Deterministic decisions, in depth.
The practical fork away from both giant generalist LLMs and narrow specialist models — sovereign, auditable, and measured on real workloads. Here is the full architecture and reporting.
Two camps. A third path that owns the decision.
The generalists race toward trillion-parameter models that re-reason everything from scratch, stochastically, from a prompt. For a repeatable, high-stakes operational decision they are the wrong instrument: multi-second latency, per-call cloud cost, outputs that differ run-to-run at the same input, no native audit trail, and a documented tendency to manufacture decisions where the correct answer is often "do nothing."
The specialists fine-tune a smaller model per task — cheaper, but fragmented: a model per task, each frozen between retrains, none aware of the full operational state, none governed by a shared auditable policy layer. It scales in headcount, not leverage.
Verificate's third path. For operational, repeatable, high-stakes decisions we build domain-scale Decision Transformers that cover the decision space of an entire vertical — trained on the complete forward projection of every relevant attribute. The DT owns the decision via an argmax over output logits on a full, uncapped state vector. A general LLM is deliberately secondary: it interprets intent and narrates the deterministic output. It never decides.
| Generalist LLM | Specialist fine-tune | Verificate — Domain DT | |
|---|---|---|---|
| Decision owner | LLM (stochastic) | small model, per task | DT — deterministic argmax |
| Same input → same output | No | Usually | Yes — mathematically identical |
| State it sees | retrieved chunks / context | hand-picked features | full forward projection of every attribute |
| Latency per decision | seconds (cloud) | ms (GPU) | ~0.6–5 ms on a commodity CPU core |
| Audit lineage | none native | none native | per-decision, in the substrate |
| Adapts to drift | quarterly fine-tune | periodic retrain | continual, single-digit-second loop on CPU |
| Data egress | third-party cloud | varies | zero — sovereign, on-prem / on-device |
| LLM's role | everything | none | interpret + narrate only |
The credible players cluster at two poles, leaving the third open. Frontier-LLM agents put a stochastic model on the decision — flexible, but non-reproducible and unauditable at the boundary. Rules-engine and decision-intelligence platforms reach for determinism the other way, but cannot see the full state, don't learn optimal actions from outcomes, and degrade into an unmaintainable thicket. Verificate occupies the gap all three leave: the only synthesis that is simultaneously deterministic, learned, state-complete, auditable, and sovereign.
Not a feature store. An active execution substrate.
A feature store materialises features; a RAG index retrieves chunks. Neither executes policy, forward-projects the complete state, nor deterministically re-wires downstream behaviour when a goal changes. Ours does all three — its core invention is a forward-projection engine.
Executable attributes
Goals, policies, thresholds, routing gates and evaluation criteria are stored as versioned attributes inside the same store as the data they govern — not in application code. Policy is data; the store is a coordination language, not passive memory.
Cascading recomputation
Change a high-level goal, threshold or rule and every downstream computation updates deterministically, with no code deployment. Measured at O(changed) cost — 411× cheaper than full re-materialisation, 13–814× vs a strong incremental baseline on a real 251k-cell workload.
Full forward projection
At training and inference every qualifying attribute is projected into the DT state vector. The loader refuses to train on a truncated state — the model always sees the complete picture. This is the opposite of feature engineering, and it is what lets a single DT cover a vertical's decision space. ~0.7 ms median projection at 315k live entities.
Self-improving controller
A meta-over-object controller governs the substrate's own improvement against a monotone attainment objective with a frontier bonus, enforcing safety and stability as hard constraints. Self-modification is admitted only on a strict attainment increase with zero regression — correcting a documented variance-reward inversion that rates value highest at low performance.
The DT sees every attribute. The LLM only speaks.
covered fact → answer with provenance, zero LLM / retrieval call · <300 ms warm · identical state → identical action, every decision writes audit lineage
Live pricing, second by second
A 32B-class pricing DT paired with a 32B narrator R gives a rep — or a dynamic-pricing engine — auditable discounting guidance while the customer is live: the DT reads the full projected state (inventory, margin, customer history, competitive pressure, demand, campaign constraints) and returns the optimal discount ceiling deterministically in milliseconds; the narrator explains why in one sentence; the lineage records exactly which policy attributes drove it. A RAG loop is too slow and non-reproducible; rules can't span the state; a narrow model can't see it.
Construction material flow
Given the complete state of every material movement across warehouses and job sites, the DT recommends the next action — reallocate, expedite, re-source, hold, escalate — with a quantified expected outcome in ~0.6 ms M, and the narrator turns “reallocate 40 pallets of rebar, ETA drift +2 days” into guidance a project manager acts on. The same pipeline materialises a 1M-entity corpus in 2.63 s and retrains in ~54 CPU-seconds.
The governed proposal layer — the LLM's other narrow role
Beyond narration, the LLM may propose a policy change — a new reward or evaluation formula in natural language — which the substrate validates, materialises and gates before it can touch a live decision. In an engineering run on a real 251k-case substrate, a governance model proposed a reward; the engine materialised it as a typed executable attribute, back-computed it over history in 561 ms, bit-exact, with full lineage — while a malicious injection proposal was rejected by the compiler guard M. On real protected demographics, a governed iterative loop drove a socioeconomic-proxy-biased reward from an adverse four-fifths ratio of 0.05 → 1.07 M. Prompt engineering becomes a typed, auditable, reversible schema change — the LLM proposes, the engine decides and records.
Production-capable, not researchware.
Measured on real, public, non-toy workloads with released harnesses — present to make the vision credible, not to oversell.
On a real 251k-case procurement log, the continually-retrained flywheel beats a frozen policy (signed-rank p≈0.009, positive across every seed). A label-permutation placebo collapses the gain 389× — the value is freshness, not volume. M
411× cheaper than full re-materialisation; 13–814× vs a strong incremental baseline. A compliance officer raising a risk threshold re-wires the live decision surface in seconds, with full lineage. M
Zero argmax mismatches between the C++/ONNX production engine and the Python reference, with ~20× tighter tail latency. The deterministic decider on the critical path. M
Reactive recompute + retrain + hot-swap closes in single-digit seconds on one commodity CPU core. 1M-entity retrain: 54 CPU-seconds. Per-adaptation energy ~10⁴× below an LLM LoRA fine-tune. No GPU anywhere. M
Via a randomised trial plus a ground-truth heterogeneity sweep, a falsifiable rule: a learned policy beats a fixed rule only in the regime of pricing, discounting, and material re-sourcing. Below that threshold, a fixed rule is honestly optimal. M
On OULAD, the policy shows no significant disparity by disability or socioeconomic deprivation — auditable via lineage, not asserted. Zero-shot LLM advisors on the same task show severe bias that evades standard safety tooling. M
The substrate is production-capable today at reference scale; the path to domain scale is engineering, not invention. The Living Decision Flywheel reproduction harness runs with one command — flywheel run all-local — that reproduces all results from committed public-dataset-derived data, with no database or cluster required.
Licensable, sovereign, production-grade.
Read the papers. Run the code.
Seed a category-scale Decision Transformer.
Book a call to discuss academic or commercial access — a deterministic, auditable decision engine trained on your operational data and deployed inside your boundary.