Unexplainable decisions should trigger a conservative default in high-risk settings
Aliases: explanation gate · fail-safe default · unexplainable decision
What it is
A conservative default under unexplainability blocks normal authority for a high-consequence action when input provenance, rule basis, or audit record is insufficient, choosing stop, slow, hold, or authorisation. Explainability functions as an evidence-completeness gate rather than post-hoc decoration.
Why it happens
Missing explanation can reveal absent perception, rule conflict, operation outside validation, or broken logging. Continuing treats unknown risk as acceptable. Conservative behaviour limits reachable consequences and buys review time, although the correct default depends on dynamics because an abrupt stop can also harm.
Studying it
Missing input, version mismatch, rule conflict, and logger failure can compare continue, stop, degrade, and confirm on weighted consequence, false trigger, recovery, and comprehension. Completion alone is inadequate. The gate must reject empty plausible text.
Where it stops holding
Explainable is not necessarily correct, and a model lacking a simple rule is not necessarily unsafe. Thresholds follow consequence. A road vehicle may need lane-holding deceleration rather than abrupt stop; life-support assistance may need continuity. Defaults require scenario risk analysis.
Applying it
- Define mandatory evidence fields and consistency checks before consequential commitment; missing evidence blocks normal execution.
- Choose stop, slow, hold, or handoff from current dynamics and environment, exposing trigger and recovery.
- Inject faults to verify canned explanations cannot bypass the gate, measuring false triggers, exposure time, and safe recovery.
Related
- Same group: X3.07.1 Critical decisions must be traceable to their inputs and rules · X3.07.2 Post-hoc and real-time explanations have different timing requirements · X3.07.3 Overly technical explanations are equivalent to no explanation for field personnel
- Adjacent: X3.08 Communicating uncertainty · X5.03 Emergency stop
- Search terms:
conservative default·fail-safe·explanation gate