Confidence-Gated Inference

Mechanism

The confidence gate sits at the boundary between input acceptance and inference production. Every input admitted to the runtime carries a confidence vector decomposed by source: a sensor-pedigree component, a freshness component, a corroboration component, and a chain-of-custody component. The gate consumes the vector together with the inference request and looks up, in the governance class's threshold tier table, the minimum confidence each component must reach for the requested inference to proceed. The lookup is deterministic and content-addressed; the tier table is declared in the policy reference and changes only when the policy hash changes.

When every component of the input's confidence vector meets or exceeds its tier-bound minimum, the inference is admitted to the resolution layer and proceeds under the anchored-resolution mechanism described separately. When any component falls below its minimum, the gate closes. A closed gate does not produce a low-confidence output, a hedged answer, or a refusal; it produces a structured inquiry. The inquiry is a typed record naming the failing component, the magnitude of the shortfall, the tier under which the gate was evaluated, and the specific evidence that would, if supplied, allow the gate to open. The inquiry is routed back to the requesting principal through the same identity-layer binding under which the original request arrived.

The architectural commitment is that an inquiry is not a soft refusal. It is a request for additional input, and the runtime is prepared to resume the original inference when that input arrives, binding the resumed inference to the original request through the lineage. This means an inquiry has the same auditability properties as an inference: it carries a sealed substrate binding, references the policy hash under which it was emitted, and can be reproduced by an auditor reconstructing the runtime's state at the moment of gating.

The gate is not a single threshold but a per-tier table because different inference classes carry different consequence profiles. A query that produces an informational answer has a lower threshold than one that produces a recommendation, which has a lower threshold than one that produces a commitment to act. The patent specifies that the tier mapping is declared by the governance class, not learned, so that licensees can certify compliance by reviewing the table rather than by empirically probing the runtime. The mapping must be monotonic in consequence severity: a higher-consequence inference may not be configured with a lower threshold than a lower-consequence one, and the runtime rejects policy references that violate this invariant at load time.

Operating Parameters

Four parameter families fully describe a confidence gate's behavior. The first is the threshold tier table itself, which lists, for each inference class declared by the governance class, the minimum acceptable value for each component of the input confidence vector. Tables typically contain between three and seven tiers, with safety-critical deployments declaring more tiers to provide finer differentiation between consequence levels. The patent specifies a minimum of three tiers for any deployment that produces commitments, because a single tier collapses the consequence distinction the gate is designed to enforce.

The second family governs the inquiry format. Inquiries are typed, and the type system declares for each failing-component pattern which inquiry shape the gate emits. A sensor-pedigree shortfall produces an inquiry asking for the missing pedigree certificate; a freshness shortfall produces an inquiry asking for re-acquisition; a corroboration shortfall produces an inquiry asking for an additional independent source. The shapes are declared in the policy reference and may not be free-form, because downstream consumers depend on parsing them deterministically.

The third family governs resumption. When an inquiry is satisfied, the gate must decide whether the original confidence evaluation may be reused with the new evidence patched in, or whether the entire vector must be re-evaluated against the tier table. The patent's default is full re-evaluation, on the grounds that partial reuse opens a class of attacks in which an adversary supplies high-confidence evidence for the failing component while the previously-passing components have decayed below threshold. Deployments may opt into partial reuse only under governance classes that explicitly declare it, and only for inquiry types that the class names as safe for reuse.

The fourth family governs the inquiry budget. A runaway sequence of inquiries against an unrelenting low-confidence input is itself a failure mode, and the patent requires that each inference request carry a bounded inquiry budget declared by the governance class. When the budget is exhausted, the gate emits a terminal refusal and closes the request, recording the budget exhaustion in the lineage so that an auditor can distinguish a confidence-gated terminal refusal from an anchored-resolution non-execution. The two are structurally different events with different downstream consequences.

Alternative Embodiments

A first alternative embodiment introduces a tier-promotion mechanism in which an inference request may be downgraded to a lower-consequence tier if its inputs fail the original tier's threshold but pass a lower one. Promotion is not automatic; the requesting principal must accept the downgrade in a structured response to a downgrade-offer inquiry. This embodiment is preferred in interactive deployments where the principal can meaningfully choose between a less binding answer and additional evidence collection, but it is forbidden in deployments where the consequence tier is fixed by external regulation.

A second alternative replaces the per-component minimum with a weighted aggregate that admits inputs whose vector sum exceeds a tier-bound aggregate threshold even if individual components fall short. The patent contemplates this embodiment but flags it as suitable only for low-consequence tiers, because aggregation allows a strong corroboration component to mask a weak chain-of-custody component, which is precisely the substitution the per-component formulation is designed to prevent.

