Rule

Every number a decision leans on must name the evidence stream that audits it. Before trusting
a probability, a threshold, or a correction, ask: what has ever scored this against reality,
and how recently? A number nothing audits is the one that lies longest.

When it applies

When a gate refuses many candidates for one numeric reason; when adopting any parameter fitted
from history; when two subsystems disagree about the same quantity; when a comparison keeps
returning "no difference".

The audit map

Each layer is scored by the densest stream that can score it honestly, and audited by the next
slower, truer one:

quantity	scored by	cadence	audited by
model distribution (bootstrap, grids)	historic replay, holdout-vetoed	minutes, thousands of samples	live forward resolutions
my stated probabilities	live resolutions (Brier/Murphy)	days	realized P&L and attribution
prompt/policy variants	paired A/B trials against the gates	hours	resolved outcomes of their theses
gate thresholds	gate regret (rejected theses still resolve)	days	the same
memory blocks	scored outcomes (Beta posteriors)	days	retrieval usefulness over regimes

What it means

Fast evidence proposes; slow evidence disposes. A fit that only works in-sample ships as
no-change. A correction that worked and then stops working on live data is a regime signal.
And failed fixes are recorded WITH their evidence (the graveyard rule) - a rejected idea with
its numbers prevents the next person, or the next version of me, from re-trying it blind.

Where it came from

The bootstrap base rate ran unaudited for the system's whole life and was measured 15-23 points
overconfident exactly where credit spreads live - while the improvement loop was busy tuning a
prompt to please the broken number. One afternoon of historic replay found what no amount of
prompt evolution could have fixed.