CI/CD for model knowledge: when a model's source of truth changes, detect the drift, build grounded evidence, choose the least-cost intervention, and promote or rollback the resulting model state.
SENSE → BUILD → DECIDE → VERIFY
Mac = cockpit | GitHub = code | B2 = data | Lambda/Modal = workers
Workers are disposable. Code and data are not.
Start with:
- Project DELTA system contract
- DELTA roadmap and current status
- Developer workflow
- Experiments registry
The active vertical slice is e1_vllm: version
fidelity and failure modes across vLLM documentation revisions.
./scripts/lab run --target modal \
--config configs/experiments/e1_vllm/c0_base_eval_qwen35_08b_modal.yamlWorld changes produce versioned evidence.
Interventions are promoted only after recovery and regression checks.
The cheapest intervention satisfying the quality constraint wins.
Workers, local caches, failed candidates, and rejected interventions.
Pinned source revisions, dataset manifests, run evidence, promotion decisions, and model-state lineage.