Observation-Relative Monitorability of Bounded Temporal Implication

A material conditional "P implies Q" has a timeless truth value, but knowing whether it holds does not: every physical or computational check of the conditional is an act of observation, carried out at finite resolution, with each event separated from its observation by delay, drift, and jitter. This produces a paradox of observed implication -- an internally sound monitor can contradict the timeless truth value in both directions (a spurious falsifier, or a phantom trigger), not because the instrument is broken but because observation has a class. We make this precise for the temporal implication G(p -> F[<=k] q) and establish a single headline result: the monitorability of the bounded implication is joint in the pair (formula, observation model) -- the observation model being the sampling regime through which the signals are read (when the monitor looks, and the delay, drift, jitter, and finite resolution between an event and its observation) -- and is not a property of the formula alone. A change of observation model alone can move the property across the safety / cosafety / liveness / coliveness classification. The observation space partitions into a machine-checked taxonomy of eight classes; a monotone ladder of declared observation theories grades which classes become True-attributable; and internally sound observers that disagree hold different rungs rather than contradicting each other. Preprint; a version has been prepared for submission. Engineering companion: "Controlling the Observation Class of a Runtime Monitor" (co-deposited). --- Version note (this version, v3): this version incorporates a citation-integrity review of the manuscript: inaccurate or unsupported attributions were corrected (including the placement of the paper's timing regimes within the Dwork-Lynch-Stockmeyer synchrony taxonomy) and bibliography entries were cleaned up. It also corrects the threshold-detectability condition (CC3) of Definition 4.2, whose inequality was stated in the wrong direction relative to the paper's own failure mode F5 (noise-induced false positive): the condition now requires the decision threshold to lie strictly above the sensor's noise floor, and Case 4 of the proof of Theorem 4.3 was updated accordingly. The statement of Theorem 4.3 is unchanged, and no new result is introduced.