1. The Light-Time Problem

Mission planning in low-Earth orbit can lean on a human in the loop. A controller in Houston or Darmstadt sees telemetry, evaluates options, and issues a go or no-go command with a round-trip delay of well under a second. That operating model degrades smoothly with distance and then breaks. At the Moon, round-trip light time is roughly two and a half seconds; at Mars at opposition it is on the order of minutes; in the outer solar system it is tens of minutes to hours. Long before those extremes, the delay exceeds the time available to react to a thruster underperformance during a burn, a sensor dropout during descent, or a debris-conjunction warning during proximity operations.

The conventional response is to pre-compute everything: ground teams generate trajectory options, contingency trees, and abort criteria before the critical event, uplink them, and let the spacecraft select from a fixed menu. This works when the anomaly space is known in advance. It leaves open the case where the spacecraft encounters a combination of conditions no pre-computed branch exactly covers, which is the case the architecture described here addresses: an on-board faculty for constructing, evaluating, and committing a plan that was not on the menu, together with a governed way to establish that a self-generated plan is validated for execution rather than a projection treated as reality.

The requirement, then, is not faster optimization. It is an on-board architecture that can hold many candidate trajectories and abort scenarios in parallel, evaluate each against physical constraints and mission policy, and commit one to execution under a defined gate, with the speculative options kept structurally separate from the maneuvers the spacecraft actually flies.

2. The Architectural Requirement

Space mission planning has three properties that the described planning substrate satisfies structurally rather than procedurally.

First, parallel alternatives with independent state. A descent profile, a powered-flyby option, and an abort-to-orbit branch each carry their own projected state vector, propellant ledger, and constraint evaluation. Collapsing them into a single weighted estimate destroys exactly the information an abort decision needs, namely how each distinct option behaves at the margin.

Second, a clean separation between what has been projected and what has been committed. A spacecraft that confuses a forecast state for its true state executes maneuvers predicated on conditions that exist only inside a simulation. Orbital mechanics are unforgiving, propellant is finite, and most maneuvers are irreversible. The disclosure treats the separation between speculation and verified state as a structural invariant of the substrate rather than as a coding convention subject to a concurrency bug or a corrupted flag.

Third, a governed transition from candidate to committed. In the described embodiments, promotion of a branch to execution passes a defined validation gate and leaves a record of why this branch, against which observations, under whose authority. Where an anomaly response is otherwise reconstructed after the fact from telemetry, an architecture that emits that record by construction supplies the same evidence directly to safety review and accident investigation.

The Forecasting Engine disclosed in United States Patent Application 19/647,395 was designed around exactly these three properties. The remainder of this article maps its disclosed mechanisms onto the mission-planning domain.

3. The Pre-Computed-Menu Model and Its Coverage Assumption

The pre-computed-menu model is procedurally rigorous and bounded by what was anticipated. Ground teams produce abort criteria, trajectory-correction-maneuver plans, and contingency trees, validate them through review boards, and uplink them. When the flight matches an anticipated scenario, the spacecraft selects the appropriate branch and the model works as intended.

The model degrades when the encountered conditions do not match any pre-computed branch precisely. Historic close calls and losses, from descent-engine anomalies to navigation-state errors to late-detected conjunctions, share a structure: a combination of intensity, timing, and geometry that no single uplinked plan exactly addressed. A human team, given enough light-time margin, improvises against the closest template. Beyond a few seconds of round-trip delay, there is no human team in the loop to improvise, and the spacecraft has only the menu.

Adding more pre-computed branches scales the same shape. The contingency space grows combinatorially, the uplink budget is finite, and validation effort scales with the menu. More fundamentally, a larger menu remains a selection among plans anticipated on the ground. The architecture described here adds a different faculty alongside it: an on-board capability for building, evaluating, and safely committing a new plan, with the same separation between speculation and verified state that a ground review board enforces by procedure but that an autonomous spacecraft must enforce by architecture.

4. What the Forecasting Engine Provides

The Forecasting Engine of United States Patent Application 19/647,395 specifies a planning graph as a first-class cognitive structure: a root node anchored to the agent's current verified state, with speculative branches representing candidate mutation sequences and their projected outcomes. Applied to a spacecraft, each branch is a candidate trajectory, abort scenario, or mission modification, carrying its own projected state vector, propellant and timing ledger, and constraint evaluation.

The disclosed forecasting cycle proceeds in six phases, which map cleanly onto mission planning. In initialization, the engine reads the spacecraft's current verified state and either refreshes an existing planning graph or instantiates a new one. In speculative mutation simulation, it projects each active branch forward, computing the candidate maneuver's consequences. In slope projection and validation, each branch is checked for continuity against the agent's governing trust constraints, so that branches which would produce a discontinuity, the analog of a maneuver that violates a hard constraint, are flagged ineligible. In the policy compatibility check, each remaining branch is evaluated against the applicable policy configuration; branches that violate policy are pruned. In emotional reinforcement tagging, branches are prioritized or deprioritized under affect modulation, with deprioritized branches retained for introspective analysis rather than discarded. In branch marking and pruning, each branch is classified.

