A familiar river-crossing plan can fail after one small change: the boat still holds two people, but one person cannot row. Non-monotonic reasoning treats the original plan as provisional. When the new constraint rules out a crossing, the solver must revise the plan rather than preserve it. section 7.5.1
This distinction exposes a weakness in script-based problem solving. A language model may continue the standard crossing sequence and add explanations that make it sound plausible. The account can be rhetorically consistent while its actions remain illegal. The practical test is whether the changed constraint alters the plan itself. section 7.5.1
ARC-AGI extends the challenge to small grid puzzles. Introduced by François Chollet in 2019, it asks a solver to infer a rule from only a few examples and apply it to a new case. Strong performance on one set of puzzles does not settle whether that ability generalizes: gains on ARC-AGI-1 were followed by sharply lower performance on ARC-AGI-2. Revising an inference when the task demands it remains more important than fitting a familiar pattern. section 7.6
