The Decision Transformer
that owns the decision.
Domain-scale Decision Transformers, trained on the complete forward projection of every attribute. The decision is a deterministic argmax on a full state — reproducible to the bit. The LLM only interprets and narrates. It never decides.
Two camps. A third path.
Generalist LLMs re-reason everything; specialist fine-tunes go blind to the full state. The DT is a learned, deterministic policy over the whole picture.
| Generalist LLM | Specialist fine-tune | Verificate DT | |
|---|---|---|---|
| Decision owner | LLM (stochastic) | small model, per task | DT — deterministic argmax |
| Same input → same output | No | Usually | Yes, to the bit |
| State it sees | retrieved chunks | hand-picked features | full forward projection |
| Latency / decision | seconds (cloud) | ms (GPU) | ~0.6–5 ms, CPU |
| Audit lineage | none native | none native | per-decision |
| Adapts to drift | quarterly fine-tune | periodic retrain | continual, seconds on CPU |
| Data egress | third-party cloud | varies | zero — sovereign |
Not a feature store. An active execution substrate.
Executable attributes
Goals, policies, thresholds and evaluation criteria live as versioned attributes inside the same store as the data they govern. Policy is data, not application code.
Cascading recomputation
Change a goal or threshold and every downstream computation updates deterministically — no code deploy — at O(changed) cost, 411–814× cheaper than re-materialising.
Full forward projection
Every qualifying attribute is projected into the state vector. The loader refuses to train on a truncated state — the model always sees the complete picture.
Governed proposal layer
An LLM may propose a new reward or rule in natural language; the engine validates, materialises and gates it before it can touch a live decision — bit-exact, with lineage.
// the LLM is on the interpretation and narration ends — never on the decision path
The technical deep dive.
The full architecture and reporting: the forward-projection engine, the governed proposal layer, worked examples (live pricing, construction material flow), and every measured result with released reproduction harnesses. M measured · R roadmap