The Clinical Trial Principle: Every Deployment Is an Intervention
A properly planned clinical trial is the best technique for assessing effectiveness. It also discovers harms. Apply the same evidence hierarchy to software deployment — from unit tests (preclinical) through post-market surveillance.
The Clinical Trial Principle
A properly planned and executed clinical trial is the best experimental technique for assessing the effectiveness of an intervention. It also contributes to the identification of possible harms.— ICH E2A
Every infrastructure deployment is an intervention. The Clinical Trial Principle requires that deployments follow the same evidence hierarchy as drug development.
The Phase Mapping
| Phase | Clinical Trial | Software Equivalent |
|---|---|---|
| Preclinical | Mechanism of action | Unit tests, cargo test |
| Phase I | Safety in small group | 5-session crash and latency test |
| Phase II | Efficacy signal | 10-session effectiveness measurement |
| Phase III | Comparative advantage | 20-session A/B with control baseline |
| Phase IV | Post-market surveillance | Ongoing monitoring via autopsy records |
Two Sentences, Two Truths
Sentence 1: Effectiveness. "Properly planned" means prospective design with endpoints defined BEFORE the trial. "Best experimental technique" means superior to observation. "Effectiveness" means: does it achieve what we intended at the consumer boundary?
Sentence 2: Harms. "Also contributes" means harms are discovered as a BYPRODUCT of proper trials. You cannot discover harms through observation alone — you need controlled exposure. Uncontrolled deployment discovers harms through user suffering. This is why Phase I exists: controlled harm discovery before broad exposure.
The Autoimmune Principle
An incorrect safety mechanism is worse than no safety mechanism. A false-positive antibody attacks self.
Design for the harm case FIRST: what does failure look like? What does autoimmune attack look like? A test suite that passes on broken code is an autoimmune disease — it provides false confidence that actively prevents you from seeing the problem.
Five Controls
- Selection bias — do not cherry-pick easy cases for testing
- Confounding — isolate the intervention from other changes
- Temporal ambiguity — measure BEFORE and AFTER (prospective, not retrospective)
- Observer bias — automated measurement, not self-assessment
- Regression to the mean — control group (current state baseline)
Why This Matters for PV
Pharmacovigilance professionals already know this framework — it is the foundation of their training. Applying it to software deployment is not a metaphor. It is recognizing that the same epistemological structure governs both domains: you are making an intervention, you need evidence of effectiveness, and you will discover harms you did not anticipate.
The question is whether you discover those harms in a controlled Phase I with 5 sessions, or in an uncontrolled post-market release with thousands of users.