What a decision system must prove inside pharma

A pharmaceutical company should not evaluate a decision system with a collection of clever prompts. It should give the system one live decision and ask it to survive the organization.

That means surviving scientific review, data rights, model governance, finance challenge, executive compression and the arrival of new evidence. Most demonstrations stop after the first step.

Choose a decision with consequences

An indication sequence, target investment, licensing case or portfolio allocation is a better test than “summarize the field.” The evaluation team should specify the decision owner, the permitted evidence, the valuation date, the acceptable analytical methods and the release authority.

Then ask a harder question: what scientific input can change the answer? A molecular liability might reduce a benchmark PoS. Human genetics might strengthen a target hypothesis. A serious safety signal might veto an advance recommendation. A trial change might alter the competitive scenario without changing the intrinsic value. The system should show these distinctions rather than decorate every new fact with urgency.

Separate authority by layer

Evidence access does not confer authority to change an assumption. Code execution does not confer authority to release a decision. ARiDA separates acquisition, deterministic computation, synthesis and approval.

Governed lanes are restricted to sanctioned tools and evidence sources. Structured records enter a source ledger. Scientific engines and calculators operate on frozen inputs. Output gates check source binding, required objects and release conditions. A writer can explain accepted results, but it cannot manufacture a missing economic value or turn an unresolved safety issue into a confident recommendation.

This separation matters in ordinary details. ClinicalTrials.gov provides public structured records, but registry fields can be revised and are supplied by sponsors (ClinicalTrials.gov). Regulatory interpretation must preserve jurisdiction, document version and product context; the relevant authority is the current guidance, not a summary that has circulated internally for years (FDA).

Demand continuity

Pharma decisions unfold over weeks and return months later. The plan, evidence files, calculations, comments and approvals must survive interruptions. ARiDA uses persistent sessions and file-backed workspaces, but the important product behavior is above that infrastructure: the case can be reopened with its state intact.

When work is divided among scientific, clinical, patent, market and valuation specialists, each receives the same decision frame and returns durable artifacts. Collection is a controlled event. The final synthesis is based on those artifacts, not on a model’s memory of earlier prose.

Test failure, not only success

A serious evaluation should remove a source, provide incomplete chemistry, submit an unsupported number and interrupt a run. The desired behavior is often a gap, a blocked release or an unchanged calculation. NIST’s framework explicitly includes validity, reliability, transparency, monitoring and safe failure among the characteristics and practices of trustworthy AI systems (NIST).

The procurement question is therefore concrete. Can the system preserve organizational judgment while making the science computationally active? If it can only write, search or chat, it is solving a smaller problem.