The causal ladder
A product hypothesis usually begins with molecular identity and proceeds through a sequence: the molecule reaches a tissue, engages a target, changes a downstream biomarker, alters a biological process and ultimately produces a meaningful human outcome. Each arrow is an empirical proposition. Chemistry can establish identity, but identity does not establish exposure; exposure does not establish engagement; and engagement does not establish benefit.
Writing the ladder explicitly prevents category errors. An enzyme assay answers whether material inhibits an enzyme under specified conditions. A pharmacokinetic study answers where and for how long material is present. A controlled human study answers a different question about outcomes in a defined population. These results can support one another without becoming interchangeable.
Negative transfer is scientifically valuable
The PF-04457845 trial is an unusually clear teaching example. Investigators observed the intended molecular pharmacology and associated lipid changes, but the clinical pain endpoint did not show the expected benefit over placebo. That does not make the enzyme measurements false. It demonstrates that the inference from those measurements to that outcome, condition and study design did not hold.
First-in-human STR-324 work provides another reason to keep layers separate. Safety, pharmacokinetics and exploratory pharmacodynamics can be characterized even when a straightforward dose-dependent functional signal is absent. A disciplined development system preserves all of these findings and updates only the propositions they directly address.
Combination systems add interaction terms
When several components are combined, single-agent evidence is necessary but insufficient. Components can alter one another's exposure, metabolism, target occupancy or downstream response. The combination may be additive, sub-additive, antagonistic or synergistic. Those possibilities are not adjectives to select; they are models to compare against an experimentally measured response surface.
A factorial design can estimate main effects and interactions while preserving controls. Concentration matrices, replicate agreement and prespecified selection rules reduce the chance that one attractive data point determines the result. Once a combination is selected, the exact selected material still requires integrated safety and human testing because a computational winner is not a clinical certificate.
Evidence-matched communication
The same hierarchy should govern public communication. Established general biology can be described with source links. A design intention should be labeled as a design intention. Candidate-specific performance requires candidate-specific measurements, and comparative claims require a qualified active-comparator design. This is not weaker communication; it is information with a precise type.
The NOVA-4 distribution compiler enforces that type boundary before publication. Claims, evidence objects and surfaces are serialized together. If a statement requires human product evidence and no matching object exists, the compiler rejects the statement while allowing accurate educational material to proceed. That produces a scalable library without turning scale into evidentiary drift.