Operator Intent: Graduated Fidelity Tiers for Mixed-Fleet Coordination

Problem: Mixed-Fleet Coordination Lacks an Intent Substrate

Coordination between autonomous units depends on each unit knowing what its neighbors intend. Current architectures handle this by dividing the population in two. Cooperative agents — those running compatible stacks under a shared protocol — exchange intent over a high-bandwidth machine-to-machine channel, typically expressing planned trajectories, capability envelopes, and short-horizon planning graphs. Everything else — human drivers, legacy aircraft, foreign drones, animals, weather, adversarial entities — is treated as an opaque hazard whose behavior must be guessed from raw sensor data. The two populations live in disjoint architectural regimes, and the boundary between them is where systems fail.

The brittleness manifests in characteristic ways. An autonomous vehicle that coordinates tightly with another autonomous vehicle still treats the human driver in the next lane as an opaque trajectory generator, even when that driver is signaling explicit intent through turn signals, brake lights, lane positioning, and gaze direction. A drone equipped with cooperative airspace data is paralyzed when a non-cooperative drone enters its operating window because the inference apparatus is bolted on as a separate subsystem rather than fused into the same intent substrate. A defense system distinguishing combatants from noncombatants has no architectural framework that scales smoothly from broadcast intent (allied units), to inferred intent (unknown civilian or commercial units), to hostile-intent classification (adversarial units), with each transition governed by appropriate credentialing.

The structural consequence is that mixed-fleet operation gets relegated to one of two poor configurations: either constrained environments where everyone is artificially made cooperative (geofenced shuttle routes, segregated drone corridors, fully autonomous commercial fleets in dedicated lanes), or fully-segregated infrastructure where mixing is prevented architecturally (banning Class B drones from civilian airspace, requiring AV-only zones, restricting autonomous trucking to dedicated rights-of-way). Neither configuration scales to the actual deployment environments where autonomy matters most: existing road networks with millions of human drivers, shared low-altitude airspace, and contested battlespaces containing civilian populations.

The deeper architectural gap is the absence of a unified intent substrate that admits varying fidelity from different broadcasters while preserving a single coherent decision surface for the consuming unit. Intent is not binary; it is a continuum of how much the broadcaster has chosen, or has been required, to disclose. A coordination architecture that cannot consume that continuum cannot operate in the world as it actually exists.

Core Primitive: Three Fidelity Tiers Fused Under Composite Admissibility

The operator intent primitive consumes intent at three distinct fidelity tiers, each with its own data model, credential requirements, and admissibility weight. Tier 1 is full cognitive-state broadcast: cooperative agents publish their planning graphs (trajectory rollouts at 5 Hz to 20 Hz over horizons of 2 to 30 seconds), intent fields (target waypoints, mission objectives, formation roles), and capability envelopes (current actuation latency, sensor health, payload mass, energy reserve). Tier 2 is structured partial-fidelity bus extraction: vehicles, devices, or aircraft publish a limited but architecturally specified subset — turn signals, brake lights, hazard flashers, lane-change indicators, ADS-B Out broadcasts, transponder modes, route plans filed with traffic authorities, formation orders distributed among allied units. Tier 3 is behavior-inferred attribution: operating units derive intent from sensor cues including instantaneous trajectory and acceleration, gaze and gesture (for human operators visible to camera systems), wheel position, formation geometry, historical pattern, and consistency with previously declared plans.

All three tiers produce intent observations into a unified composite admissibility evaluator. The evaluator does not pick one tier and ignore the others; it weights the contributions and integrates them. Tier 1 broadcasts arrive with a credential signed by the broadcaster, are weighted most heavily (typical default weight 0.7 to 0.9 of the composite estimate when fresh and uncontradicted), and degrade with age according to a configurable decay function. Tier 2 observations carry the weight associated with their specific channel (a turn signal weighted 0.4 to 0.6, an ADS-B intent broadcast weighted 0.5 to 0.8, a filed flight plan weighted 0.3 to 0.5 because it expresses long-horizon intent that may have been revised). Tier 3 inferences carry weight inversely proportional to the inference function's recent verification record, typically 0.1 to 0.4.

