Daily pulse — evals

~2 min read · deterministic analyzer, replayable golden set

The pulse analyzer is deterministic, which means the testing question is "does the analyzer surface the same signals on the same window two days in a row?" not "did the model say something plausible?" The 15-case golden corpus covers healthy windows, cash-floor breaches, corridor spread drift, vendor concentration, runway compression, and cadence anomalies, plus 2 refusal cases on malformed windows. Every case fixes the window and the decision-log slice, so any change to the analyzer is visible as a diff in the surfaced signals.

Coverage

Daily pulse · 15 casesHealthy / cash-floor / corridor / vendor-concentration / runway / cadence cases plus 2 refusals on malformed windows.

15runtime golden cases

6signal families covered

2refusal cases · malformed window

replayableold pulse against today's analyzer

Replay is the property that makes the pulse defensible. Because the brief is a decision row, an old pulse can be run against today's analyzer to see whether the recommendations would have changed — that catches signal-threshold drift the same way evals catch prompt drift on auto-tagging. Reproduce with pnpm eval -- --skill daily-pulse; reports write to evals/results/<timestamp>.md.

Read next

• Overview — the analyzer architecture and signal taxonomy.
• Worked example — the 8–15 May 2026 pulse, end to end.
• Eval harness — the cross-skill harness (in Operations).