Behavior-Inferred Intent as Governed Observation

Mechanism

A unit observing another unit's behavior runs an inference function over a vector of cues — trajectory, signaling, role, prior interactions, environmental context — and produces a candidate proposition of the form "operator B's intent class is X." Rather than write that proposition into a shared belief store as a fact, the mechanism wraps it in an observation envelope. The envelope carries the inferring unit's credential, the inferred operator's identifier or pseudonym, the proposed intent class, a scalar confidence, the version reference of the inference function, and the supporting cues used. The envelope is signed by the inferring unit and published into the same observation channel that carries first-party declarations.

The envelope is the unit of consumption. A consumer reading the channel does not see "B intends to merge"; it sees "unit A inferred at confidence c, using inference function v, from cues {...}, that B's intent class is merge." Consumer policy decides what to do with that record based on the confidence, the inferring unit's authority, and the consumer's own threshold. A high-stakes downstream action — for example, yielding a right-of-way in a multi-agent traffic scenario — may require multiple corroborating inferences from independently credentialed units, each above its own threshold. A low-stakes action may use a single inference at moderate confidence. The inference function's output never bypasses these thresholds by being silently coerced into a fact.

The inferred operator — B in the above example — is itself a consumer of the channel and can publish a signed counter-observation: "I am B; the inference that I intend to merge is incorrect; my declared intent is X." The counter-observation is processed through the same admissibility evaluator that processes any other observation. Whether it overrides the original inference depends on policy: in many configurations, a first-party declaration of one's own intent outweighs a third-party inference; in others (for example, where deception is plausible) the policy weights both and produces a combined posterior. The point is structural: the inferred party has a credentialed pathway to dispute, and downstream policy is defined over the joint record of inference and counter-observation rather than over a single overwritten belief.

Lineage is preserved end-to-end. Each derived state in a downstream consumer that depends on the inference carries a reference back to the originating envelope. If the inference is later retracted by the inferring unit, contradicted by a higher-confidence corroboration, or overridden by an admitted counter-observation, the dependency graph identifies which downstream states are affected and propagates the change. There is no point in the pipeline where the inference loses its provenance and becomes an unattributed assumption.

Operating Parameters

The envelope schema fixes the field set: inferring credential, inferred operator identifier, intent class drawn from a published taxonomy, confidence on a normalized scale, inference function version, cue set, and timestamp. Confidence is calibrated against the inference function's published evaluation record, so that a confidence of 0.8 emitted by function v1.4 is comparable across deployments that share that version. The cue set is included by reference rather than by raw payload where the cues are sensor data of nontrivial size, with content-addressed handles that consumers can dereference if they have the authority to do so.

Consumer-side parameters include the per-action confidence threshold, the corroboration count required at each threshold tier, the authority weighting that scales an inferring unit's confidence by its credential class, the staleness window after which an inference is no longer admissible, and the override policy that resolves conflicts between an inference and a counter-observation. Override policy is itself parameterized: pure first-party precedence, weighted combination, adversarial-aware combination that down-weights first-party declarations when they conflict with multiple independent inferences, and configurable hybrids.

Inferring-side parameters control the publication discipline. A unit may suppress publication below a minimum confidence; may rate-limit publication per inferred operator to avoid flooding; may redact cue references where the cues themselves are sensitive. A unit must, however, sign every published envelope, must record its publications in its own audit log, and must honor retraction requests against its own envelopes. The asymmetry between publish-discretion and audit-mandatory is deliberate: a unit can choose silence, but it cannot publish anonymously or unrecorded.

Alternative Embodiments

The inference function can be embodied as a deterministic rule set, as a learned classifier, as an ensemble, or as a chain of an LLM proposing intent classes followed by a calibration stage that maps the proposal onto the published taxonomy and confidence scale. The envelope does not depend on the function's internals; it depends on the function having a stable version reference and a published evaluation record that grounds the confidence scale. Inference functions can be private to a unit or shared across a fleet; in shared embodiments, the version reference points to the shared function and the inferring unit's credential remains distinct from the function's authorship.

The observation channel can be embodied as a broadcast mesh in which all credentialed units consume all envelopes, as a topic-partitioned channel keyed by inferred operator identifier, or as a query-response channel in which inferences are persisted and served on demand. Counter-observations can be embodied as inline retractions on the same channel, as separate retraction streams, or as challenge transactions that pair an original envelope identifier with a signed dispute. Adversarial-aware embodiments may require challenges to themselves carry corroborating cues rather than being bare denials.

