	22 KiB
documents	318	11.1 MiB
episodes	105	944.8 MiB
results	3	237 KiB
tasks	1	108 KiB
traces	2,106	17.8 MiB
unstable	1	327 KiB
videos	6,301	10.51 GiB
views	41	36.65 GiB
total	8,880	48.12 GiB