STEP 01
ON YOUR OWN WORK
Train
Tune on the data you reviewed, with consent and license scope already attached. Checkpoints and lineage hold from the first step.
→DEVELOPERS · INFRASTRUCTURE
Promotion is a signed event.
A train, register, and deploy path where no model reaches production without a reviewer record attached. Train on your own reviewed work, sign the release, and keep the weights — not a rented endpoint.
TRAIN · REGISTER · DEPLOY
PROMOTION
Signed event
No model reaches production without a reviewer record attached: who approved it, against which evals, on what date.
TRAINING DATA
Consent attached
Every dataset carries who made it and under what rights — the answer when an EU AI Act auditor asks where it came from.
WEIGHTS
Yours to keep
Start open, tune on your own reviewed work, keep the weights. A model you run, not a rented endpoint.
HOW IT WORKS
Train. Register. Deploy.
One governed path from your reviewed work to a signed release. Each step keeps the consent chain and the reviewer record attached — so the model that ships can survive an audit.
STEP 02
WITH A SIGNATURE
Register
No model moves to production without a reviewer record attached. Each promotion is signed: who approved it, against which evals, on what date.
→STEP 03
KEEP THE WEIGHTS
Deploy
Ship the approved model with rollback and traffic controls wired. The weights are yours to run — and every regression you catch becomes a gate the next release must pass.
WHAT THE PATH HOLDS
Proof, attached at every step.
From the training data to the signed release to the weights you keep — every part of the path carries who made it, who reviewed it, and under what rights.
STAGE · 01
TRAINING DATA
Every dataset carries who made it.
Datasets arrive with consent, reviewer record, and license scope already attached. When an auditor asks where the data came from, the answer is in the run.
CARRIES · Consent · review · rights
STAGE · 02
TRAINING RUNS
The run keeps its lineage.
Train on your reviewed work with checkpoints, tracked configs, and lineage held from the first step. The next reviewer picks the run back up where it stopped.
CARRIES · Checkpoints · configs
STAGE · 03
MODEL REGISTRY
Promotion is a signed event.
No model moves to production without a reviewer record attached. Each transition is signed: who approved it, against which evals, on what date.
CARRIES · Stages · reviewer · sign
STAGE · 04
RELEASE GATES
Failures are remembered.
Every regression you catch becomes a gate the next release must pass. The same mistake does not ship twice.
CARRIES · Gates · regression bank
STAGE · 05
SERVING
Roll out, roll back.
Approved builds move into serving with versioning, traffic controls, and rollback in place. Promotion never leaves the signed path.
CARRIES · Versions · rollback
STAGE · 06
OWNED WEIGHTS
You keep the weights.
Start open, tune on your own reviewed work, and keep the weights that result. A model you run, not a rented endpoint you lose access to.
CARRIES · Open · tuned · yours
THE PATH
From reviewed work to signed release.
Four steps. The consent chain enters with the data and stays attached through the signed release — so the proof is there when an auditor looks.
01
Reviewed work in
Datasets enter with consent, license scope, and reviewer record already attached. The provenance chain starts before the first training step, not after.
↳ Consent · review · lineage
02
Train on your own work
Tune on the data you reviewed, with checkpoints and tracked configs. Runs stay reproducible and the weights stay yours.
↳ Checkpoints · configs · weights
03
Register with a signature
Approved builds move into serving with rollback and traffic controls. Each promotion is signed with the reviewer record that approved it.
↳ Registry · signed gates · rollback
04
Evidence downstream
Release events, the reviewer record, and the consent chain reach operators in Control Center alerts and signed webhooks — where auditors already look.
↳ Signed webhooks · alerts · audit trail
INFRASTRUCTURE
Bring the work. Keep the model.
Train on your reviewed work, sign the release, keep the weights. The consent chain and the reviewer record stay attached the whole way — so your data survives an audit and your model stays yours.
↳ STARTS WITH
Your reviewed work, a model to tune, and a release you need to defend under audit.
↳ LEAVES WITH
A signed promotion with the reviewer record attached, regression gates set — and the weights in your hands.