These are the non-negotiables. They encode truth over compromise into rules a dev team can follow without having to re-derive the stakes every time. Break one and it isn't an OpenDiabetic cook.
The Diabetic Crockpot Cooks menu has exactly one tier: 5-cap (premium, A+, full discipline). No experimental side-cooks, no proofs-of-pipeline dressed as dishes. A cook that fails a gate is re-cooked (v+1) or killed — never demoted onto the menu at a lower cap. Flagship-grade or it doesn't get plated.
We own the rigs and the electrons. Cooks are judged on one axis: is it best-in-class? We never trim epoch length or steps to save time, energy, or money. Training length is a quality choice, governed by early-stopping (the overcook guard), not the clock. Energy is tracked as a receipt (provenance), never as a constraint.
Never cook before a full rig evaluation. Inventory the GPUs, evict slop, confirm the silicon is healthy and idle, then cook. A contended GPU spills to CPU and silently corrupts the run. The rigs are senior managing directors — preserve the hardware (cap power for thermals on long burns).
"Open" means the recipient can verify it themselves: schema, real samples, exact size, SHA-256 — inline, before download. Metadata-only "open" is trust-me-bro. And dedup honestly: publish the true unique count, never the inflated sum of overlapping parts.
Every dataset and every cook carries a deed/receipt — tribunal-graded, hash-chained, publicly anchored. Provenance is verifiable independent of us. Verifiability is the moat.
A model is not "good" because the loss fell. It must beat the base model on a held-out A/B with deterministic gates — think-off, deterministic decode, rule-scored, no LLM-judge. The A/B result is a receipt. No beat-base, no ship.
Always smoke-test the full pipeline (load → LoRA → steps → eval → save → merge) before the full burn. The canary's job is to fail cheap. Also smoke-test trainer-init on the full dataset — some bugs are scale-dependent and a small canary will miss them.
Every cook is a flightsheet anyone can audit: intent, config, corpus SHA, canary fixes, in-flight receipts, beat-base A/B, sign-off. We show the math. The "AI trust-me-bro" era is over.
Private health data never leaves the box. Models flow down, receipts flow up, PHI crosses never. Training corpora are open/synthetic/cited — never patient records. The patient-facing edge organizes and reminds; it does not diagnose. The clinician-grade anchor may train on physician-grade QA but is a separate model class with a stamped intended-use. → The Firewall
"OpenDiabetic is a slow cooker — why? because lives depend on us. Truth over compromise."