A failure explanation becomes search control only when it changes a later branch decision
Type: kb/types/note.md · Tags: foundations, learning-theory, deploy-time-learning, self-improving-systems
An explanation produced after a failed or disappointing branch may be correct yet inert. It becomes part of search control only when it is retained and later changes a branch decision: what is tried, how the idea is scoped or prioritized, which probe is selected, or when the branch is continued, suspended, or abandoned.
Rejecting the original idea is not required. An explanation can preserve it by narrowing its applicability or identifying a different test. That counts as an update when the narrower account predicts a different later choice.
An explanation that leaves every later branch decision unchanged has not yet improved search control. It may still be useful commentary, but it does not separate learning from rationalization.
Scope
- The later choice may occur in the same episode or a later one, provided the retained explanation is part of the causal path.
- Changing wording or confidence without changing a branch decision does not satisfy the claim.
Relevant Notes:
- Lightweight search control allocates further search without licensing adoption — grounds: identifies the branch decisions an explanation can change
- Theory-mediated self-improvement needs interpretation and retention — extends: supplies the stronger requirement that a retained theory-state change affect later operation
- Knowledge storage does not imply contextual activation — exemplifies: treats a retained failure explanation as a concrete case of stored knowledge becoming active only when it changes later behavior
- Backtracking keeps lightweight search control provisional — mechanism: restores an earlier usable state so a retained explanation can redirect search instead of remaining commentary
- Citing retained theory at the decision point is a mediation trace — enables: supplies a checkable record that the retained explanation entered the causal path to the later branch decision
- A search controller is tested by what it brings to stronger evaluation — extends: distinguishes evidence that an explanation changed routing from evidence that the resulting control improves search