Vendor and Product Reality
Mobileye occupies a structurally privileged position in the autonomous-driving stack. Its EyeQ system-on-chip family sits behind the windshield in vehicles produced across a broad set of automakers, and Mobileye has publicly described an active REM-contributing fleet on the order of millions of vehicles. Each REM-enabled vehicle extracts compact semantic descriptors of road geometry, lane markings, signs, signal heads, and drivable-surface boundaries during ordinary driving and uplinks those descriptors to Mobileye's cloud within a deliberately small per-kilometer data budget. The cloud aggregator stitches descriptors from many vehicles into a vector map that Mobileye calls the Roadbook.
The Roadbook is redistributed to consuming systems, including Mobileye's own driver-assist and autonomy stacks and third parties that license the map data. Mobileye's public materials describe map-data and related data services as a business line distinct from silicon and system sales. The product is real, the contributing fleet is real, and no competing crowdsourced HD-mapping program approaches REM's contributing-vehicle count. Nothing in this article disputes that. REM is a strong, mature mapping substrate, and the Forecasting Engine described below is a consumer-side governance layer that presupposes a good map rather than competing to produce one.
Scoping the Comparison Honestly
It would be a straw man to fault REM for lacking a planning-governance mechanism. REM is a mapping program, and a mapping program is not supposed to arbitrate a downstream agent's speculative reasoning. The Roadbook is best understood, from the Forecasting Engine's point of view, as a source of verified prior state: an accurate description of the road as observed. The architectural question the Forecasting Engine addresses lives one layer up, inside whatever autonomy or driver-assist agent consumes that map.
That consuming agent must do something REM does not attempt and does not claim to do. Given the map as ground state, the agent must project multiple hypothetical futures, alternative maneuvers, other vehicles' likely trajectories, contingencies for occlusions and construction, and it must do so without any of those speculative projections silently being treated as fact. A map cell is verified. A projected maneuver is not. Conflating the two is the failure mode that a planning-governance layer exists to prevent. REM does not address that layer because it is not REM's job. The Forecasting Engine of 19/647,395 is a substrate for exactly that layer, and the comparison is scoped strictly to it.
What the Forecasting Engine Provides
The Forecasting Engine disclosed in 19/647,395 is an agent-level (or zone-level) cognitive substrate that constructs, evaluates, modulates, and manages planning graphs. It is not an external service, a shared map utility, or a centralized scheduler; it operates on the consuming agent's own state under the agent's own policy constraints. Applied to a vehicle autonomy stack consuming a map such as the Roadbook, its architecture supplies four elements that a mapping substrate does not.
First, structural separation of speculation from verified state. The spec makes planning graphs an architectural invariant distinct from the agent's verified execution memory. A branch that projects a successful maneuver does not thereby make that maneuver true; the projection remains speculative until it is promoted through governance. Snapshot isolation means a planning graph reads verified state (including the current map) as its root node but is not perturbed by concurrent changes, so evaluations are deterministic with respect to the state at construction time.
Second, branch classification. The engine classifies each speculative branch as eligible, introspective, delegable, or pruned. An eligible branch has passed slope validation and policy compatibility and is a viable promotion candidate; an introspective branch is structurally viable but affectively deprioritized and retained for self-examination rather than execution; a delegable branch transfers to a child planning graph; a pruned branch is removed. This gives the consuming agent an explicit, auditable taxonomy over its own hypothetical maneuvers.
Third, a containment boundary with immutable speculative markers and a delusion boundary. Planning-graph content is tagged at instantiation with a speculative marker that cannot be modified from inside the speculative domain. The promotion interface strips that marker and retags content as verified only after governance validation. The delusion boundary is the architectural guard against the pathological condition in which speculative content is treated as verified reality, precisely the condition an autonomy stack must never enter when it plans over a live road map.
Fourth, confidence-gated dispatch. A branch does not execute because it was projected; it executes only when it is the highest-ranked eligible branch and passes the promotion interface's governance evaluation, which is not waivable by the agent's affective state, personality configuration, or operational urgency. Forecasting is speculative and structurally contained throughout; it never becomes action until governed dispatch.
Composition Pathway With REM
The relationship is compositional, not competitive. REM continues to produce and license the Roadbook exactly as today. The Forecasting Engine sits in the consuming autonomy or driver-assist agent, treating the Roadbook (or any comparable map) as verified prior state and applying contained speculative planning above it. Where the map supplies what the road is, the Forecasting Engine governs how the agent reasons about what to do next without letting that reasoning contaminate the map-grounded verified state it acts on.
The spec also discloses a vehicle-adjacent embodiment that clarifies the boundary. In a fleet-level affective-state aggregation embodiment, autonomous vehicles share affective-state metadata and a fleet policy coordinator adjusts fleet-wide policy bounds for defined geographic regions or road segments, with the explicit constraint that fleet-level adjustment operates on policy bounds and never overrides an individual vehicle's safety governance. This is a governance-of-planning mechanism, not a mapping mechanism, and it composes above a map substrate rather than replacing one. It is disclosed as an embodiment of the platform, not as a description of REM.
Enablement and Embodiment Scope
A skilled implementer can build the approach described here from the disclosure. The core is a substrate that (a) maintains planning-graph structures in a computational domain separate from verified execution memory, (b) tags all planning-graph content with an immutable speculative marker at instantiation, (c) classifies branches as eligible, introspective, delegable, or pruned through slope validation, policy compatibility, affective prioritization, and pruning, (d) exposes a single promotion interface that strips the speculative marker and commits a branch to verified memory only after governance validation, and (e) refuses any alternative speculative-to-verified pathway. Confidence-gated dispatch follows from ranking eligible branches and admitting only the leading candidate that passes governance.
The approach admits broad variation. The forecasting engine may be instantiated at the agent level or the zone level. Branch expansion may be modulated by personality and affective fields, compressing planning horizons under temporal pressure or broadening search under novelty appetite. The verified prior state at a planning graph's root may be any map or world model, an HD vector map such as the Roadbook, a locally built occupancy map, or a fused multi-source model; the containment and promotion guarantees are independent of the map's provenance. Delegable branches may spawn child planning graphs for multi-agent coordination. Fleet embodiments may aggregate affective state to adjust policy bounds per road segment while preserving individual-vehicle safety governance. None of these variations relaxes the containment boundary, the immutable speculative marker, or the single governed promotion pathway.
Disclosure Scope
The mechanisms attributed to the invention in this article, structural separation of planning graphs from verified execution memory, immutable speculative markers, the containment and delusion boundaries, branch classification into eligible, introspective, delegable, and pruned, governance-gated promotion, and confidence-gated dispatch, are disclosed in United States Patent Application 19/647,395. This article is a dated public description of that disclosure and its embodiments as of its publication date.
All statements about Mobileye, REM, EyeQ, and the Roadbook are provided as external market and technical context based on Mobileye's public descriptions of its products. They are not claims of United States Patent Application 19/647,395, and nothing here should be read as asserting that the filing covers, incorporates, or limits Mobileye's technology. REM is characterized only as an accurate, mature mapping substrate; the invention is positioned solely on the contained-speculation and governed-dispatch axis and does not assert any deficiency in REM's mapping function. Product and company names are the marks of their respective owners and are used for identification only.
