Interpret Results

Recommended reading order

1. failed_checks
2. not_evaluated_checks
3. gate_outcomes
4. evidence_tier

This avoids over-indexing on the final label before understanding the causal reason for it.

Evidence tiers in practice

• cross_model_confirmed: strongest cross-model evidence pathway
• single_model_confirmed: accepted within one model context
• causal_plus_robustness: meaningful signal, but incomplete acceptance evidence
• causal_tested_unstable: causal evidence exists but instability remains
• suggestive: weak/incomplete support
• rejected: evidence did not satisfy the contract

Failure decomposition workflow

For a bundle set:

1. Count failed gates across all claims.
2. Group failures by task/model/lane/provider.
3. Separate missing-evidence failures from contradiction failures.
4. Prioritize experiments that target the highest-concentration failure modes.

Interpreting not_evaluated

not_evaluated often indicates missing slice or unavailable transfer evidence, not necessarily negative evidence.

Treat it as a data-completeness signal.

Resubmission strategy

• Fix structural incompleteness first (slices, schema, manifest integrity).
• Then target robustness/method sensitivity failure clusters.
• Re-run deterministic review before resubmitting.