Offline Retraining on L0

L0 is where every model is retrained: a slow, expansive, reproducible compute substrate that regenerates ground truth before any artifact is compiled toward the machine.

L0 as the retraining substrate

Retraining is intentionally exiled to the L0 Foundation layer, the non-real-time compute base of the stack. L0 is allowed to be slow and thorough: multi-physics Monte Carlo neutronics, GPU-cluster PINN training, and petabyte-scale replay of pulse histories all run here. No retraining job ever runs on the L1 control plane, which must stay deterministic.

Every retraining job is a first-class, reproducible unit of work. It pins its dataset version, its base-model version, its random seeds, its container image, and its hardware profile, then emits a signed artifact plus a lineage record. Two runs of the same job on the same inputs must produce byte-identical or tolerance-identical models; this is enforced, not hoped for. See reproducible training runbooks.

Retraining triggers

• Scheduled cadence (e.g. after each breeder campaign block)
• Drift breach flagged by drift monitors
• New physics regime explored (breeder TBR lever 1.1/1.5/1.8, burner plug-density scan)
• Incident-driven retrain after a postmortem
• Active-learning request for high-value samples

# Retraining job manifest (pinned for reproducibility)
job = RetrainJob(
  base_model = 'burner/ambipolar-potential@3.2.1',
  dataset    = 'l0://datasets/burner_pulses@2027Q2.v14',   # content-addressed
  seed       = 20270415,
  image      = 'kronos/train:cuda12-pinn-0.9@sha256:...',
  budget     = {'gpu_hours': 4096, 'nodes': 32},
  emits      = ['model.safetensors','lineage.json','card.md'])

Because the machines are not yet built, L0 retraining today runs largely against simulation and twin-generated data. Once the breeder reaches first tritium near 2030, real pulse evidence begins to dominate the datasets, but the substrate and the discipline stay the same. L0 remains the only place weights change.