Schema revisions must generate new predictions, not just…
When adjusting a schema after disconfirming evidence, require the adjustment to generate new testable predictions rather than merely explaining away the original failure.
When This Fires
- After a schema's prediction fails and you're revising the schema
- When explaining away disconfirming evidence
- During any belief revision where you want to verify the revision is genuinely improving
- When a schema keeps "explaining" failures without preventing them
Common Failure Mode
Revising by adding exceptions: "My schema still works, it just doesn't apply when [the condition that produced the failure]." This protects the schema from the specific failure but generates no new prediction — it's an ad hoc rescue rather than a genuine revision. The test: does the revision predict something it didn't predict before?
The Protocol
After disconfirming evidence triggers schema revision: (1) Revise the schema to account for the failure. (2) From the revised schema, generate at least one new testable prediction: "Given this revision, I now predict [specific outcome] in [specific conditions]." (3) If you can generate a new prediction → progressive revision. The schema is genuinely improving. (4) If you can only explain the past failure without predicting anything new → degenerating revision. The revision is a rescue operation, not an improvement. Go deeper — find the actual flaw rather than patching around it.