Causation Without a Global Manipulator: Record-Mediated Difference-Making, Internal Intervention, and Counterfactual Dependence in a Consistency-Constrained Information History

The first companion paper established an invariant causal partial order without a global clock, but an order is not causation: that one event lies in another's causal past does not settle that it made a difference to it or produced it. This paper supplies the causal-specification layer that the series synthesis left open, and partially discharges the record-mediated-causal-access subproblem, while declining the reduction of causation to records. It specifies type-level difference-making and productive mediation; prevention, enabling, omission, and token-level actual causation, which require further contingency or directional specification, are scoped out. The discipline is to separate things routinely conflated: the world-side causal relation, its model-side representation, an embedded observer's record-mediated estimate, the observer's attribution, and the observer's entitlement. A causal claim is made adjudicable by a declared, kind-indexed specification, rather than settled from outside the history. The signature commitment is that no physical intervention realization is performed from outside the history: every manipulation and controlled comparison is itself a history-internal physical process with provenance, resource costs, failure modes, and records where traces are retained. Three contributions follow. The formal intervention transformation - graph surgery on a model - is separated from its physical realization by an internal controller, and a directional realization relation measures how far the controller realizes the formal operation, using three quantitative tolerances - target-setting, decoupling, and undeclared side effects - together with a structural path-admissibility condition. The production lineage is separated from the record lineage, and difference-making is represented by a model-side interventional kernel, which does not constitute the world-side relation and which an observer estimates only through records. The verdict is a typed package - applicability, model adequacy, a model-relative world-support profile, an observer-side epistemic profile, specification admissibility, and, where declared, aggregated support and entitlement - rather than a single magnitude. The results are conditional: relative to a declared specification, the paper supplies a discipline for adjudicating causal claims, not a reductive metaphysics of causation and not a proof that the world possesses a unique causal structure.