When Evidence Becomes a Veto: Procedural Objectives and Decision-Authority Inflation in LLM Agents

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

An LLM agent can correctly recognize that an intermediate observation is useful and still use it for the wrong decision. We study decision-authority inflation: the autonomous promotion of a non-dispositive signal into a rule for whether justified downstream work should continue. Controlled planning tasks separate a legitimate blocker L, which truly can stop the target, from an observation N, which may inform interpretation but should not veto continuation. Qwen3-8B frequently lets unfavorable N terminate the target under implicit decision context. Clarifying downstream value, feasibility, or decision authority sharply reduces this pruning while the model continues to branch on N, separating information use from action authority. A length-matched factorial isolates a strong interaction: in the key L+,N− cell, evaluative future evidence yields 50% STOP, whereas settled evaluative evidence yields 0% and opaque future evidence 2%; the primary graded difference-in-differences is +12.54 STOP-vs-PROCEED logits, or 1.48 pooled within-scene standard deviations. Internally, the future-by-evaluative conjunction forms a stable low-rank representation before final decisions diverge, then becomes increasingly coupled to a generic STOP computation. Projection removal can strongly change false stopping, but matched same-operation controls show that this representation is not a unique causal authority direction. Additional experiments show domain and measurement sensitivity, judge-dependent preference for overgated plans, a preference-training shortcut that deletes the evidence branch instead of calibrating it, and preliminary spontaneous re-verification of already-settled states. The resulting picture is not a single “overcautious” circuit. It is a broader failure in how agents assign procedural roles to represented information.