Evaluation

Phase 09 graded one thing: when the risk engine cuts a trade because of large past losses, does the explainer name the real cause of those losses using only headlines it actually cites? It does not grade returns, and it claims no trading edge.

Why a pre-registration

The plan, the baseline and the pass rule were committed before any label existed or any explanation was judged (phase09-plan.md). Three deviations were forced along the way and each is logged in the engineering log before the step it changed: labels written by Claude instead of by hand, a test set rebuilt on headline dates, and the judge check scored by Claude.

Regression harness

The test set (eval/golden_days.json), the labels and a 1,415-headline fixture are committed. A CI test re-runs retrieval over the fixture and fails if the rankings the labelled test set was built on ever change, until the test set is rebuilt on purpose. The judged part is not re-run in CI because it spends money.