01Identity
A model is more than its weights.
Dataset version, training configuration, source code, base model, GPU profile, evaluations, safety results, cost, latency, runtime and approval status form one identity.
Draft → monitor → retrain
Governed Model Lifecycle
Every PDICON model moves through draft, training, evaluation, candidate, human validation, production, monitoring and retraining states with a complete versioned record.
StatusModel governance design
Reviewed2026-08-20
Direct answer
Every PDICON model moves through draft, training, evaluation, candidate, human validation, production, monitoring and retraining states with a complete versioned record.
Stage-gated sequence
Expansion follows evidence.
- 01Gate 1Draft task specification
- 02Gate 2Train against versioned dataset
- 03Gate 3Run technical and safety evaluation
- 04Gate 4Register candidate artifact
- 05Gate 5Complete human validation
- 06Gate 6Deploy approved runtime
- 07Gate 7Monitor and collect corrections
- 08Gate 8Retrain from governed evidence
02Promotion
Evaluation precedes production.
A trained checkpoint becomes a candidate only after automated gates and becomes production only after qualified human validation.
03Operation
Monitor behavior and outcome.
Runtime quality, drift, cost, latency, corrections and project outcomes determine whether to retain, roll back or retrain.
Operating matrix
Evidence moves through explicit controls.
| Subject | Input | Intelligence operation | Human / policy control | Output |
|---|---|---|---|---|
| Draft | Task and dataset proposal | Specification review | Owner approval | Train-ready job |
| Candidate | Checkpoint and metrics | Evaluation review | Domain validation | Promotion decision |
| Production | Approved model and runtime | Monitoring | Rollback policy | Operational evidence |
Design boundary
What the system will not pretend to be.
No mutable mystery endpoint
Production version and configuration remain resolvable.
No metric-only approval
Qualified users inspect representative workflow behavior.
No silent retraining
New datasets and candidates repeat the governed lifecycle.