confidence-calibration

Run the SteerConf three-pass protocol: generate conservative, neutral, and optimistic versions of the analysis, then measure agreement across passes. The consistency level maps to a confidence score from 0.00 to 1.00, combined with data quality and source diversity factors.

SteerConf is a prompt-only calibration technique that halves calibration error without infrastructure. It generates three analysis passes with different assumption stances and derives confidence from their agreement, weighted by verified datapoints and source diversity.

The score combines three weighted components: base agreement across passes (60%), ratio of verified datapoints (25%), and source diversity (15%). Scores above 0.80 indicate high consistency, while scores below 0.20 signal conflicting interpretations requiring human review.

Yes, the protocol integrates with credit memos, deal screening, CRE underwriting, and financial models. Each domain has specific calibration drivers, such as independent data source counts for credit memos and T12 quality for CRE underwriting.

Scores in the 0.20-0.49 range indicate high uncertainty where the three passes diverge, and scores below 0.20 indicate conflicting conservative and optimistic conclusions. Both cases flag insufficient data or fundamental ambiguity requiring human judgment.