The composite admissibility evaluator combines the tier observations with environmental context, governance policy, and the consuming unit's own dispositional state to produce a single coherent intent estimate per neighbor. The estimate is not a point prediction; it is a distribution over possible intents with explicit uncertainty, and it carries provenance pointers to the contributing observations so that downstream consumers can audit how the estimate was constructed.

Tier consumption is itself an architectural primitive. An operating unit can be configured to consume any subset of tiers based on its policy: a high-trust military formation may admit only Tier 1 from credentialed allies and treat all Tier 2 and Tier 3 as advisory; a civilian AV may weight Tier 2 most heavily because Tier 1 broadcasters are rare in mixed traffic; a regulated drone operator may be required by airspace policy to consume Tier 1 from any UTM service supplier in range. The same physical neighbor can be observed across all three tiers simultaneously, with the evaluator handling agreement and disagreement between tiers structurally rather than as an exception path.

Mechanism 1: Behavior-Inferred Intent as Governed Observation

Tier 3 — behavior-inferred attribution — is the lowest-fidelity tier and the one most prone to error and abuse. Conventional sensor-based situational awareness produces these inferences as private internal state of the consuming unit; the inference is consumed by the consuming unit's own planner and discarded. The operator intent primitive treats Tier 3 inferences fundamentally differently. Each inference is structured as a governed observation: a published artifact carrying the inferring unit's credential, the inferred operator's identifier (or a privacy-preserving pseudonym when identification has not been authorized), a reference to the specific inference function and version that produced the inference, the supporting cues with their timestamps, and a temporal commitment specifying when the inferred behavior is predicted to manifest.

Because Tier 3 inferences are published observations rather than private state, they are subject to the same provenance and audit machinery as any other observation in the system. Other observers in the mesh can integrate the inference (raising or lowering its admissibility based on their own observations of the same operator), contradict it (publishing an opposing observation that the evaluator combines with the original), or challenge it on procedural grounds (objecting that the inference function was not credentialed for this operating context).

Crucially, the inferred operator has structural standing to challenge through governance-credentialed retraction. A unit that has been classified — by another unit's Tier 3 inference — as 'preparing to merge' or 'evading checkpoint' or 'intent unknown' can publish a retraction observation that, when properly credentialed, structurally reduces the admissibility weight of the original inference. This produces a feedback channel that current architectures lack entirely. In legacy telematics, an operator misclassified by an inference engine has no architectural remedy; the misclassification flows downstream to insurers, traffic authorities, and law enforcement databases without a mechanism for structural retraction. The intent primitive replaces appeal-by-litigation with retraction-by-protocol.

The retraction credential is itself tiered. A self-retraction by the inferred operator carries one weight; a retraction by an authority credentialed over that operator (parent fleet operator, employer, or regulatory body) carries another; an automated retraction triggered by verification feedback (the predicted behavior did not occur) carries another still. The composite admissibility evaluator combines retractions with the underlying inferences in the same machinery it uses for intent observations themselves.

Mechanism 2: Verification-Feedback Inference Evolution

Each Tier 3 inference carries a temporal commitment: a prediction that a specific behavior will follow within a specified window. The window is bounded by the inference function — typical ranges are 250 ms to 5 s for vehicle micro-maneuvers, 5 s to 60 s for routing-level intent, 1 to 30 minutes for mission-level intent. After the window expires, the actual behavior is observed and compared to the prediction. The agreement is recorded as a verification observation against the originating inference, with its own credential and provenance.

Verification observations accumulate over many inferences from the same function. The evaluator maintains a running posterior over the function's accuracy, conditioned on operating context (urban vs. highway, daylight vs. night, dense traffic vs. sparse, civilian vs. military, etc.). Functions that consistently match observed behavior gain admissibility weight in the contexts where they perform well; functions that consistently miss are demoted. New inference functions can be proposed by any credentialed contributor and admitted under sandboxed evaluation (their inferences are recorded but weighted at zero in production decisions) until their verification record justifies promotion.

