1. Vendor and Product Reality
Cruise LLC, founded in 2013 and majority-owned by General Motors following GM's 2016 acquisition, was, until late 2023, one of two leading commercial robotaxi operators in the United States, alongside Waymo. Its flagship deployment was a paid, driverless ride-hail service in San Francisco, operating a fleet of retrofitted Chevrolet Bolt vehicles equipped with a multi-modal sensor stack of lidar, radar, cameras, and ultrasonic sensors feeding a perception, prediction, planning, and control pipeline. Cruise also developed the Origin, a purpose-built vehicle without traditional driver controls, on the GM BEV3 platform, and had announced or begun expansion efforts in several additional cities before the events described below.
The architectural shape was the canonical driverless AV stack: a perception layer fusing sensor inputs into a tracked-object world model, a prediction layer projecting the trajectories of other road users, a planner generating candidate maneuvers under cost functions, a controller commanding actuators, and safety monitors authorized to trigger a minimal-risk-condition (MRC) maneuver such as a pullover. A remote-assistance operations center allowed human operators to provide guidance under defined stuck-vehicle protocols, and a fleet-management plane handled dispatch, charging, and maintenance. Deployment authority in California was held externally by the California Public Utilities Commission (CPUC) and the California Department of Motor Vehicles (DMV), with the National Highway Traffic Safety Administration (NHTSA) holding federal vehicle-safety oversight.
Cruise's strengths were real: a deep engineering bench, in-house sensor and compute development, a purpose-built driverless vehicle, and genuine operational experience running paid driverless service in a dense US city. Within its operational design domain it was a credible peer to Waymo. On October 2, 2023, a Cruise vehicle in San Francisco was involved in an incident in which a pedestrian, initially struck by a separate human-driven vehicle, was thrown into the Cruise vehicle's path; the Cruise vehicle then performed a maneuver that dragged the pedestrian a further distance before stopping. California regulators subsequently found that Cruise had not fully disclosed the dragging behavior, and the DMV suspended Cruise's driverless deployment and testing permits. Cruise then voluntarily paused its driverless operations more broadly. This article does not relitigate those findings. It reads the episode for one architectural property: the response option set available to regulator and operator was structurally shaped by the execution stack itself.
2. The Architectural Gap
The property the Cruise stack did not expose, in common with the driverless AV category generally, is a first-class distinction between continuing to execute and being deliberately held in a non-executing but still-reasoning state, chosen by the system itself and by its overseers as a revocable, graduated condition rather than an all-or-nothing switch. The vehicle was certified to operate at full authority within its approved operational design domain, or it was not operating. The stack contained safety-integrity mechanisms, redundant sensors, monitor-actuator architectures, and a defined MRC behavior. What it did not contain was a structural gate that treats the permission to act as something continuously re-earned and revocable at any moment, distinct from failure, and distinct from a complete shutdown.
When regulators moved to suspend the deployment, the practical option set for both regulator and operator collapsed toward full operation or full halt. Intermediate postures such as operating on a curated subset of routes, operating only in daylight, operating with elevated pre-authorization for unusual maneuvers, or operating in a mode where commands are computed but not committed to actuators were not first-class operating states with a defined, auditable transition path. They were ad-hoc engineering efforts, each requiring fresh validation to be safety-defensible. That is not a criticism unique to Cruise; it is a general property of how production AV stacks are built and certified. The architecture treats execution as the default and a monitor-triggered stop as the exception, rather than treating execution itself as a permission that must be continuously satisfied.
The gap matters because incidents are a certainty across any AV deployment. The question an architecture answers is what response it can structurally support short of revocation. When the only defined states are run and halt, the pressure on a regulator bends toward suspension, because the alternative is unverified continuation. This produces predictable dynamics: boom-and-bust deployment cycles, capital that grows risk-averse toward driverless infrastructure, and a public conversation that flattens to a single yes-or-no axis rather than a graduated trust posture of the kind aviation and rail have long used. Closing the gap is not a tuning problem inside the planner or controller. It requires a structural stage sitting between intent and actuation that can withhold execution authority while cognition continues, and that treats that withholding as a governed, recoverable state rather than a fault.
