How the platform works
Specification, contract, collection, validation, delivery, evaluation and the next collection. Each stage leaves an artifact you can check.
From brief to next batch
When a policy is evaluated, its failures become the next specification, and targeted collection is measured against an equal-budget control. A negative result is reported, not hidden.
Before collection
- Task specificationSkill, embodiment, streams
- Acceptance termsReview window, codes
- Rights and site approvalIn writing
During collection
- Pilot sampleReviewed by you
- Supervised captureEvery attempt logged
- ValidationAccepted or coded
After delivery
- Versioned datasetManifest, checksums
- EvaluationFailure profile
- Next batchAgainst a control
You hold accepted data, and what to collect next.
Run around accepted data
The operator. Paid, trained on the versioned SOP and identified by pseudonym. Pay never silently depends on a buyer's rejections.
The supervisor. Named for each site. Owns resets, safe stops, intervention logging and the attempt log, failures included.
The reviewer. Independent of collection. Applies the contract's acceptance criteria and its reason codes.
- One validated cell before any replication.
- Active manipulation and reset time reported separately.
- A paid trial and a defect taxonomy before volume.
- Uploads and checksums verified on site.
- Existing operations preferred over new facilities.
- Partner production lines left untouched.
One capture contract per task
| Field | What it records |
|---|---|
| Episode ID | Unique and immutable |
| Task and version | The SOP version the attempt was run against |
| Site and operator | Operator by pseudonym, with the permission record |
| Modality | Footage, staged demonstration, teleoperation or engineering data |
| Embodiment | Robot and end-effector identity, calibration hash |
| Observations, commands, state | Synchronized, with per-stream timestamps |
| Objects and fixtures | Which parts and fixtures were in the cell |
| Start, end and reset times | Active manipulation and reset reported separately |
| Interventions | Takeovers, safety stops, manual repositioning |
| Failures and recoveries | Labelled segments within the attempt |
| Outcome | From a check the model under test cannot influence |
| Checksums | For every stream file |
| Review result | Accepted, or rejected with its code |
| Split | Train, diagnostic or holdout |
Accepted, or rejected with a reason
- Stream synchronization R01
- All required streams present and aligned within the contracted tolerance.
- Calibration lineage R02
- Calibration and firmware hashes match the session record.
- Outcome verifiability R03
- The terminal outcome is confirmed by an independent check, not inferred from video.
- Intervention disclosure R04
- Every human assist is flagged; none is edited out.
- Rights lineage R05
- Site, worker and customer permissions cover this unit and its intended use.
- Task conformance R06
- The attempt followed the SOP version, objects and fixture it claims.
Rights are part of the data
Site authorization. Written approval of the specific workflow, camera views and data fields. A membership or an introduction is not permission.
Worker participation. Informed consent and explicit terms; operators are pseudonymized in delivered data.
Customer IP excluded. Owned parts and open designs by default. No customer drawings or production records.
Scoped licences. Training and evaluation use, exclusivity, retention and territory, written per contract.
Pay for accepted data
<dataset>/<version>/
manifest.json
schema/capture_contract@1.json
splits/{train,diagnostic,holdout}
episodes/<episode_id>/...
rejected/<episode_id>.reason
sop/<task>@<version>/
checksums.sha256
acceptance_reportFormat. Agreed at specification to match your training stack; the schema is versioned with the data.
Versions. Immutable and checksummed. A refresh batch is a new version, never a silent edit.
Splits. Diagnostic and final holdout reserved up front; the holdout never chooses the next batch.
Acceptance. A report per batch: accepted units, rejections by code, active and reset time.
Evaluation decides the next batch
- Task success
- All SOP criteria met within the declared attempt and time limits, without disallowed help.
- Recovery
- Success after a permitted autonomous correction, reported apart from first-attempt success.
- Intervention
- Operator takeover, safety stop, manual repositioning or unplanned reset. Never edited out.
- Speed
- Wall-clock execution and reset time, not sped-up video. Failed trials stay in the denominator.
- Economics
- Accepted episodes per paid operator-hour, review minutes and rejection causes.
Embodiment-agnostic, purchase-gated
Commercial gate
The buyer accepts modality, country and embodiment before hardware is committed.
Mechanical gate
Real sample parts fit payload, jaw width, reach and joint limits; a two-arm envelope is collision-free.
Software gate
A supported record and replay path; state, action, time and camera calibration captured; firmware pinned.
Delivery gate
A written equipment list and arrival commitment, with spares on hand.
Safety gate
A site risk review, restrained fixtures, low speed, a trained supervisor and an accessible safe stop.
