01
Who made it
Every datapoint carries the contributor who created it, with identity-verified consent attached.
On the recordCUSTOMERS · STORIES YOU CAN RUN
Stories you can run, not logos you can’t verify.
Real programs, anonymized by field. Each one shows the work that came in, the record it left behind, and the gate it set for the next release. Bring yours and run the same loop.
BY FIELD
Your vertical
Post-training, model risk, manufacturing, clinical, robotics, compliance.
THE RECORD
Who, who, what
Who made it, who reviewed it, and under what rights.
THE WEIGHTS
Yours
The layer you run, not a marketplace that locks you in.
WHAT EVERY RECORD CARRIES
Who made it, who reviewed it, under what rights.
The same three facts the market now demands of training data. They travel with every datapoint and every decision — so the record survives an audit and the EU AI Act provenance clock.
02
Who reviewed it
The reviewer who signed the hard case stays named on the record, not lost in a screenshot thread.
On the record03
Under what rights
The license the work was created under travels with it — so the record survives an audit.
On the recordIdentity · verified consentReviewer · named on the recordRights · attached to the datapoint
FIND YOUR FIELD
Each story starts with the pain that triggered the work.
Post-training and evals, model risk under examination, manufacturing release, clinical second reads, robotics episodes, training-data provenance. Jump to your field and read the same loop you would run.
POST-TRAINING & EVALS
A frontier AI lab
An escaped regression had nowhere to go.
The failure became the next gate.
Eval tooling stopped at scores and traces. RLHF adjudication, the reviewer who signed it, and the regression it caught now live on one record — and that record holds the next release until it passes.
One record
per release, reviewer named
↳ RELEASE GATE
MODEL RISK UNDER EXAMINATION
A regulated decisioning program
A score does not survive examination.
A governance packet does.
Borderline credit, fraud, and AML decisions left a risk score and little else. Now each one carries the reviewer, the escalation, and the regression memory — one examiner-ready packet instead of a screenshot debate.
One packet
per audit cycle, escalation attached
↳ SIGNED PROOF
MANUFACTURING & QUALITY
A quality-engineering team
Detection is not release.
The part passes on a record, not a hunch.
Visual defect detection flagged parts but never approved a release. An escaped defect now becomes a gate: the reviewer signs the call, and the same miss is caught again before the next part ships.
One gate
per escaped defect, signed off
↳ RELEASE / HOLD
CLINICAL SECOND READS
An academic medical center
No record proved who read the hard case.
The second read went to the right reader.
Uncertain inferences routed back to the radiologist qualified for the body region. The rubric, the override, and the reader who signed it stay attached to the case — a record a clinician will trust.
One case
routed to the qualified reader
↳ REVIEW RECORD
ROBOTICS EPISODES
A humanoid program
Failure data came with a gig-platform mess.
Capture, review, export — one chain.
Task-diverse manipulation and failure episodes captured in real environments, with safety review and rights-cleared consent on every clip. Demonstration capture, teleop review, and export lineage stay on one workflow — a second source you can defend.
100–500
reviewed episodes per pilot band
↳ EPISODE LINEAGE
TRAINING-DATA PROVENANCE
A team procuring under the EU AI Act
78% can't validate training data. 77% can't trace it.
Every datapoint traces to its origin.
When enforcement begins August 2026, a high-risk model has to show where its training data came from. Each record here carries who made it, who reviewed it, and under what rights — a neutral second source you can defend under audit.
Aug 2026
EU AI Act provenance clock
↳ PROVENANCE TRAIL
WHY BUYERS COME HERE
“We needed a second source we could defend under audit. The work comes in, the reviewer signs it, and we keep the weights.”
BRING THE WORK
Bring the work. Keep the proof. Own the model.
These are program patterns you can run today. When evaluation moves to procurement, a private reference opens under NDA — named, with the numbers.