A radical argument for evidence-first scientific infrastructure. The thesis attacks inherited bottlenecks, not the accomplishments or fallibility of individual researchers.
Scientific methods were built and refined for human institutions with finite memory, attention, coordination and time. Their principles can remain valuable without retaining every institutional procedure. A Wikipedia article explaining a practice is not necessarily claiming that the practice is optimal; this site criticizes the design constraints of the enterprise rather than inventing such a claim.
The mathematics analogy is productive but not complete: a theorem can be checked relative to axioms, whereas a claim about nature must survive measurement, calibration, causal identification and independent contact with the world.
The unit of scientific progress should be a versioned claim plus its evidence, assumptions, executable derivations and unresolved challenges. Papers can continue to explain results, but no longer constitute the only way results enter the scientific record. Negative outcomes, failed apparatus, calibration changes and corrected premises must be able to revise the graph directly. Each dependent conclusion becomes queryable when upstream support changes.
AI makes exploration cheap. Restricting curiosity to preregistered hypotheses would sacrifice a genuine capability, but calling results confirmatory after extensive reuse of the same data would industrialize p-hacking. Every analysis path and data exposure should be recorded. Confirmation must rely on valid selective inference, prospectively specified tests, or truly independent observations. The cost of generating a hypothesis is not the strength of evidence for it.
A theory should face a coordinated search for counterexamples, rival mechanisms, failed auxiliary assumptions, simulation artifacts and out-of-distribution conditions. The most valuable challenge is not the loudest dissent: it is the challenge that would separate live competing explanations. Independent agents and instruments should be intentionally diverse so that repeated conclusions are not just repeated training-data mistakes.
Knowledge systems need not collapse compatible and incompatible explanations into a single official story. Competing models can coexist with clearly exposed domains, predictions, uncertainties and dependency histories. Consensus remains a useful sociological and decision signal, but it must not substitute for the evidence that produced it. Forking is disciplined when rival models make distinct, auditable claims.
A formal proof can certify a derivation relative to axioms and a precise encoding; it cannot establish that a biological treatment works. Replication of a measurement does not validate the causal story attached to it. An epistemic system must use different certificates for logical proof, numerical reproduction, calibrated measurement, causal identification, generalization and ethical authorization. Each certificate has its own failure modes.
Editorial selection, acceptance and rejection can determine which scientific narratives reach readers. They should not determine whether experimental evidence exists. Investigation registries, immutable evidence records and automatically generated methodological reports can make all outcomes accessible. Journals may retain their best functions—explanation, synthesis, criticism and pedagogy—without monopolizing scientific existence.
No set of methods should be exempt from comparison simply because scientists inherited it. Research infrastructures should measure their own error-correction speed, predictive calibration, replication rates, coverage and costs. Competing experiment-selection algorithms, statistical policies and auditing systems should be benchmarked under controlled conditions. A scientific method earns confidence through demonstrated performance.
Institutional legitimacy governs research permissions and social responsibilities. It cannot make a mathematical statement true or an empirical claim reliable by decree. Equally, a machine's computational prowess cannot authorize experiments on people or decide acceptable risks for a society. Technical knowledge claims, legal permission, ethical legitimacy and political authority must be represented separately.
A powerful AI can search over experimental designs, models and formal languages, but its greatest contribution may be discovering that a field is asking a question in the wrong variables. Competing ontologies should be evaluated through their explanatory reach and discriminating consequences. Novelty alone establishes nothing. The point is to remove unnecessary limits on hypothesis space without removing evidence requirements.
When evidence is insufficient, the infrastructure must be capable of saying so indefinitely. A model can be predictive yet causally unidentified; a theorem can be formally valid yet misaligned with the intended informal problem. Unsupported generalizations, unverifiable claims and contested definitions should remain visible and queryable. An AI-native scientific system should make this honesty easier—not optimize it away.
Human beings retain legitimate authority over ethical permissions, rights, funding priorities and acceptable risk. Machine verification cannot independently establish which human values should govern experimentation. Scientific consensus remains useful as a practical summary, but is not identical to empirical warrant. A powerful AI that repeatedly invents attractive mistakes is not scientific progress.
We do not propose turning every research field into mathematics. We propose representing every field with appropriate evidence types, explicit assumptions and independently testable consequences.