LAB.02 / Health & Pharma

A plausible answer is not an acceptable answer.

In pharmaceutical or cosmetic formulation, a system that is right ninety-five per cent of the time and cannot tell you which five per cent it got wrong is unusable. This laboratory works where the cost of error is real and every recommendation has to justify itself.

01 / Philosophy

What this laboratory studies, and why.

Object

A development team formulates by iteration. Each product crosses dozens of versions before the one that ships, and every step encodes a decision: why an active was reduced, why an acid was removed, what was being sought. That reasoning is never written down. What remains is a spreadsheet of numbers.

The central question of this laboratory is the division of labour between the deterministic and the generative. The model may propose; the norm is applied by code. Limits come from a database transcribed from the official source, never from a model's memory, and validation runs reproducibly without consuming a model at all.

The open problem is how much of the decision can be delegated before traceability becomes theatre.

Why

The interesting question here is not whether a model can propose a formula. It can. It is who answers when the proposal breaches a limit and nobody noticed. Until that has an architectural answer — which part of the system is never allowed to be wrong, and why — any deployment is risk handed quietly to the formulator.

Establishing it beforehand is cheaper than discovering it in an audit. And part of it can only be learned on a real development record: how much of a formulator's reasoning was ever written down, how much is reconstructable, and how much is irreversibly gone.

How it works
Deterministic layer

Regulatory limits, allergen thresholds and authorised preservatives, transcribed from the official register. A formula validated twice returns exactly the same result. No invention is possible because no model is involved.

Generative layer

Interprets why a formulator made a decision, drafts the reasoning behind a proposal, infers the function of an unknown raw material and writes the questions the team should answer. Everything it produces is labelled as hypothesis.

Rule A figure without a verifiable source does not enter the database. It is reported as not verifiable rather than assumed correct — the error that ruins an audit. When an ingredient arrives as a blend without a declared proportion, non-compliance cannot be asserted: the warning drops in severity and requests the supplier's data sheet.
02 / Projects

Systems and open lines.

  1. P-01
    In production

    Formulation history as a knowledge base

    The system reads the full development record — including the per-sample laboratory sheets nobody was opening: organoleptics, measured pH, density, manufacturing method, formulator comments — and reconstructs the sequence of decisions, not just the sequence of formulas. It then proposes the next iteration ranked by confidence, with the reason for each change.

    • Interpreted differences between versions, each marked as data, inference or hypothesis
    • A raw-material dictionary that grows: function, typical phase, in-house range, every commercial name seen
    • Rules inferred with their supporting evidence and a declared confidence level
    • Deep analysis once, on the mature version; delta analysis on the intermediate ones
  2. P-02
    Open

    Demonstrating without disclosing

    A marketed formulation is a laboratory's most sensitive asset, yet the system has to be shown to third parties. We work with an explicit anonymisation map: products, raw materials and suppliers are substituted consistently so that the technical structure stays intact — proportions, phases, relations, the logic of the iterations — while the real formula stops being reconstructable.

  3. P-03
    Open

    Diagnostic and medical imaging support

    An open line with no system in production. We state that plainly because in this domain the temptation to imply experience one does not have is high and the consequences of believing it are serious. What we do carry over from formulation is the architectural principle: the model may order, prioritise and describe, but the threshold separating one clinical decision from another cannot depend on it.