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
Every PDICON model moves through draft, training, evaluation, candidate, human validation, production, monitoring and retraining states with a complete versioned record.
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
Dataset version, training configuration, source code, base model, GPU profile, evaluations, safety results, cost, latency, runtime and approval status form one identity.
A trained checkpoint becomes a candidate only after automated gates and becomes production only after qualified human validation.
Runtime quality, drift, cost, latency, corrections and project outcomes determine whether to retain, roll back or retrain.
Operating matrix
| 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
Production version and configuration remain resolvable.
Qualified users inspect representative workflow behavior.
New datasets and candidates repeat the governed lifecycle.