The Engine Is Not the Product
Most agentic architectures are organized around the language model. The model is the asset, the integration point, and the thing the system is named after, and everything else is plumbing arranged to feed it context and carry away its output. Inference control inverts that hierarchy. The model becomes a subordinate component: a proposal generator whose output is evaluated by an authority that sits above it and does not depend on it. Once the substrate decides and the engine only proposes, the engine's identity stops mattering to the guarantee the system makes, and an interchangeable component is, by definition, not where the durable value lives.
This reframing is not rhetorical. It is the direct operational consequence of the separation between proposal and authority, and it is what makes the inference engine not merely swappable for another engine but replaceable by a simpler mechanism altogether.
Why Proposal and Authority Separation Makes Substitution Possible
At every step, the inference engine proposes a transition and the execution substrate evaluates it for admissibility against policy, lineage continuity, entropy bounds, and temporal validity, producing a deterministic admit, reject, or decompose outcome before anything is committed. Because the governance guarantee is supplied by the gate rather than by the engine, the engine does not need to be trusted, and an untrusted component can be exchanged without disturbing the property the system relies on. The substrate is also model-agnostic by construction: anything that can order a set of structured candidates given a structured state can serve as the engine, from a large language model to an embedding scorer to a rule matcher. Swapping one engine for another therefore changes nothing about admissibility, and an anchor can upgrade its engine simply by verifying that the new engine produces valid proposals against the anchor's tests.
The same separation permits a deeper substitution. Because the engine's job at a stable, well-traversed anchor is only to reproduce a decision the lineage has already demonstrated, the engine can be replaced not by another model but by a cheaper mechanism compiled from that lineage. The companion disclosure on maturation-driven engine substitution develops this from the discovery side: a matured anchor substitutes its resident model with a lineage-compiled lookup table, distilled small model, rule set, or similarity scorer, with deterministic escalation for queries outside the substitute's envelope. This article is the inference-control reading of the same fact: it is the proposal-authority separation that makes the substitution safe, because the admissibility guarantee is unchanged regardless of what is doing the proposing.
The Economic Reframe
If the engine is a subordinate, substitutable component, then the economics of the system change in three ways. Vendor lock-in disappears, because the durable value is in the substrate, its index, governance, and lineage, not in a tether to a particular model provider whose pricing and availability the system cannot control. Model upgrades become drop-in, because adopting a more capable engine requires only that it pass the anchor's proposal validity check, with no change to the traversal protocol, the object schema, or the governance. And operating cost declines with maturity rather than scaling with traffic, because the most-traversed parts of the system migrate to compiled substitutes whose marginal cost approaches a table lookup, while the frontier model is reserved for genuinely novel queries.
Implications
The practical consequence is that an organization building on this substrate is not betting on a model. It is building a governed cognitive asset that can absorb whatever model is best at any moment and shed it for the next without re-architecting, and that becomes cheaper to run the more it is used. The model market can churn, prices can move, and new architectures can arrive, and the substrate treats each as a component swap behind a stable admissibility guarantee. The thing that compounds in value is the index and its lineage; the model is a tenant, not the landlord.
Disclosure Scope
The separation of the inference engine as proposal generator from the execution substrate as authority is disclosed in the cognition filing (U.S. Application No. 19/647,395 and its international counterpart) at Section 10.5, and the model-agnostic applicability of the inference engine, including engine replacement by verification against an anchor's proposal tests, at Section 10.11. This article frames those disclosed mechanisms from the inference-control perspective, establishing that the proposal-authority separation is what renders the engine a subordinate, substitutable component, and connects that property to the maturation-driven engine-substitution mechanism disclosed in the companion article. The scope extends to engine classes and substitution policies not enumerated whose admissibility guarantee is supplied by the execution substrate independently of the engine in place.