Lineage tracking can be embodied as inline reference fields on each downstream state, as a side-car dependency graph maintained by an audit service, or as a reconstructable property derivable from a content-addressed event log. The choice depends on the latency budget for retractions to propagate; inline references propagate fastest, side-car graphs are easier to reason about at scale, and reconstructable embodiments minimize hot-path overhead at the cost of slower retraction processing.

Composition With Adjacent Mechanisms

Inferred-intent observations compose with declared-intent declarations in a single ordered channel. Declarations are first-party envelopes; inferences are third-party envelopes; both are signed; both carry confidence, although declarations typically operate at a maximum confidence within the inferring party's own authority. Consumer policy treats them as records of the same kind with different authority weightings, which lets the same threshold and corroboration logic apply uniformly.

The mechanism composes with the evidence-based gating layer described in companion disclosures: a downstream agent's authorization to act on inferred intent at a given confidence may itself be a gated capability, with the gate referencing the agent's track record of acting correctly on inferred-intent records of similar profile. It composes with the feedback-asymmetry layer: an agent that acts on low-confidence inferences and produces sustained validation failures will see its authorized surface contract immediately, while widening its surface to act on lower-confidence inferences requires the slow, confirmed positive path.

The mechanism also composes with multi-party arbitration. When two inferring units produce contradictory inferences about the same operator, arbitration consumes both envelopes and either selects one, combines them weighted by authority and confidence, or defers and emits a structured request for additional cues. The arbitration outcome is itself a published envelope carrying its own lineage, which preserves the property that no fact in the system lacks a reviewable origin.

Distinction From Prior Art

Conventional sensor-fusion and situational-awareness systems produce intent estimates as private state inside the perceiving system. The estimates may be excellent statistically; they are nonetheless not artifacts that other parties can read, that the inferred subject can challenge, or that downstream regulators can audit as the source of an action. When the estimate is wrong, the resulting decision often has no traceable artifact distinguishing "we were told B intended to merge" from "we inferred B intended to merge." The structural collapse of these two cases into a single internal proposition is what the disclosed mechanism prevents.

Conventional intent classifiers built atop machine-learned models likewise emit confidences, but the confidences are not bound into a credentialed envelope, are not addressable by the inferred party, and do not feed a structural retraction pathway. Bayesian belief systems do represent probabilistic propositions, but they typically operate over an internal belief network rather than over a multi-party signed observation channel; the belief is a private posterior, not a publishable artifact under a credential. The disclosed mechanism's contribution is the binding of the probabilistic representation to a credentialed, challengeable, lineage-tracked observation envelope that downstream policy must key against — not the probability of the intent in isolation, but the joint object of inference, source, lineage, and dispute history.

Disclosure Scope

The asymmetry between a unit's discretion in publishing and its mandatory audit of its own publications has a further consequence in adversarial environments. A unit cannot publish many low-confidence inferences silently to flood a target's reputation while disclaiming responsibility, because every publication is signed and recorded against its credential. A unit attempting to retract its publications wholesale to evade accountability cannot do so because the audit log is append-only with respect to its own history; retractions are themselves logged events that reference the original envelope rather than erasing it. The combination of credentialed publication, signed counter-observation, and append-only audit produces a system in which inference is a first-class governed activity with attached responsibilities, not a private cognitive operation whose outputs leak unattributed into shared state.

The disclosed mechanism encompasses the observation envelope schema, the calibration of confidence against a versioned inference function, the credentialed publication discipline, the counter-observation pathway available to the inferred operator, the consumer-side parameter set governing thresholds, corroboration, authority weighting, staleness, and override, and the lineage-preserving binding between inference envelopes and downstream derived state. The mechanism is independent of any specific inference function family, of any specific observation transport, and of any specific intent taxonomy; equivalent embodiments that preserve the envelope-as-distinct-from-fact property, the credentialed-and-signed publication, the challenge pathway, and the lineage propagation are within scope. The mechanism does not claim novelty in intent estimation per se; it claims novelty in the structural treatment of inferred intent as a governed observation event whose downstream effects are conditioned on confidence and lineage rather than on coercion to ground truth.