Tesla FSD Supervised Lacks Structurally-Defensible Stage Architecture
by Nick Clark | Published April 25, 2026
Tesla FSD Supervised operates the largest commercial L2 supervised-autonomy deployment in the world, with V12 and V13 end-to-end neural networks executing on Hardware 3 and Hardware 4 across millions of vehicles and several billion cumulative miles. The architectural element conspicuously absent from the FSD stack — a structurally-defensible stage architecture that gates the transition from L2 supervised operation to L3 conditional autonomy and on toward L4 driverless Robotaxi service — is precisely what governed actuation supplies.
FSD Reality
Tesla FSD Supervised is, as of V13 and the early Robotaxi pilot rollouts, the largest commercial deployment of supervised driving autonomy ever fielded. The vehicle fleet comprises millions of cars on Hardware 3 and Hardware 4 compute, with V12 having transitioned the stack to a single end-to-end neural network for planning and control and V13 extending that architecture with higher-frame-rate perception, expanded training data, and tighter integration with Tesla's Dojo and HW4 inference pipelines. Cumulative supervised mileage now exceeds several billion miles, and Tesla discloses crash-rate telemetry against human baselines on a recurring cadence.
Despite this scale, the stack operates under SAE Level 2 categorization: the human driver is at all times the legally-responsible operator, and the vehicle requires hands-on-wheel and eyes-on-road monitoring enforced by cabin cameras and steering-torque sensing. The transition Tesla has publicly committed to — first toward an unsupervised Robotaxi service in geofenced metropolitan areas, then toward broader L3 conditional autonomy on consumer vehicles — collapses a regulatory and architectural gap that the present FSD codebase does not structurally address. The neural-net policy treats supervised and unsupervised operation as a continuous distribution rather than as discrete commitment regimes, and the handoff between human and machine authority is implemented as a UI-level alert rather than as a credentialed state transition with reversibility guarantees.
Tesla's Robotaxi service, now operating in limited geofenced deployments in Austin and the Bay Area, exposes the gap. Each Robotaxi ride is, by NHTSA and state-DMV framing, a commercial driverless operation — but the underlying stack is the same end-to-end network used in supervised consumer FSD, distinguished operationally only by remote-monitoring overlays and route constraints. The architectural delta between supervised consumer FSD and unsupervised Robotaxi operation is therefore policy and process, not structure. That asymmetry is what regulators, plaintiffs' counsel, and insurers interrogate first, and it is what governed actuation addresses.
Stage Architecture
Governed actuation is the primitive of graduated actuation modes — a structural commitment that every actuation event resolves through one of a fixed set of named, credentialed stages, each with declared reversibility properties, declared authority bindings, and declared rollback semantics. Applied to FSD, the primitive maps cleanly onto the L2-to-L3-to-L4 transition that Tesla's product roadmap requires. A supervised lane change executed under L2 is not the same kind of event as a Robotaxi unsupervised lane change executed under L4 service, even when the underlying neural-network output is bit-identical: the two events differ in who is the legally-responsible operator, what reversibility window applies, what telemetry must be retained for post-hoc adjudication, and what handoff path is admissible if the vehicle exits its operational design domain.
Stage-gated commitment provides the substrate for this distinction. Each actuation — steering torque application, throttle command, brake command, lane-change initiation, intersection traversal — is admitted through a stage descriptor that names the active autonomy tier, the credentialed operator (human driver, remote monitor, or autonomous policy under defined ODD), and the reversibility class of the action. Reversibility-aware execution maps onto transition-class actuations: events that change which party holds operational authority, such as an L3 autonomy disengagement or an L4 minimal-risk-condition pullover, become structurally distinct from steady-state actuations and carry their own admissibility and audit obligations.
Crucially, the primitive does not require Tesla to abandon end-to-end neural-network control. The neural network continues to produce the actuation policy; the stage architecture wraps the policy output in a credentialed envelope that records, for every actuation, which stage was active, which operator was bound, and which reversibility class applied. The envelope is what makes the same neural-network output legally and architecturally distinguishable across L2 supervised, L3 conditional, and L4 driverless contexts.
Tesla Position
Adopting governed actuation gives Tesla three concrete advantages in the L3-and-beyond regulatory engagement that is now imminent. First, the stage descriptors provide a defensible audit trail for NHTSA and state-DMV inquiries: every actuation event is reconstructible against the stage that authorized it, and the question of whether a crash occurred under supervised or unsupervised operation becomes a structural query rather than a forensic reconstruction. Second, the reversibility-aware execution layer provides the structural basis for L3 minimum-risk-maneuver compliance under emerging UNECE and FMVSS frameworks, which require that a conditional-autonomy system be able to bring the vehicle to a safe state within a bounded time window when the human driver fails to take over.
Third, and most importantly for the Robotaxi rollout, the primitive provides a clean architectural separation between the consumer-FSD product line and the commercial-Robotaxi service line. Today the two share a stack and are distinguished by operational policy; under governed actuation they share a stack and are distinguished by stage credentials, with the credential bindings themselves being auditable artifacts. That separation is what allows Tesla to expand Robotaxi geofences, raise consumer-FSD autonomy tiers, and respond to regulatory interrogation along three independent axes rather than as a single coupled product. The architectural-compliance substrate is, in short, what converts Tesla's scale advantage in supervised miles into a defensible position for the unsupervised transition.
The product-management consequence is equally direct. Tesla's current FSD product roadmap couples consumer-vehicle autonomy upgrades, Robotaxi geofence expansion, and Hardware 4 retrofit programs into a single release cadence, with regulatory engagement managed reactively across the bundle. Under governed actuation, each axis has an independent credential surface: a consumer-vehicle V13.x release exposes new actuations under the L2 supervised stage, a Robotaxi service expansion binds the same neural-network outputs under L4 driverless stage credentials with named remote-monitoring authorities, and an L3 conditional-autonomy pilot in selected jurisdictions binds a third stage with its own minimum-risk-maneuver and handoff-window obligations. The three axes can move at independent cadences without architectural coupling, and the regulatory conversation in each jurisdiction can engage the relevant stage credentials directly rather than requiring an end-to-end review of the FSD stack. That decoupling is the structural prize.
