What Experiential Makes Auditable
This final chapter answers the opening question: what changes when a model gateway treats production traffic as evidence rather than as anonymous request volume? Experiential's answer is a connected operating loop. Real interactions become immutable evidence; real transitions ground a bounded evaluation world; fit evidence freezes a routing policy; the gateway serves that policy under explicit authority; and completed runtime interactions can become a new, bounded source for optimization. The practical judgment is not whether the project sounds ambitious. It is whether a workload benefits from these boundaries enough to accept their costs, and whether the available evidence supports the claims an operator wants to make.
One loop, several boundaries of trust
Experiential solves one coherent problem across two traditionally separate domains. Offline systems need a defensible way to compare model behavior and choose a policy. Online systems need to enforce who may call which model, what the request is allowed to contain, how much work may be spent, and what happens when a provider fails. If the offline choice has no serving authority, it can be bypassed or applied to the wrong model. If the gateway has no evidence loop, its routing rules remain static guesses or hand-maintained preferences. Experiential connects the domains without pretending that one artifact can answer every question.
The first boundary is provenance. A source is declared, normalized under an explicit convention, and stored with source identity, digests, model evidence, and retained normalization issues. Representative tasks and later artifacts point back to the exact inputs and producing revision. This does not make the traces complete or unbiased. It makes the basis of a later choice inspectable. A routing decision can be traced to a named evidence graph rather than to a mutable export called “latest.”
The second boundary is between observation and prediction. The read-only retrieval artifact accepts real action-to-subsequent-observation transitions. It excludes terminal assistant-only turns and removes excluded lineages before scoring. A grounded world model can use those transitions to predict a next user or environment observation, but its prediction remains a generated rollout. The simulator can persist that rollout for comparison without feeding it back into the retrieval index. This boundary lets Experiential use simulation for coverage while limiting the risk that a model's own guesses become its apparent history.
The third boundary is temporal. Fit evidence may influence the router bank and policy; held-out evidence is opened only after the policy is locked. The workflow can combine observed and simulated cells in one evaluation plan, but it does not let a report quietly tune the policy it is supposed to test. World-model fidelity is measured separately from router fitting because confidence in the simulator and authorization of a routing decision are different judgments. The result is not context-free proof of optimal routing. It is a policy whose evidence timing and source types remain visible.
The fourth boundary is runtime authority. A frozen policy is activated with its bank, feature extractor, pricing snapshot, candidate aliases, and embedder identity. Drift is rejected rather than reconciled by intuition. A request may receive a sticky episode decision, but the state is bounded and process-local. The gateway then authenticates and authorizes before decoding, routing, continuation lookup, or provider work. It checks capabilities, reserves conservative cost before each physical dispatch, and records the attempt and settlement. The online path does not grant the offline policy an exemption from current identity, catalog, or budget rules.
The fifth boundary is feedback. A completed routed interaction can be sealed from a validated runtime journal prefix into an immutable production trace snapshot. That snapshot can feed a bounded SFT dataset and a Tinker configuration only after the source prefix, base model, connection metadata, schedule, and cost ceiling are verified. The possibility of future model optimization therefore closes the loop, but it does not collapse serving into training. A completed response is not automatically a good training example, and a recursively verified training artifact is not automatically evidence of better behavior.