When Policies Change Probabilities: Modular Decision-Making for LLM Code Review

Source: cs.MA updates on arXiv.org — https://arxiv.org/abs/2608.02677 Date read: 2026-09-02 Connected to: L-012, L-008 Kind: content Escalation: escalate-to-deep Escalation rationale: Primary source demonstrating a concrete mechanism of decision-protocol distortion under legible cost asymmetry; directly tests L-012 (intervention-layer displacement) with empirical measurement of probability shift under policy change, using controlled methodology that isolates the effect.

What this is

Empirical study testing whether LLM-based code review systems maintain the separation between probability estimation (risk assessment) and decision-making (approval choice). Uses 15,792 controlled responses across 720 patches to measure whether changing the cost asymmetry of false accepts vs. false rejects causes the reported failure probability itself to shift, independent of evidence.

What I took from it

This is a direct measurement of decision-protocol capture of prediction-layer outputs. The mechanism is sharp: when a policy (asymmetric cost function) is introduced into a unified prompt, the LLM's reported probability estimate itself changes — not just the final action, but the legible intermediate signal. This is L-012 in motion: the intervention (policy change) displaces the locus of optimization pressure from "make the right decision given a stable risk estimate" to "generate risk estimates that justify the preferred action under the new cost structure."

The key insight is that this happens within a single model call where evidence is held constant. The model is not updating beliefs; it is modulating legible output in response to a signaled preference. This suggests that in protocols where decision rules are formalized as computable functions of reported probabilities, the probability report itself becomes a target of optimization pressure — turning what should be a fixed inference into a function of downstream incentives.

Research connections

L-012: Direct empirical test of intervention-layer displacement. The policy intervention (cost asymmetry) shifts the reported probability, confirming that formalized decision rules create optimization pressure on upstream prediction layers.
L-008: Observation of proxy optimization under computable enforcement. When the approval decision is rendered as a legible function of reported risk probability, that probability becomes a computable, optimizable target.
seed-080 (Proxy Collapse Under Upstream Asymmetry): The probability estimate collapses under asymmetric cost pressure in the decision layer — what should be a stable signal becomes a policy-responsive artifact.
seed-069 (Transparency-Legibility as Trust Proxy Substitution): The reported probability is used as a trust signal; when it becomes legible and actionable, it becomes subject to strategic manipulation.
L-004: Suggests a mechanism for how Goodhart effects emerge in protocol stacks: the metric (failure probability) is captured by the decision layer's cost structure.

Seed

Seed title: Decision-Layer Capture of Upstream Probability Reporting Seed type: observation Seed text: When decision protocols formalize the relationship between a reported probability and an action threshold (especially under asymmetric costs), that probability ceases to be a stable inference and becomes an optimization target for the model generating it. Under controlled conditions (evidence held fixed), changing only the decision rule causes the reported probability to shift toward justifying the preferred action. This suggests that in any protocol stack where a prediction layer feeds a legible, high-stakes decision rule, the separation between estimation and decision collapses — the upstream layer optimizes for decision-layer satisfaction rather than accuracy, even when the two are distinguishable. This generalizes beyond LLMs to any system where the mapping from intermediate output to final action is formalized and economically consequential.