Adaptive decision support under frozen rules: guarantees, failure modes, and component admission

Preprint, not peer reviewed. Submitted to PLOS ONE. Like fault-tolerant control, adaptive decision support must keep named rules fixed while it learns. For one decision-support system that ranked options with fixed rules and adjusted the ranking from reported outcomes, we asked what a bound on those adjustments guarantees and costs, and whether more complex components improve decisions. The primary estimand was the difference in expected net utility per decision (the modeled chance of success or the declared stopping value, minus question costs) between the full system and a simpler system with a simple learner under the same bound. For rule-based scores that do not depend on learned values, we proved that, if learning can move each option’s score by at most B, it can reorder two options only when their rule-based scores differ by at most 2B. For larger gaps, the rules’ order cannot change: fabricated reports cannot reverse it, and true reports cannot correct it when it is wrong. Simulations showed the first effect: under fabricated success reports for an option more than 2B below the top, no bounded learner ranked it first, whereas an unbounded learner chose it in 814 of 1,500 decisions on average. The full system did not improve decisions: it scored lower before and after a revision (afterward, −0.0087; 95% confidence interval −0.0136 to −0.0039) and showed no established advantage over a pooled learner when one action was best in every context. Known-answer simulations showed that the learning rule as released changed no decision, although all 76 existing tests passed. An admission rule based on added value almost always rejected both of two nearly identical useful components. In this architecture, protection against fabricated feedback comes from the bound, which simple learners share, so more complex components must outperform them under that bound. All evidence was computational; no physical system was tested.

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-13
DOI
https://doi.org/10.5281/zenodo.24207064
Primary Topic
Decision-Making and Behavioral Economics
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Adaptive decision support under frozen rules: guarantees, failure modes, and component admission

Patrick Iben
Zenodo (CERN European Organization for Nuclear Research)
Decision-Making and Behavioral Economics
preprint

Adaptive decision support under frozen rules: guarantees, failure modes, and component admission

Patrick Iben
preprint en

Abstract

Preprint, not peer reviewed. Submitted to PLOS ONE. Like fault-tolerant control, adaptive decision support must keep named rules fixed while it learns. For one decision-support system that ranked options with fixed rules and adjusted the ranking from reported outcomes, we asked what a bound on those adjustments guarantees and costs, and whether more complex components improve decisions. The primary estimand was the difference in expected net utility per decision (the modeled chance of success or the declared stopping value, minus question costs) between the full system and a simpler system with a simple learner under the same bound. For rule-based scores that do not depend on learned values, we proved that, if learning can move each option’s score by at most B, it can reorder two options only when their rule-based scores differ by at most 2B. For larger gaps, the rules’ order cannot change: fabricated reports cannot reverse it, and true reports cannot correct it when it is wrong. Simulations showed the first effect: under fabricated success reports for an option more than 2B below the top, no bounded learner ranked it first, whereas an unbounded learner chose it in 814 of 1,500 decisions on average. The full system did not improve decisions: it scored lower before and after a revision (afterward, −0.0087; 95% confidence interval −0.0136 to −0.0039) and showed no established advantage over a pooled learner when one action was best in every context. Known-answer simulations showed that the learning rule as released changed no decision, although all 76 existing tests passed. An admission rule based on added value almost always rejected both of two nearly identical useful components. In this architecture, protection against fabricated feedback comes from the bound, which simple learners share, so more complex components must outperform them under that bound. All evidence was computational; no physical system was tested.

Zenodo (CERN European Organization for Nuclear Research)
Decision-Making and Behavioral Economics
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon