Interpret a result

Start with the route or verdict, then read the rationale and provenance. Do not start with the Sharpe ratio.

A disciplined reading sequence

1. Confirm source and claim fidelity.
2. Check whether the result is a verdict or routing state.
3. Inspect synthetic, venue, frequency, and survivorship flags.
4. Read power and minimum detectable effect.
5. Verify search cohort, denominator, and family.
6. Inspect fold, regime, CPCV, bootstrap, parameter, and tail diagnostics.
7. Distinguish modeled from measured costs and capacity.
8. Check confirmation and holdout status.
9. Read the final rationale.

Underpowered is not dead

An underpowered result should include a resolution estimate, such as more out-of-sample observations or greater cross-sectional breadth. Use that estimate to decide whether further data collection is worth the cost.

Needs review is not watch

needs_review is an integrity stop. It can accompany spectacular metrics when the implementation, binding, or scale is implausible. Never interpret it as a weak positive verdict.

Compare within an epoch

Metrics from different engine versions, data fingerprints, or corpus epochs may not be directly comparable. Use receipts and migration lineage when comparing historical and current results.