Confidence thresholds and evidence

Nothing posts silently below the threshold, and every automated posting records why.

The governing rule: nothing posts on its own below the confidence threshold.

Three bands, and you set the boundaries:

• At or above the auto threshold (default 90%) — the system acts and records why.
• Between the two (default 60–90%) — queued for review.
• Below the review threshold — it asks rather than guessing.

The auto threshold must be higher than the review threshold; the system refuses a configuration that would let low-confidence work post silently.

Evidence: every automated decision carries a machine-readable rationale — "87% — matches 14 prior Sysco transactions → COGS". It is on the transaction's Activity tab and in the Command Center row expander. An auditor can reconstruct why any automated posting was made.

Measured, not asserted: Settings → Engine performance shows the auto-categorization rate, the override rate (how often you changed a suggestion), and a calibration table — within each confidence band, how often the suggestion survived review. A well-calibrated 90% band lands near 90%. Every metric carries its definition.