How Firm Is an AI Verdict in Finance? A Firm or Not-Firm Reading for Banks, Insurers and Stock Markets
Year: 2026
Type: article
Abstract: AI programs now give a score for almost every decision in finance: a loan, a card or UPI payment, a name checked against a sanctions list, an insurance claim, a trading signal, a suspicious-trade alert. Each case gets one score from one pass, and nobody is told how firm that score is. This note describes a simple check that sits on top of any such AI program. The data arrives already processed, as a credit report, a payment record or a price feed, and it stays fixed (a machine of the second kind). The controller takes three to four setting vectors of the AI program's own non-learned settings, which the maker has already validated. For each setting value it sets X, a little below X and a little above X, and runs the same processed data each time. If the scores stay together, the verdict is firm. If they spread out, it depends on a setting. A computer-decided shunting action then moves along the validated setting vectors to one where the verdict is firm. Only if none is firm does that case go to the front of the human check list, while all other cases carry on. The learned weights, the data and the final cutoff are never changed. This is different from Lean proof checking, which proves that an action or a statement follows from rules written in advance. The firm score checks no rules. It tells how steady an AI program's own score is when its non-learned settings are moved, and a bank can use both together. The note covers six fields: loans and credit cards, card and UPI fraud, anti-money-laundering and KYC, insurance, trading signals, and exchange surveillance. Each field has a table of related work and how this method differs, and there is a plain explanation of how it differs from Bayesian networks and Monte Carlo dropout. Patent pending. Keywords firm or not-firm verdict, machine of the second kind, non-learned settings, spread of scores, shunting action, controller, setting vectors, Lean proof checking, banking AI, credit scoring, fraud detection, anti-money-laundering, insurance, trading signals, market surveillance, model risk
Author: Vaithyanathan, Prakash
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7230234894
