Failure Mode Taxonomy

Failure exports use a small taxonomy so invalid estimates can be compared across estimators and suites without over-interpreting estimator-specific error strings.

Code	Meaning	Typical source	Public interpretation
insufficient_signal	The estimator did not have enough usable data or scale support.	insufficient_signal_for_* failure reasons.	Treat as a validity failure, not as evidence against LRD in the record.
estimator_exception	The estimator raised an exception that was captured as an invalid estimate.	exception:<type>:<message> failure reasons.	Treat as implementation or numerical fragility that needs inspection.
missing_uncertainty	A point estimate exists, but requested intervals were unavailable.	Failure-summary exports when coverage or CI metrics are requested.	Accuracy metrics may still be usable; coverage and CI-width metrics are incomplete.
invalid_fit	The estimator returned no admissible point estimate for another reason.	Unknown or uncategorised failure reason.	Treat as an invalid estimate and inspect raw diagnostics before drawing conclusions.

The taxonomy is intentionally coarse. Estimator-specific failure_reason values remain available in raw estimate exports for debugging, while summary tables should use these categories when grouping failure rates.

Reporting Rules

• Do not convert an invalid estimate into a numeric score by imputation.
• Report validity or invalid-rate metrics beside accuracy and robustness metrics.
• Separate missing uncertainty from invalid point estimates whenever coverage metrics are included.
• For observational suites, failure rates describe estimator behaviour only; they do not validate or falsify the underlying data source.