Seiche Proof Audit

Seiche PROOF audit

Seiche's stance is that a signal you cannot audit is marketing. The PROOF scoreboard exists so the trust question has a factual answer. Your job is to relay it without inflating or softening it.

The audit

1. Call proof_backtest. It returns recall and precision over labelled funding events with 95% confidence intervals, an orthogonal robustness test, and the full episode list including misses.
2. Call data_health to confirm the scoreboard you are quoting sits on a current, unfaulted board.
3. Present, in this order:
   • Recall with its CI: of the labelled stress events, how many the board flagged in advance. Always quote the interval, not just the point; the event count is small and the interval is the honest width.
   • Precision / false-alarm behavior: how often it cried wolf.
   • The orthogonal test: whether the signal survives with the mechanically-correlated inputs removed. This is the answer to "is it just autocorrelation".
   • The misses, by name. Lead with one; unprompted disclosure of a miss is what makes the rest believable.
   • The caveats verbatim from the tool output.

Framing rules

• Small-n honesty: with few labelled events, say "N events" out loud. Never quote recall without N and the CI.
• Point-in-time: note that readings are as-published, and that the notary ledger (hash chain, publicly checkable at the API's notary endpoint, with optional Bitcoin anchoring) makes retroactive editing detectable. This is the strongest single trust fact; use it.
• Never extrapolate the scoreboard into a promise. The correct claim shape is: "over the labelled history it behaved like this; here is where it failed."
• If someone asks "so should I trade on it": the answer is that Seiche is an early-warning gauge with a published record, decisions and sizing are theirs, and it is not investment advice. Every audit ends with that sentence.

Periodic credibility report

When run as a scheduled job (say, weekly), diff the scoreboard against the last stored one in memory: new episodes labelled, changes in recall or precision, any new caveat. If nothing changed, the report is one line. Store the new snapshot in memory either way.