Design scientific decision systems around methods, not prompts

Prompts are useful for interpretation. They are a poor place to hide a scientific method.

Put repeatable logic in code

ARiDA uses language models to plan, acquire, reconcile and explain. Stable decision logic lives in typed calculators and governed engines. Target scoring has explicit dimensions, coverage rules, contradiction penalties, safety vetoes and archetype-specific thresholds. Indication ranking has declared criteria, gates and weight sensitivity. Valuation has separate modules for reference classes, Monte Carlo rNPV, real options, information value and portfolio construction.

The distinction matters. A model may explain why a safety finding is relevant. It should not improvise whether a severity score of 0.86 crosses the approved veto threshold. A model may locate a molecular record. RDKit should compute the fingerprint (RDKit).

Give evidence a schema

Open Targets demonstrates why scientific evidence needs typed sources and relationships rather than a pile of text (Open Targets). ARiDA extends that discipline into the decision path. Records declare their kind, entity, source references and ledger identifiers. Calculators specify required inputs. Missing fields produce gaps or identity behavior.

This produces useful restraint. Incomplete chemistry does not create portfolio correlation. Missing economics does not create rNPV. Analyst-owned PoS does not change without an explicit action.

Make orchestration enforce the method

Governed programs encode lane graphs, tool policies, required outputs and release gates. The agent can choose how to pursue a bounded task, but it cannot redefine the contract while running. Deterministic fallbacks remain deliberately below launch thresholds when they lack enough evidence.

NIST’s framework calls for systems to be evaluated in context, documented and able to fail safely (NIST). Structured tools make those properties testable. Prompt wording does not.

The aim is not to remove expert judgment. It is to reserve judgment for the places where it belongs: interpreting evidence, choosing assumptions, setting thresholds and authorizing action.