Results
Each job produces a summary and one manifest per episode. SE3 Labs shares these with you when the job finishes. Every number in the summary is recomputable from the manifests and the station's recordings.
Episode outcomes
The operator scores every episode that ran to a normal end:
| Outcome | Meaning |
|---|---|
success | The operator judged the task completed |
failure | The operator judged it not completed |
indeterminate | The operator could not decide |
invalidated | The episode did not count; see the reason below |
An invalidated episode is not scored. By default the station re-runs it, up to a retry limit, so it does not eat into the requested episode count.
Invalidation reasons
| Reason | What happened |
|---|---|
MISS_BUDGET | The number of consecutive steps without an accepted chunk exceeded the station's budget. A deadline miss, a structurally invalid chunk, and a chunk rejected by the safety checks each count as one missed step |
STREAM_LOST | The client stream was lost during the episode |
HOLD_EXPIRED | No chunk was accepted for longer than the hold grace period; the arms returned to the home pose |
TORQUE_CEILING | Measured joint or gripper effort exceeded the configured ceiling for the configured number of control cycles |
OPERATOR_PAUSE | The operator paused the station during the episode. The episode is re-run after the operator resumes |
OPERATOR_ABORT | The operator aborted the job |
LEASE_LOST | The station daemon revoked the runner's lease |
MISS_BUDGET, STREAM_LOST, and HOLD_EXPIRED are the reasons a policy
can influence. There is no separate reason for a collision or other safety
rejection: a rejected chunk contributes to MISS_BUDGET, and the rule that
rejected it is counted in the episode manifest's rejection_rules. See
timing and the protocol rules.
Job summary
| Field | Meaning |
|---|---|
job_id | The job |
state, terminal_code, terminal_message | SUCCEEDED, FAILED, or CANCELLED, with why |
task | id, version, instruction, parameters |
model | The identity you declared in your manifest, marked declared_by: customer |
episodes | Count by outcome |
episodes_requested | How many were asked for |
success_rate | Successes over scored episodes |
success_rate_wilson95 | Wilson score interval at 95% |
invalidated_by_reason | Count per reason |
policy_rtt_ms | Your round-trip statistics across all steps: n, mean, p50, p95, max |
steps, deadline_miss_rate | Total steps and misses over steps |
chunks_rejected, actions_clamped | Safety totals |
episode_duration_s, reset_duration_s | Statistics across episodes |
preflight | Round-trip statistics and invalid-chunk count from preflight |
git_sha | Station software revision |
Job terminal codes:
| Code | Meaning |
|---|---|
COMPLETED | Every requested episode was run |
PREFLIGHT_FAILED | Preflight validation or latency measurement failed |
PROFILE_UNSUPPORTED | The manifest requested something the station does not offer, or the protocol versions differ |
STREAM_LOST | No client connected within one hour of job creation, or no client reconnected after a mid-job disconnect |
OPERATOR_ABORT | The operator aborted the job |
LEASE_LOST, DAEMON_UNAVAILABLE, INTERNAL_ERROR | Station-side failures |
Episode manifest
One per episode, including retries. A retry of episode index 3 is
ep_0003_r1.
| Field | Meaning |
|---|---|
episode_id, index, attempt, job_id | Identity |
task, model | The task definition and the customer-declared model identity |
outcome, reason, note | Score, invalidation reason if any, and the operator's free-text note |
end_condition | How a scored episode ended: policy (your EndEpisode), operator, max_steps, or timeout. Empty for an invalidated episode |
end_reason | The reason text from your EndEpisode, when end_condition is policy |
steps | Observations sent to you |
valid_chunks | Chunks accepted for execution |
misses, late, invalid | Deadline misses, late arrivals, and chunks that failed validation |
rejected, clamped | Safety rejections (whole chunk dropped) and clamps (action kept, limited) |
rejection_rules | Count per safety rule that fired |
rtt_ms | Round-trip statistics for this episode |
started_unix, ended_unix, duration_s | Episode start and end times (Unix seconds) and wall-clock duration |
reset_duration_s | Time the operator took to reset the scene before the episode |
walk_settled | Whether every arm settled at the start pose before the episode began |
recording | Path of the episode recording on the station |
Recordings
The station records every episode in full, including every camera at full resolution, the targets it executed, the requested and approved actions of each chunk, and every safety verdict. Recordings stay with SE3 Labs; ask your contact if you need one for a specific episode.