System use provides evidence of theory fit and causal usefulness, not independent warrant

Type: kb/types/note.md · Tags: learning-theory, deploy-time-learning, evaluation

Putting a claim to work in a live system creates evidence that static inspection cannot supply. If the claim changes proposal, diagnosis, branch allocation, modification, recovery, transfer, or later operation, and those changes survive consequential use, the result bears on whether the claim is integrated into the working theory and causally useful there. It does not by itself establish that the claim is independently warranted.

The distinction follows because a claim's warrant does not determine its fit in a working theory. System use primarily exposes the relational side. A claim that sounds coherent but never changes behavior has weak evidence of operative fit. A claim that changes search, helps distinguish rival repairs, reduces later repair, or transfers to a new case has stronger evidence that it participates usefully in the theory's operation.

But the live system can reward its own misconceptions. Earlier implementation choices may already assume the claim; neighboring retained claims may make it appear indispensable; an evaluator may share the same error; or the current task distribution may never expose the claim's bad scope. Successful use can therefore be self-confirming.

Independent factual, formal, source, and scope evidence must remain able to overturn the claim. Depending on the claim, that may include empirical observation, derivation, source review, contradiction tests, preregistered predictions, held-out demands, or transfer outside the cases that shaped the current theory. These checks test warrant at the strength their own evidence licenses rather than inheriting authority from the system's success.

The evidence can run in both directions

System use is not merely positive evidence. A claim may fit poorly when it repeatedly fails to affect decisions, increases search or repair cost, breaks under later demands, or loses to a rival formulation under a controlled comparison. Those outcomes can motivate rescoping or removal even when the claim remains independently warranted.

Conversely, system failure can expose a warrant problem without identifying one automatically. A failed modification may reveal that the guiding claim was false, that its scope was wrong, that another premise was missing, that the interpreter misapplied it, or that the evaluator accepted the wrong candidate. Read-back must assign the failure at the granularity the evidence supports rather than treating every bad outcome as a refutation of the theory.

The useful division is therefore not "system evidence versus real evidence." System-building consequences are real evidence for integration and causal usefulness. They become evidence for truth, validity, or warranted scope only through an additional argument connecting the observed consequence to that claim.

Scope

  • The claim concerns evidence produced by operating or modifying a system with a working theory. It does not require that the system itself perform the independent warrant assessment.
  • Causal usefulness is relative to an objective, task distribution, horizon, and comparison. A useful claim under one regime need not fit another.
  • Independent warrant can itself depend on operational evidence when the claim predicts system behavior. The requirement is an explicit evidential connection, not separation by data source.
  • The note does not define a complete oracle for theory fit. It only separates what system-use evidence establishes directly from what requires further warrant.

Relevant Notes: