When a rule can be read two ways, which example settles it?
Turns competing readings of a pure rule into the boundary example that separates them, a recorded choice and an exported regression test.
Demonstrated
In the worked example, two readings of “within 30 days” differ on 16 of 48 finite inputs, and the two ordinary examples miss both edits.
In practice
Ambiguity hides at boundaries. Ask about the case where the readings differ, not the ordinary one.
Why this question matters
Two people can read the same rule and agree on every ordinary example. The disagreement only shows at the edges, which is where code and agents tend to fail.
What I built
A disagreement matrix across candidate rules, deletion-minimal conflicts, and the chosen reading exported as a machine-readable test, with a Python CLI and a local browser workspace.
A design choice
I compute where the candidate readings disagree, ask about the smallest example that separates them, and export the chosen reading as a regression test. Asking about typical cases would be easier, but in the worked example the two ordinary examples cannot tell the readings apart. Keeping the answer as a test means the clarification survives the conversation it came from.
Where the evidence stops
Human usefulness, productivity and live-model evaluation have not been run. The worked example is authored. It follows ClarifyGPT as a predecessor and is not a replication.