The structural consequence is that the inference apparatus is closed-loop and falsifiable. Every Tier 3 inference is a prediction; every prediction expires; every expiration produces a verification observation; verifications drive function evolution. This eliminates the brittleness of frozen-at-training-time classifiers in domains where the underlying behavior co-evolves — adversarial driving, evolving formation tactics, novel drone designs — because the inference functions are continuously re-evaluated against the world they are trying to predict.

Verification operates at multiple time scales. Micro-verification (single-prediction granularity) modulates per-instance admissibility weight in real time. Aggregate verification (rolling windows of hundreds to thousands of predictions) drives function-level promotion and demotion. Long-horizon verification (months to years) drives the fleet-wide retirement of inference function families and the introduction of replacements. The same machinery serves all three.

Mechanism 3: Multi-Authority Intent Composition

Mixed-fleet coordination rarely operates under a single regulatory authority. A civilian drone descending through low-altitude airspace may be subject simultaneously to FAA UTM rules (in U.S. airspace), to local-jurisdiction noise ordinances, to the operator's own corporate policy, and to mutual-aid agreements with adjacent fleets. A military asset operating in a coalition zone is subject to its national rules of engagement, the coalition mission policy, the host-nation airspace authority, and DoD Directive 3000.09 on autonomy in weapon systems. The intent primitive composes intent across these authorities rather than picking one and ignoring the rest.

Each authority is represented by a credentialed observer. The credential specifies which intent fields the authority is authorized to constrain (an FAA UTM service supplier may constrain altitude and routing but not payload composition; an EASA U-space class definition may constrain identification and tracking obligations but not mission objectives; DoD 3000.09 doctrine may constrain target classification but not navigation). When intent observations are produced by an operator, the multi-authority composer combines them with the relevant authorities' constraints to produce a constrained intent surface — the subset of intents that satisfy every applicable authority.

The composer handles authority conflict structurally. When two authorities issue contradictory constraints (a coalition mission policy authorizes a route that the host-nation airspace authority forbids), the conflict is published as a governance observation rather than silently resolved. Resolution policy is itself credentialed: the operator's credential specifies a precedence hierarchy among authorities, the conflict is logged with full provenance, and downstream consumers see both the constrained intent and the conflict that produced it. This replaces the brittle 'fail-closed when any authority objects' default with a structural mechanism that supports live multi-authority operation.

The regulator-as-credentialed-observer pattern generalizes to non-state authorities. A fleet operator coordinating across multiple subsidiaries can credential each subsidiary as a constraining authority over its own assets. A consortium of drone operators sharing a low-altitude corridor can credential the consortium as a coordinating authority over the shared corridor. A standards body that publishes a behavioral specification can credential its specification as a constraining authority for subscribers. The same multi-authority composer serves all of these patterns with configuration changes rather than re-implementation.

Mechanism 4: Risk-vs-Hostility Bifurcation Under Due Process

Current insurance-based usage telematics — Cambridge Mobile Telematics, Nauto, Lytx, the OEM driver-monitoring stacks — conflate competence-based risk (an operator who is unsafe due to inattention, fatigue, intoxication, or skill limits) with intent-based hostility (an operator who is deliberately attempting harm). The conflation is architectural: the same scoring engine that produces actuarial premiums also produces flags consumed by law enforcement and corporate safety teams, with no structural separation between 'this operator is incompetent' and 'this operator is adversarial.' Operators flagged for risk reasons have no mechanism to contest the assertion that they are hostile, because the architecture does not distinguish the two assertions.

The intent primitive separates them by construction. A risk profile is constructed from observed behavior under normal-operation assumptions — a Bayesian estimate over the operator's competence parameters (reaction latency, attention span, skill envelope) — and is used for actuarial purposes, training intervention, and capability-based admissibility (e.g., gating high-stakes maneuvers). A hostility profile is constructed from behaviors structurally indicative of adversarial intent (deliberate counter-flow on a one-way thoroughfare, targeting trajectory toward a protected asset, weapon-deployment cues, formation patterns characteristic of coordinated attack) and is used for protective response.

The two profiles use different evidence, different inference functions, and different downstream consumers. They are not interchangeable. A high risk profile does not imply hostility; a low risk profile does not preclude hostility; risk and hostility coexist as orthogonal axes of the operator's intent characterization.