A third alternative introduces a deferred-evidence mode in which the gate may admit an input provisionally on the strength of an obligation to supply the missing evidence within a bounded window. The provisional admission is sealed to the obligation, and any inference produced under the provision carries a deferred-evidence flag in its lineage. If the evidence does not arrive within the window, every inference produced under the obligation is rolled back through the lineage and the principals to whom outputs were emitted are notified through the same channel that delivered the original results.

A fourth alternative integrates the gate with an external evidence service that can answer some inquiries automatically by querying authoritative directories. Automated inquiry resolution is policy-restricted to inquiry types whose answers can be verified end-to-end without principal involvement, and the resulting evidence carries a service-binding in its pedigree component so that the original gate can re-evaluate against the augmented input.

Composition

The confidence gate composes upstream with the input acceptance layer, which is responsible for producing the confidence vector that the gate consumes; downstream with the anchored resolution mechanism, which receives only inputs that have passed the gate; and laterally with the governance class, which provides the threshold tier table, the inquiry shapes, and the inquiry budget. The lineage log threads through all three couplings, recording each gate evaluation, each inquiry, each resumption, and each terminal refusal under the policy hash active at the moment of the event.

The gate's relationship to the slope-constrained simulator is indirect but important. Forecasts produced by the simulator are themselves inputs to subsequent inference requests, and they carry confidence vectors derived from the envelope evaluation. A forecast whose envelope evaluation produced a discontinuity record arrives at the gate with a sensor-pedigree component reflecting that discontinuity, which causes the gate to close for high-consequence inferences and emit an inquiry asking for ground-truth confirmation. This is the structural pathway by which speculative discontinuities propagate into inquiry rather than into action.

Prior Art Distinction

Existing systems that condition output on input quality fall into two categories, neither of which captures the structure the patent describes. The first category produces hedged outputs, attaching a confidence score or a verbal qualifier to a result generated regardless of input quality. This approach leaves the consumer to interpret the qualifier and is known to produce outputs that downstream systems treat as authoritative despite the hedge. The second category refuses to answer when input quality falls below a threshold, returning an error or empty response. This approach is brittle because it provides no mechanism for resolving the underlying input deficiency and forces the principal to redesign the request.

Confidence-gated inference is distinct in three structural respects. First, the threshold is bound to a governance tier rather than to the model, so the same model serving multiple consequence tiers has different gates without retraining. Second, gate closure produces a typed inquiry naming the specific evidence required, transforming the gate from a refusal into a request that can be answered. Third, the gate is forbidden by patent specification from emitting fabricated outputs in any configuration, which is not a property prior systems claim because their architectures do not separate input acceptance from inference production at the structural level.

Implementation Notes

Reference implementations of the confidence gate decompose the threshold check into a vector comparison primitive and a tier-table lookup primitive, with the two combined by a small dispatcher that handles the failure path. The vector comparison is constant-time for a fixed component count, and the tier-table lookup is a content-addressed read against the active policy reference; together they place the gate in the runtime's hot path with predictable latency. The patent specifies that the gate's worst-case latency must be bounded independently of input content, because variable-latency gates leak information about input quality through timing channels and that information can be exploited by an adversary attempting to craft inputs that pass the gate by minimal margin.

Inquiry routing is implemented as a typed channel rather than a free-form callback. Each governance class registers, at policy load time, the set of inquiry types it permits and the routing destination for each type. The runtime rejects at startup any policy whose registered routes are incomplete or contradictory, and the patent specifies that runtime route changes are not permitted; route changes require a new policy hash and a re-load. This discipline ensures that an auditor reconstructing an inquiry sequence can determine route-by-route which destination handled which inquiry under which policy, without reference to runtime state that may have been overwritten.

Resumption is implemented as a continuation of the original inference request, not as a new request. The original request's substrate binding, its lineage references, and its inquiry history are carried forward into the resumption attempt, which means a single principal-facing inference can survive an arbitrary number of inquiry rounds within its budget while remaining a single auditable event. When the budget is exhausted or the principal abandons the inquiry sequence, the runtime emits a sealed terminal record that summarizes the inquiry history and closes the request. Terminal records are themselves anchored under the same governance class as the original request and are produced exactly once per request, regardless of how many inquiries preceded them.

Disclosure Scope

This article presents the confidence gate at the level required to evaluate its role in the Cognition Patent. The full specification — including the confidence vector schema, the inquiry type system, the resumption protocol, the inquiry budget enforcement rules, the deferred-evidence rollback procedure, and the integration with the external evidence service — is provided in the patent's inference control chapter together with reference threshold tier tables for the deployment domains the inventor has tested. Licensees should consult the formal specification when implementing for regulated deployment, as several details necessary for certification are deliberately summarized rather than reproduced here.