Emerging Technologies and Intelligent Systems: Theory, Applications and Innovations
Threshold, Not Architecture: Where Visible-Error Suppression Actually Comes From in Consumer-Facing AI
Authors: Somesh Rahul, Raghuram Malpe Pai, Krishna Suman Dara, Mohammed Kamran Syed and Anudeep Rentala
Publishing Date: 15-11-2026
ISBN: 978-81-975670-4-9
Abstract
Machine learning evaluation treats classification errors as symmetric. That inheritance comes from research benchmarks, where a false positive and a false negative cost the same thing, and it survives into products where they don’t. In a consumer application a false negative is usually invisible. The detector declines and the image renders normally, so the shopper never learns an affordance was attempted. A false positive is visible and attributable, and it’s expensive in a way that compounds. We call that asymmetry trust forfeiture. A common architectural response is two-stage gating, where one threshold decides whether to show an affordance at all and a second decides what to rank inside it. We set out to quantify how much that architecture buys. It buys nothing. This is a modelling study, not a measurement of real users. We build an agent-level model of longitudinal trust and sweep it rather than reporting a single point. Three findings. First, the asymmetry itself holds up: a threshold derived by maximising F1 loses to a high threshold in 72 of 72 cells of a four dimensional parameter sweep, median retention gap 47.9 points. Second, two-stage gating contributes nothing once you compare it against a single threshold at a matched abstention rate. It’s worth 0.02 points of retention, and the visible-error rates are identical to three decimal places. The reason is algebraic. Gating on (c ? ????1) ? (c ? ????2) over one score is c ? ????2. We then re-tested the architecture the way it’s actually deployed, with two distinct correlated scores from a detector and a matcher, and it still doesn’t win: across correlations of 0.0, 0.4 and 0.7 its paired advantage over gating on the matcher alone is +0.009, +0.004 and ?0.014 points, and all three 95% intervals cross zero. Third, classical cost-sensitive thresholding, derived properly from the stated cost matrix rather than assumed, outperforms the proposed architecture: 100.0% retention at 0.06% visible error, against 99.7% and 0.28%. Our practical conclusion is narrower than the one we expected to write, and more useful. Spend the engineering effort on deriving the threshold and on choosing which score to gate on. The staging is an implementation convenience, not a source of accuracy.
Keywords
Trust in automation, confidence calibration, deployed machine learning, cost sensitive learning, selective prediction, consumer electronics, human-machine interaction.
Cite as
Somesh Rahul, Raghuram Malpe Pai, Krishna Suman Dara, Mohammed Kamran Syed and Anudeep Rentala, "Threshold, Not Architecture: Where Visible-Error Suppression Actually Comes From in Consumer-Facing AI", In: Harish Sharma and Mukesh Saraswat (eds), Emerging Technologies and Intelligent Systems: Theory, Applications and Innovations, SCRS, India, 2026, pp. 68-87. https://doi.org/10.56155/978-81-975670-4-9-4