Tier-Weighted Admissibility

Mechanism

The architecture's admissibility evaluator receives, for each candidate intent, a tuple containing the intent payload, the operator credential, the operator's tier assignment, and a set of supporting observations. The evaluator reads the policy reference to obtain the envelope authorised at that tier — expressed as bounds on action surface, time horizon, target population, resource consumption, and irreversibility class — and tests whether the candidate intent fits within those bounds. Intents inside the envelope are admitted directly. Intents at or beyond the envelope edge are not rejected outright; instead, the evaluator switches into evidence-stacking mode and asks whether enough corroboration exists to lift the candidate over the threshold.

Evidence stacking is a deterministic accumulation. Each supporting observation carries its own credential and tier, and contributes a weighted increment to a composite admissibility score. Tier 1 observations — full cognitive-state broadcasts from cooperative agents — contribute the largest increments because the broadcasting agent reports its own intent with full fidelity. Tier 2 observations — structured partial signals such as turn signals, beacon emissions, or constrained protocol fields — contribute moderate increments because the channel is narrow but structurally interpretable. Tier 3 observations — behaviour-inferred attributions derived from the agent's perception stack — contribute the smallest increments because inference can be wrong. The composite score is compared against a tier-bound admission threshold, and the intent is admitted only if the score exceeds the threshold.

The output of the evaluator is itself a credentialed observation written into the agent's lineage: the candidate intent, the tier, the envelope at the time of evaluation, the supporting observations that contributed, the resulting composite score, and the admission decision. Any auditor can replay the decision from the ledger because the policy reference, the credential set, and the observation set at the moment of evaluation are all retrievable.

Operating Parameters

The default tier weights are governance-credentialed and configurable per consumer. A defence-context consumer operating against an adversarial population applies different weights than a civilian-context consumer operating against a cooperative population: the former discounts Tier 1 broadcasts more aggressively because broadcasts can be spoofed, while the latter weights Tier 3 inferences less aggressively because most agents in the environment are cooperating. Each weight curve is recorded in the policy reference under a credentialed revision so that admissibility decisions remain reproducible across policy changes.

The envelope shape is parameterised per tier and per intent class. A high-tier operator may have a wide envelope for routine intents and a narrow envelope for irreversible ones; a low-tier operator may have a narrow envelope across the board but with intent-class-specific exceptions where domain expertise has been credentialed. Envelope edges are expressed as continuous functions rather than hard cutoffs, so that intents near the edge degrade gracefully into the evidence-stacking regime rather than flipping abruptly between admitted and rejected.

The admission threshold is itself tier-conditioned. Higher-tier operators face lower thresholds because their baseline credential already supplies most of the warrant; lower-tier operators face higher thresholds because more of the warrant must come from corroborating observation. This asymmetry encodes the structural intuition that credential and corroboration are partially substitutable: an intent supported by a strong credential needs less corroboration, and an intent supported by abundant corroboration needs less credential, but every admitted intent must pass some combined warrant.

Alternative Embodiments

In one embodiment the weights are static and policy-set; in another they are adaptive, updated by a credentialed learner that observes downstream outcomes and adjusts the weight curves to minimise admission errors. The adaptive variant requires that every weight update is itself a credentialed policy revision so that historical decisions remain reproducible against the weights in force at the time. A third embodiment supports per-domain weight overlays — a single deployment can host civilian-context and defence-context weight curves simultaneously, with the active overlay selected by the operating context observation rather than by global configuration.

A fourth embodiment introduces cross-tier disagreement detection. When Tier 1, Tier 2, and Tier 3 contributions about the same entity agree, the composite score increases beyond what any single contribution would supply, because independent observation streams confirming the same intent provide a multiplicative rather than additive warrant. When they disagree, the architecture not only fails to admit but emits a credentialed disagreement observation that downstream consumers — adversarial-aware operators, security monitors, fault detectors — can subscribe to. A vehicle whose Tier 1 broadcast disagrees with its Tier 2 partial signals or its Tier 3 behaviour inference is structurally suspect: possibly compromised, possibly malfunctioning, possibly adversarial. Disagreement, in this embodiment, is itself a first-class signal.

A fifth embodiment exposes the evaluator's score and supporting evidence to the operator, allowing iterative intent refinement: an operator whose intent fails to clear the threshold may supply additional corroboration or narrow the intent payload until admission becomes possible. This embodiment is appropriate where operators are sophisticated enough to participate in the warrant negotiation and where time permits the round-trip.

Composition

Tier-weighted admissibility composes with the operator-intent broadcast and observation primitives described elsewhere in the cognition patent. The same fidelity tiers that govern broadcast — Tier 1 full cognitive-state, Tier 2 structured partial, Tier 3 behaviour-inferred — also govern the weights applied during admission. There is no parallel tier system; the architecture maintains a single tier vocabulary across emission, observation, and admissibility, so that an operator's tier means the same thing whether it appears as the source of an intent or as a contributor of corroborating evidence.

