Answers are grounded in this paper and cite their sources.
This conceptual paper introduces Separation of Diagnostic Revision Depths, a distinction among four ways to investigate a failed expectation. An inquirer can adjust a value within the current model, reopen a condition previously treated as settled, expand the model’s boundary to include an omitted factor, or revise its account of how one quantity or event depends on another. Each operation holds different things fixed. A value that brings a prediction into agreement with an observation establishes that the model can fit that observation at that value. The assumptions retained in making the adjustment still require their own grounds. The paper assesses a diagnosis by asking what evidence supports the chosen revision and what competing revisions would predict. The four operations can overlap or occur in different orders. A small adjustment may have the strongest support. The paper places this distinction alongside Imre Lakatos’s account of research programs, narrowing the question to what a particular response to a single discrepancy can establish. Two historical cases give the distinction concrete form. To explain the anomalous advance of Mercury’s perihelion, Urbain Le Verrier proposed an unseen planet or smaller bodies, Asaph Hall proposed changing the exponent in the law of gravitation, and Albert Einstein accounted for the advance using a new theory of gravitation. In the Mars case, Johannes Kepler adjusted a circular model to fit Tycho Brahe’s observations of planetary positions at opposition. A check against distances exposed a further difficulty. After correction, some predicted positions missed the observations by as much as eight minutes of arc, beyond their estimated precision of two minutes. This gave Kepler grounds to reopen the assumption of circular motion. He subsequently adopted an ellipse. Part of the Recursive Inquiry and Encounter series within On Human Modeling, the distinction has a prospective application in human–AI evaluation. When an AI forecast underestimates a project’s duration, a reviewer might adjust a task estimate, check whether a promised input arrived, include an omitted approval, or revise which tasks the model treats as able to run in parallel. The project record can help the reviewer decide which change the evidence supports.
No takes yet. Share an insight, caveat, or question.
Alexandra Cary (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: