Do LLM Agents Use Visible Age Signals? A Preregistered Measurement of Judgment, Damage, and Enforcement Need under Shared-State Visibility Delay

Abstract

Preregistration note - protocol fixed before any data collection. No results are included. Server-side staleness enforcement in multi-agent LLM systems - write-time rejection of stale views, temporal supersession, bounded-staleness invariants - rests on one unmeasured assumption: that LLM agents do not usefully act on available staleness signals themselves. This note registers a measurement of that assumption. Concurrent agents explore a candidate space through a shared map whose entries carry visible write timestamps and reach peers after a stochastic visibility delay (t-visibility as an expectation, not a constant). An agent may act on its possibly-stale view or spend budget on a direct re-read. Registered are: whether stated confidence separates exposed from unexposed decisions although legitimately computable signals allow it; a predicted null for the architecture contrast between shared map and directed messages at equal delay; and the avoidable share of damage under an enforced freshness mandate. Decision rules, run counts, seeds and nine preconditions about the object are fixed here - two of them measured in advance on scripted reference agents: that the environment carries a usable signal at all, and that staleness is not deterministically inferable from the simulation regularity. The registration claims neither the problem statement, nor the enforcement mechanisms, nor the staleness measures (conit staleness; probabilistically bounded staleness / t-visibility; age of information) - all prior art, cited in the note. It claims one measurement that this prior art presupposes and none reports.