The mechanism also composes with capability and permission gating. Once an intent is admitted by the tier-weighted evaluator, the resulting authorised action still passes through the capability and permission gates before execution: admission warrants only that the operator may issue this intent, not that the agent can or is permitted to perform it. The three evaluations — admissibility, capability, permission — are independent and additive; an action proceeds only when all three concur. This separation prevents the common failure mode in which a strong operator credential is allowed to override either physical infeasibility or governance prohibition.

Prior-Art Distinction

Conventional intent-fusion systems apply equal weights across heterogeneous sources, or apply ad hoc per-source weights set at integration time and rarely revisited. In environments dominated by low-fidelity sources — urban traffic where most vehicles are human-driven and only inferable at Tier 3 — equal-weight fusion produces estimates dominated by the most numerous source rather than the most informative one. The few cooperative agents broadcasting at Tier 1 are drowned out by the inferences. The disclosed mechanism inverts this failure mode by binding weights to fidelity tier rather than to source count, so that a single Tier 1 broadcast can outvote many Tier 3 inferences when their informational value warrants it.

Conventional access-control systems also typically treat operator authority as a flat predicate — either the operator can issue this command or cannot — without graduated envelopes or evidence stacking. Where graduated authority exists, it is usually expressed as discrete role tiers without the structural coupling to environmental corroboration that the disclosed mechanism provides. The combination of tier-bound envelopes, deterministic evidence stacking against tier-conditioned thresholds, and the resulting credentialed admissibility observation is not present in prior intent or authority architectures.

Worked Example

Consider an urban traffic environment in which an autonomous vehicle is deciding whether to admit a yield intent attributed to an oncoming vehicle. A Tier 1 broadcast from the oncoming vehicle declares its intent to yield. A Tier 2 partial signal — its turn indicator and a brake-light pattern — corroborates the broadcast. A Tier 3 inference from the perception stack observes deceleration consistent with yielding. All three concur, the composite score exceeds the admission threshold by a substantial margin, and the yield intent is admitted. The autonomous vehicle proceeds.

Now consider the same scenario in which the Tier 1 broadcast declares yield but the Tier 2 turn-indicator signal is absent and the Tier 3 inference observes acceleration rather than deceleration. The cross-tier disagreement detector fires, the composite score remains below threshold despite the Tier 1 contribution, and the yield intent is not admitted. A credentialed disagreement observation is published, and downstream consumers — including any adversarial-aware monitor subscribed to the channel — receive the signal. The autonomous vehicle treats the oncoming vehicle as if no yield intent had been issued, and falls back to the perception-driven trajectory the Tier 3 inference supports. The structural separation between admission and corroboration is what allows the architecture to handle the disagreement without conflating an unverified broadcast with a confirmed intent.

Deployment Considerations

Deploying tier-weighted admissibility into a heterogeneous fleet requires three coordinated choices. First, the credential authority that issues operator tiers must be itself credentialed and recorded in the policy reference, so that tier assignments are auditable rather than ambient. Second, the observation channels that feed corroborating evidence must publish their tier alongside their payload, so the evaluator can apply the correct weight without inferring the tier from channel identity. Third, the policy reference must support credentialed revisions of weights, envelopes, and thresholds without invalidating historical decisions; the architecture pins each admission to the policy revision in force at the moment of evaluation, so that subsequent revisions do not retroactively change the meaning of a recorded decision.

Mixed-fleet operation — a few cooperative agents broadcasting Tier 1, more entities visible at Tier 2, most entities only inferable at Tier 3 — is the realistic operating regime for nearly every domain in which this architecture deploys. The disclosed mechanism is calibrated for this regime by construction: weight curves are parameterised to give Tier 1 broadcasts informational dominance even when they are vastly outnumbered by Tier 3 inferences, while still allowing many concurring Tier 3 inferences to challenge a single anomalous Tier 1 broadcast. This is the structural foundation on which adversarial-aware operation, cooperative-fleet coordination, and graceful degradation under sensor loss all rest.

Disclosure Scope

The disclosure covers any embodiment in which the admissibility of an operator-issued intent is evaluated against a tier-bound envelope, with marginal intents lifted by deterministic accumulation of tier-weighted supporting observations against a tier-conditioned threshold, and with the resulting decision recorded as a credentialed observation reproducible from the policy reference. The specific weight curves, envelope shapes, threshold functions, disagreement-handling rules, and adaptive variants are illustrative embodiments. The disclosure spans civilian and defence contexts and is parameterised per consumer through credentialed policy revisions, and extends to any combination of the variants described above so long as the canonical-store discipline and the credential-and-corroboration warrant structure are preserved.