3. What Confidence Governance Provides
The Confidence Governance inventive step, disclosed in United States Patent Application 19/647,395, addresses exactly this axis. In an embodiment, execution is treated as a revocable permission rather than a default assumption: a conditional privilege that must be continuously earned by the agent's demonstrated sufficiency across a plurality of evaluated dimensions, rather than a mode that runs until an external failure interrupts it. The enforcing mechanism is a confidence governor, a structural subsystem that continuously evaluates whether the conditions for execution remain satisfied and withdraws execution authorization when they no longer do.
The confidence governor is specified as a hard gate. In an embodiment, it is not an advisory module, a monitoring dashboard, or a soft constraint the system may override through urgency or intent priority. When the governor determines that authorization should be withdrawn, execution ceases, and the system cannot restore it through self-assessment, affective escalation, or policy reinterpretation. Execution authority is granted and revoked only by the governor, and no alternative pathway to execution bypasses it. The governor operates as a composite admissibility evaluator alongside an integrity engine and a capability envelope, requiring concurrent satisfaction of confidence sufficiency, integrity compliance, and capability confirmation before a proposed action is admitted, so admissibility never reduces to a single dimension.
Confidence is disclosed as a first-class computed state variable rather than a heuristic score or metadata annotation. In an embodiment, it occupies a designated confidence field in the agent schema, is computed from structured inputs derived from agent state and task state by a defined evaluation function, and is a continuous scalar between complete assessed insufficiency and complete assessed sufficiency. Because it is continuous, the governor can implement graduated response thresholds, early-warning behavior, and differential treatment based on the magnitude and rate of confidence change. In a further embodiment, the governor performs differential rate analysis on the confidence trajectory and, when a projected trajectory indicates that confidence will cross the authorization threshold within a configurable safety margin, initiates a graceful, preemptive suspension before the crossing occurs. This preemptive, trajectory-based suspension is the structural counterpart to what an AV stack lacks: the ability to step down before, rather than react after.
Execution authorization gating is disclosed as operating in one of three states. In an embodiment these are: authorized, in which confidence is above the threshold and the trajectory triggers no alarm, and execution is permitted; suspended, in which confidence has fallen below the threshold or the trajectory has triggered a preemptive suspension, and execution is prohibited while cognitive processes continue; and locked, in which a severe integrity violation, catastrophic resource failure, or governance-mandated halt restricts both execution and certain cognitive processes pending external review. The suspended state is the architecturally significant one: the system enters a non-executing cognitive mode in which it can forecast, plan, inquire, and self-assess while its ability to actuate is structurally removed by decoupling the execution subsystem's output pathway. Suspension is thereby distinguished from failure. Recovery from suspended back to authorized requires confidence to exceed the authorization threshold by a configurable hysteresis margin, preventing oscillation when confidence fluctuates near the threshold. Recovery from locked is not reversible by the system itself and requires external authorization.
In a further embodiment, the governor applies task-class differentiation on suspension. Terminal tasks, characterized by high irreversibility and low tolerance for state corruption, are halted at the earliest safe point with partial progress preserved in a durable, governance-tagged checkpoint. Exploratory tasks are redirected toward hypothesis expansion. Generative tasks are transitioned to a lower-commitment mode that produces candidate directions without finalizing any. In an embodiment, task-class assignment is recorded in lineage and is auditable. The approach is disclosed as technology-neutral with respect to the underlying perception, planning, and control implementation, which lets a conforming system retain its domain stack and add the governor as a structural stage between intent and actuation.