Critically, hostility profile construction requires due-process credentialing. A regulatory or judicial authority must have credentialed the criteria under which hostility classification operates in the relevant jurisdiction; the classification event itself must be governed by audit-grade lineage (which inference functions, which observations, which thresholds, which authority's credential authorized the operation); and the classified operator has structural standing to challenge through the same governance-credentialed retraction machinery that governs Tier 3 inferences generally. This puts hostility classification on the same footing as protective orders, restraining orders, and other due-process-bound classifications used in civilian governance: an authority bounded the criteria, a procedure produced the classification, a record documents the basis, and the classified party can contest.

Operating Parameters and Composition with Adjacent Primitives

Typical operating parameters reflect the use case. For automotive mixed-fleet coordination at suburban speeds (15 to 45 mph), Tier 1 broadcast cadence is 10 Hz to 20 Hz over horizons of 2 to 8 seconds; Tier 2 extraction operates at 30 Hz to 60 Hz from camera and radar; Tier 3 inference functions produce predictions with 250 ms to 3 s windows. For low-altitude UAS operation (under 400 ft AGL, speeds 0 to 60 kt), Tier 1 broadcast cadence is 1 Hz to 5 Hz over horizons of 5 to 30 seconds; Tier 2 includes ADS-B In/Out at 1 Hz and Remote ID at 1 Hz; Tier 3 windows extend to 10 to 60 seconds because UAS maneuver dynamics are slower. For defense ISR in contested airspace, Tier 1 broadcasts are encrypted under credentialed key material at 1 Hz to 10 Hz; Tier 2 includes IFF Mode 5 and link-16 PPLI at variable cadence; Tier 3 inference windows span seconds to minutes depending on platform class.

The composite admissibility evaluator's tier weights are configurable per consuming unit and per operating context, with typical defaults of (0.8, 0.4, 0.2) for (Tier 1, Tier 2, Tier 3) in low-trust contexts and (0.6, 0.5, 0.3) in higher-trust contexts where Tier 3 inference functions have established verification records. Verification-feedback decay timescales typically run from 100 to 10,000 predictions per function before steady-state weight is reached.

The intent primitive composes with two adjacent primitives in the disclosure family. It composes with the marker-track primitive: intent observations carry marker-track lineage, so an intent estimate for a particular neighbor is bound to the marker-track that establishes that neighbor's identity, and identity confusion structurally invalidates the intent estimate rather than silently mis-attributing intent across operators. It composes with the governed-actuation primitive: intent estimates are inputs to the admissibility envelope that gates actuation, with low-confidence or contested intent narrowing the envelope and high-confidence credentialed intent widening it. The three primitives form a coherent stack — identity, intent, action — with provenance threaded through all three.

Alternative Embodiments

The primitive admits several alternative embodiments without departing from the core mechanism. In a degraded-communications embodiment, Tier 1 broadcasts are unavailable due to jamming or infrastructure failure; the consuming unit operates entirely on Tier 2 and Tier 3, with the composite admissibility evaluator automatically renormalizing weights and broadening uncertainty. In a privacy-preserving embodiment, Tier 1 broadcasts are encrypted under a group key that admits aggregate observation but not individual identification, and Tier 3 inferences operate on pseudonymized identifiers that can be unmasked only under credentialed authority. In a fully cooperative embodiment (geofenced AV-only zone, dedicated drone corridor), Tier 3 is disabled by policy and the system operates on Tier 1 and Tier 2 alone, with the inference function machinery available but not consulted.

The primitive scales across domains. In civilian automotive mixed-fleet coordination, Tier 2 dominates (turn signals, brake lights) and Tier 3 fills the gap for non-signaling maneuvers; hostility classification is disabled by default. In commercial fleet management, Tier 1 dominates (corporate broadcast among the operator's own fleet) with Tier 3 for non-fleet vehicles; risk profiles drive insurance integration. In emergency response, Tier 1 broadcasts among credentialed responders coexist with Tier 2 from civilian vehicles in the response zone. In defense ISR, all three tiers operate concurrently with hostility classification enabled under credentialed mission policy and audit lineage tied to applicable rules of engagement.

