Providers of high‑risk AI systems must keep records that make decisions traceable, yet for agentic systems it remains unclear what those records must contain to enable post‑hoc causal attribution. This paper first introduces an estimator framework and then specifies the conditions under which it fails. We separate the marginal total effect measured by prior work ($TE_{marg}$) from a common‑random‑numbers total effect ($TE_{crn}$) that isolates a step’s own contribution. We also add the natural direct effect ($NDE$) under a pinned downstream and compare the estimators against hand‑derived results. Both estimands fail in the same direction. Under the marginal estimand a causally inert step exhibits the identical total effect as the decisive step on every run of our planted chain – an algebraic identity rather than a coincidence. Under common random numbers the decisive step returns exactly zero on the roughly one‑in‑ten runs where the executing step flips, while its direct effect there is 0.25 and it demonstrably acts; a zero does not certify inactivity, and we note this here rather than in the limitations. We derive a coupling that keeps the direct effect estimable once contexts diverge, providing a closed‑form expression for its degradation. Experiments show that the mediated share used for natural ranking is not a true share under suppression: when direct and mediated paths oppose, the share exceeds one and ranks a suppressed component above a pure mediator. Because the required live pipeline was unavailable during the study window, we publish only the pre‑registration of the discrepancy experiment, not its result. Finally, we contribute a traceability specification to fill the gap opened by the AI Act’s timetable: Article 86’s right to an explanation has applied since 2 August 2026, while Article 12 logging and Annex IV documentation were deferred to 2 December 2027 by Regulation (EU) 2026/1744.
Review This study exposes the limitations of existing causal attribution estimators for agentic systems and proposes a coupling approach together with a traceability specification, offering concrete technical guidance for regulatory compliance.