The Boundary Refinery

The Story

The most valuable computational knowledge in the world economy is, in a word, radioactive. Pharmacokinetic models refined over decades of trials. Catastrophe models that price hurricane risk. Trading signals, yield models, actuarial formulas. Publishing them destroys their value; pasting them into a cloud prompt may violate the law. So this knowledge and frontier AI have never been able to touch — and when they try, the model simply hallucinates a confident imitation.

The standard playbook says: give the model the knowledge. Fine-tune it in, retrieve it with RAG, paste it into context. For executable knowledge, all three are category errors. Fine-tuning teaches the model to imitate a risk model — an approximation of something whose entire value is exactness. RAG retrieves a description, and then the LLM predicts the computation — precisely the failure mode PrOOPt exists to prevent. Context-pasting is simultaneously a leak and a lie.

The Boundary Refinery inverts the playbook: the model never receives the knowledge at all. It receives an interface.

How It Is Built

Proprietary simulators are registered as @CodeFunctions — a knowledge injection layer the model can invoke but never inspect. Reasoning about a tool is comfortably within the model's distribution, even when the tool's contents are permanently outside it.

public class DoseOptimizationCopilot {

    @CodeFunction(description = "Proprietary population-PK simulation: "
            + "steady-state concentration-time profile for a dosing regimen "
            + "given covariates. Validated model vX.Y; returns AUC, Cmax, Ctrough.")
    public ExposureProfile simulateExposure(Regimen regimen, Covariates covariates) { /* … */ }

    @PromptFunction(model = ModelTier.LOCAL,
            prompt = "Extract structured covariates (age, weight, eGFR, "
                   + "comedications, organ function) from this clinical note: {note}")
    public Covariates extractCovariates(
            @SensitiveData(label = "***CLINICAL-NOTE***") String note) { return null; }

    @PromptFunction(model = ModelTier.CLOUD_ADVANCED, thinking = true,
            prompt = "Design a simulation campaign to answer: {question}. You may "
                   + "only propose calls to the registered simulators...")
    public CampaignPlan planCampaign(String question,
                                     CovariateSummary deidentified) { return null; }
}

@PromptOrchestrator(
        prompt = "Answer dose-optimisation questions strictly via registered simulators.",
        planMode = PlanMode.STATIC,
        planCacheStrategy = PlanCacheStrategy.INTENT,
        parallel = true
)
public Recommendation advise(String clinicianQuestion) { return null; }

The division of labour is strict. ModelTier.LOCAL handles every step that touches raw sensitive data, inside the JVM. The cloud tier sees only de-identified, schema-shaped summaries, and contributes what it is genuinely good at: fluent scientific reasoning about which validated computations to run and what the results mean. Simulation grids fan out in parallel waves on virtual threads, and every simulator invocation — version, input hash, interpretation — lands in the audit trail, ready to staple to a regulatory submission.

Why It Is Unprecedented

Every “chat with your data” product ships your data to the model. This is “chat with your models” — and the models never leave home. The training-data boundary is permanent, and the Refinery treats it not as a problem to solve but as a seam to exploit. The expert gets frontier-AI fluency attached to ground truth she is professionally accountable for, rather than substituting for it. And when the compliance question arrives, the answer is short: the AI never computed anything. It decided which of our validated computations to run, and explained the results — and here is the complete record.

The Fine Print

The Refinery's outputs are exactly as good as the registered simulators; it amplifies validated science and validated mistakes with equal enthusiasm, so model governance remains a human discipline — now perfectly recorded. The cloud planner can propose wasteful campaigns; cost guards belong in configuration and deterministic constraint checkers, not in trust. And this is decision support for qualified professionals, not autonomous prescribing — the recommendation type should say so on the tin.