Human-in-loop authority is preserved across all embodiments. At every tier, a credentialed human operator can override, retract, or supplement the automated intent observation. The intent primitive does not displace human authority; it extends the apparatus through which human authority is exercised, providing the credential machinery, the audit lineage, and the structural retraction channel that human-in-loop oversight requires to operate at scale.

Prior-Art Distinctions

The intent primitive is structurally distinct from several adjacent technologies. It is not Cambridge Mobile Telematics, Nauto, Lytx, or comparable usage-based insurance telematics. Those products produce behavior-based usage scores from sensor data without an architectural distinction between competence-based risk and intent-based hostility, and without a due-process framework for adverse classifications. The intent primitive bifurcates risk and hostility by construction and carries credentialed retraction as a first-class observation type.

It is not Mobileye REM, HERE HD Live Map, or comparable crowd-sourced map ingestion. Those systems aggregate static and semi-static observation data into shared maps that the consuming unit reads; they do not propagate intent observations or support bidirectional retraction. The intent primitive aggregates per-operator intent at three fidelity tiers with structural retraction by the inferred operator.

It is not ATC voice clearance, ADS-B traffic alerting, or UTM service-supplier APIs. ATC voice produces a cleared-route artifact that is binding but does not propagate as machine-readable intent across the fleet. ADS-B Out broadcasts position and limited intent fields but does not support multi-tier fusion or verification-feedback evolution. UTM service-supplier APIs (e.g., the discovery and synchronization service interfaces standardized under ASTM F3548) coordinate operator declarations but do not consume Tier 3 behavioral inferences or carry due-process-credentialed hostility classification. The intent primitive is the substrate that sits beneath those interfaces and provides the cross-tier fusion they presuppose but do not implement.

It is not lethal autonomous weapons systems (LAWS) doctrine in the conventional sense. Conventional LAWS hostility classification is mission-policy-internal: the system carries its own classification criteria, applies them, and acts; the criteria and the application are not externalized as governed observations. The intent primitive externalizes the classification criteria (credentialed by named authorities), the inference functions (with verification records), and the audit lineage (signed observations). DoD Directive 3000.09 doctrine composes with the primitive as a credentialed authority over hostility-class intent fields, but the primitive is not itself a LAWS. The same machinery serves civilian, commercial, and defense deployments under their respective credentialing frameworks.

Disclosure Scope

The disclosure under USPTO Provisional 64/049,409 covers the operator intent primitive and its sub-primitives: graduated fidelity tiers (Tier 1 full cognitive-state broadcast, Tier 2 structured partial-fidelity bus extraction, Tier 3 behavior-inferred attribution), multi-fleet intent fusion under composite admissibility, multi-authority intent composition with regulator-as-credentialed-observer (including FAA UTM-class composition, EASA U-space-class composition, and DoD 3000.09 LAWS-doctrine composition), human-in-loop authority preservation across all tiers and authorities, and the bifurcation of competence-based risk profiles from intent-based hostility profiles under due-process credentialing.

The disclosure covers composition with adjacent primitives — marker-track for identity binding, governed-actuation for admissibility-gated action — and the alternative embodiments described above (degraded-communications, privacy-preserving, fully cooperative, civilian, commercial, emergency response, defense ISR). It covers the verification-feedback inference evolution mechanism, including sandboxed admission of new inference functions, verification-driven function promotion and demotion, and multi-time-scale verification operating on micro, aggregate, and long-horizon windows.

Mixed-fleet operation — between autonomous and human-driven vehicles, between cooperative and non-cooperative drones, between allied and adversarial entities, across multiple regulatory authorities — moves from special-case workarounds to a unified architectural primitive. The primitive consumes intent at whatever fidelity the broadcaster supplies, fuses tiers under credentialed admissibility, evolves under verification, composes across authorities, preserves human-in-loop oversight, and bifurcates risk from hostility under due process. This disclosure establishes the substrate; downstream implementations specialize the parameters, the inference functions, and the credentialing relationships to their respective deployment domains.