Closing the Causal Loop: Public and Private Routes in Autoregressive Models

Mechanistic interventions on autoregressive language models are often interpreted as evidence that internal state directly controls later tokens. But generation closes a causal loop: an intermediate output becomes the model's next input, so the same downstream change can arise either from contextual influence that persists after the intermediate token is fixed or from feedback through the model's own generated token. We introduce a six-cell intervention assay and a four-stage causal-validity ladder that separate these public and private routes. On an occupation-completion task adapted from prior latent-planning work, an attribution-selected sparse intervention in Gemma 3 270M with both article and future-noun influence rewrites completions such as a pilot to an aviator through an almost entirely public route: replaying the treated article reproduces the noun, restoring the baseline article rescues it, and article substitution reproduces almost all of the noun-distribution change. Yet target-directed residual interventions still shift later nouns with the article fixed at both 270M and 1B, demonstrating private-route capacity. A neutral-prompt synonym assay at 1B then reveals mediator-relative routing: when two valid alternatives require different articles, target-directed influence is public-dominant; when they share the same article, it is private-dominant. In a matched 14-family extension holding semantic family and source state fixed, crossing the article boundary shifts route dominance toward the public path in 13/14 families. Finally, for the binary article mediator, public effect magnitude factorizes exactly into mediator-policy movement and continuation leverage: near an article decision boundary, crossing the boundary produces a large noun-distribution jump while either fixed-article path moves very little. Future-token control is therefore a closed-loop causal quantity: mechanistic claims must identify not only what internal state changes, but where that influence travels and how generation amplifies it.