Input
What you provide
- Named model candidates and versions
- Locked rubric, thresholds, data slices, and measurement version
- Known regressions, release policy, approvers, and rollback owner
Models / Model Launch Check
Decide a model release against the same checks.
Compare model candidates under the same checks, replay known failures, identify blockers, collect approvals, and produce a source-linked launch file.
Available nowScoped around one candidate set and release decision.
CandidateAtlas 3.2
DestinationProduction
DecisionHold
- EvaluationSame rubric and source slice
- Regression replayOne failed case needs review
- Approval chainSafety signer remains
- RollbackOwner and procedure attached
Release gate2 requirements remainClose the failed-case review and collect the required signer.
Input, work, and output
Work
What happens
- Run the same evaluation and operating checks
- Compare deltas, confidence, cost, latency, and failed examples
- Replay known failures and expose blockers before approval
Output
What you receive
- Ship, hold, or return decision
- Required checks and completed approvals
- Launch file, rollback plan, supporting records, and retained regressions
Decision flow
Move from intake to review and handoff with clear owners at every step.
01
Candidate input
Confirm candidate versions, owners, intended release, and evaluation context.
Candidate manifest
02
Same-check comparison
Lock the rubric and source slice before scores or rank are displayed.
Comparison matrix and source runs
03
Regression replay
Replay retained failures and keep each failed case visible.
Regression outcomes and affected versions
04
Blockers
Name every failed, missing, expired, or permission-restricted requirement.
Blocker owner and recovery
05
Approval
Show who can approve, who has approved, and what remains.
Approval chain and timestamps
06
Launch file
Export the decision, evidence, destination, checksum, and rollback.
Signed release decision
What you receive
Decide a model release against the same checks. deliverables
- Candidate
- Model and version identityProvider, owner, stage, and intended destination
- Checks
- Evaluation, regression, safety, operating, and policy stateVersioned source runs and evidence
- Blockers
- Failed, missing, expired, or restricted requirementNamed owner, consequence, and recovery
- Approval
- Required signers and completed decisionsActor, role, time, scope, and expiration
- Release
- Launch destination, manifest, checksum, and rollbackGenerated launch file
What the workflow supports
| Outcome | Work | What you receive | Program fit |
|---|---|---|---|
| Ship | All required checks pass, approvals are current, and rollback is present. | Source-linked results, approvers, launch manifest, and generated time. | The launch file includes the approved candidate, environment, checks, and rollback plan. |
| Hold | A required check fails, expires, or lacks sufficient evidence. | Blocking criterion, failed source run, consequence, owner, and next action. | The owner and required next action remain visible until the blocker is cleared. |
| Return for work | The candidate needs remediation, new evaluation, or a corrected release record. | Returned items, responsible owner, expected evidence, and review route. | The candidate returns through the same checks after the required work is complete. |
Inspectable evidence
Candidate comparison exposes the decision, not a leaderboard.
Release review with descriptive states, blockers, owners, and next actions.
Interactive walkthrough
Compare two candidates under one locked launch check.
Select candidates and a launch criterion to see where each release is ready, needs review, or remains blocked.
| Required check | Candidate A | Candidate B |
|---|---|---|
| Review required | Evidence ready | |
| Blocking issue | Evidence ready | |
| Review required | Review required | |
| Evidence ready | Evidence ready |
Criterion inspector
Known regressions
- Required evidence
- Replay of retained failures under the same candidate
- Source
- Regression Bank replay record
- Decision effect
- A blocking retained failure holds the candidate.
- 1
Same checksCandidates share the same rubric and measurement version.
- 2
Blocking evidenceA failed regression remains visible and source-linked.
- 3
Release effectEvery difference states how it changes the decision.