4. Composition Pathway
A driverless AV operator, or the AV category more broadly, composes with Confidence Governance by running its domain-specialized perception-planning-control stack over the confidence-governed execution substrate rather than by re-architecting it. What stays with the operator: the sensor hardware, the perception and prediction models, the planner cost functions, the simulation infrastructure, the remote-assistance operations center, the fleet-management plane, the safety-case engineering, and the entire commercial relationship with riders and municipalities. The operator's investment in domain knowledge, urban driving patterns, construction-zone handling, and emergency-vehicle interaction remains its differentiated layer.
What the substrate adds is the gate. In an embodiment, the planner emits actuation intents to the confidence governor rather than directly to the controller; the governor computes confidence from agent-state and task-state inputs, evaluates it against the authorization threshold and against the projected confidence trajectory, and either admits execution in the authorized state or holds the system in the suspended, non-executing cognitive mode while planning and inquiry continue. Because the execution subsystem's output pathway is structurally decoupled during suspension, the controller cannot actuate while suspended regardless of the urgency of intent. Transitions among authorized, suspended, and locked, and the hysteresis-gated recovery from suspended to authorized, are recorded in the agent's lineage. A trajectory that would previously have run until a monitor-triggered stop is instead stepped down preemptively when the confidence trajectory forecasts insufficiency.
The commercial surface this opens is a defined, auditable middle posture. Rather than choosing between full authority and revocation, a deployment can, as a governed and recoverable condition, hold execution in suspension while cognition continues, recover only across a hysteresis margin, and preserve a lineage record of every transition. This article does not assert that any particular regulator would accept such a posture; the regulatory framing is external context. The claim grounded in the filing is the architecture: execution as a revocable permission, a hard confidence gate, three authorization states with hysteresis recovery, and preemptive trajectory-based suspension into a non-executing cognitive mode.
5. Enablement and Embodiment Scope
A skilled implementer could build the disclosed approach on a conventional AV or robotics stack. The confidence field is a computed scalar written by an evaluation function over structured inputs drawn from agent state and task state; the governor is a subsystem that compares that scalar and its differential against an authorization threshold and a trajectory margin; the three-state gate is a state machine over authorized, suspended, and locked with defined transition rules; the non-executing cognitive mode is realized by decoupling the execution subsystem's output pathway so cognition can continue while actuation cannot; and hysteresis recovery is a threshold-plus-margin condition on the suspended-to-authorized transition. Each of these maps onto standard software and control primitives without requiring an exotic component.
The disclosure contemplates variation across embodiments. The confidence evaluation function may weight agent-state and task-state inputs differently across domains. The authorization threshold, hysteresis margin, and trajectory safety margin are configurable rather than fixed. The composite admissibility evaluation may combine confidence with integrity and capability signals under domain-specific policy. Task-class differentiation may recognize terminal, exploratory, and generative classes and hybrids inheriting the most conservative constraints of their contributing classes. The locked state may be triggered by integrity violation, resource failure, or governance-mandated halt, with recovery requiring external authorization. The approach applies beyond road vehicles to any cyber-physical or software agent whose execution should be gated on continuously computed sufficiency, and it composes hierarchically so that governance can be applied at the level of an individual agent or across a coordinated set of agents.
6. Disclosure Scope
The inventive subject matter described here, the confidence governor, execution as a revocable permission, confidence as a first-class computed state variable, the composite admissibility evaluator, the three authorization states of authorized, suspended, and locked, hysteresis-gated recovery, preemptive trajectory-based suspension, the non-executing cognitive mode, and task-class differentiation, is disclosed in United States Patent Application 19/647,395. This article is a dated public description of that subject matter tied to that filing.
References to Cruise LLC, General Motors, Waymo, the October 2023 San Francisco incident, California regulatory actions, and the driverless AV category are provided as external market and technical context to situate the disclosed architecture against real-world execution stacks. They describe third-party products and events accurately to the best available public record and are not claims of United States Patent Application 19/647,395. Named companies and products are the property of their respective owners. Nothing here asserts a defect in any named product beyond the general, architecture-level observation that conventional AV execution stacks do not expose execution as a continuously revocable, hysteresis-recovered permission with a non-executing cognitive mode of the kind the filing discloses.