The disclosed classification is structural, not a label. A branch that is slope-eligible, policy-compatible, and positively reinforced is marked eligible for promotion. A branch that is slope-eligible and policy-compatible but negatively reinforced is marked introspective and retained for further analysis. A branch suited for hand-off can be marked delegable. Branches that fail eligibility are pruned. For a spacecraft, this is the difference between an abort branch that is ready to fly, one that is viable but flagged for closer review, one that should be handed to a subsystem agent, and one that has been ruled out, with the reason recorded.

Two further disclosed mechanisms matter directly here. Confidence-gated dispatch means that when confidence in the current state is insufficient, the agent transitions to a non-executing cognitive mode and continues speculative planning without committing state changes. A spacecraft with a degraded navigation solution keeps forecasting and ranking options but does not fly a maneuver until confidence clears the gate. Personality-modulated and affect-modulated branch expansion means the breadth and risk posture of speculation are tunable parameters over an invariant architecture: a robotic precursor mission can run with broader speculative breadth, while a crewed vehicle can run with narrower breadth and higher promotion thresholds, without changing the underlying engine.

5. Containment: Holding Projections Apart From Flown Maneuvers

The safety property that makes this applicable to crewed and high-value missions is the disclosed containment layer and its delusion boundary. In the application's terms, the containment layer is an architectural boundary in the agent's cognitive substrate, not a software flag, a metadata annotation, or a runtime check, that keeps the speculative planning-graph domain separate from verified execution memory.

The layer enforces several invariants at once. Every element of a planning graph, every candidate maneuver, every projected state, every evaluation tag, is stamped with an immutable speculative marker at the moment of construction. In the described embodiments that marker is not removable, modifiable, or overridable by operations inside the planning-graph domain. Read isolation provides that the spacecraft's execution processes, the ones that actually command thrusters and reaction wheels, query verified execution memory and receive the true state rather than a projected value from an active branch. Lineage isolation ensures that speculative branches do not write to the agent's committed lineage; only governance-validated promotions are recorded as committed state, though the forecasting process itself, the creation, evaluation, and pruning of branches, may be recorded as cognitive events.

The pathway from speculation to execution is the promotion interface. It receives a candidate branch, subjects it to governance validation, and only on success strips the speculative marker, retags the content as verified, and writes it to execution memory. The disclosure describes this interface as the route by which a projected trajectory becomes a flown trajectory, which is the structural reason that in the described embodiments a spacecraft running this engine flies a maneuver only after that maneuver has left the forecast domain through the gate.

The application also specifies the failure mode and its remediation, which is precisely what a mission safety case needs. The delusion boundary is the formally defined pathological state in which containment collapses and speculative content is treated as verified reality, for example through a corrupted or stripped marker, a breached read-isolation boundary, or a promotion that admitted content without completing validation. The disclosed containment integrity verification mechanisms, periodic containment audits over speculative markers, boundary-crossing monitors, lineage consistency checks, and behavioral coherence monitors, are described as detecting collapse before it produces observable behavior. On detection, the disclosed containment restoration protocol suspends execution authority, quarantines the affected planning graphs, runs lineage forensic analysis, reconstructs verified state from the last governance-validated checkpoint, and re-initializes the containment layer with fresh markers. For a spacecraft, that is a defined safe-mode and recovery path grounded in the same architecture that produced the plan.

6. Executive Aggregation Across Subsystems and Vehicles

A spacecraft is a system of subsystems, and many missions are systems of spacecraft. The application's executive-graph layer addresses both. The disclosed architecture distinguishes micro-planning graphs, the per-agent structures described above, from macro executive graphs, zone-level or group-level structures that an executive engine produces through intersection detection and conflict resolution across multiple planning graphs.

Within a single vehicle, a guidance agent, a power agent, and a thermal agent each maintain a planning graph. When the guidance agent's preferred trajectory-correction burn conflicts with the power agent's safe-mode posture or the thermal agent's attitude constraints, the conflict surfaces as a branch incompatibility at the executive aggregation layer rather than as an emergent surprise during execution. Across a constellation or a multi-vehicle architecture, the same mechanism aggregates plans across vehicles, detecting conjunction risks and shared-resource conflicts and identifying coordination opportunities. The disclosed delegation and forking mechanisms let a parent agent hand a delegable branch to a child agent, the delegated content entering the child's planning-graph domain with its speculative marker preserved, which maps onto delegating a maneuver-planning sub-problem to a subsystem controller or a companion vehicle without ever transferring it as already-committed state.

7. Embodiments and Deployment Options

The architecture admits a range of embodiments across mission classes:

  • Cislunar and lunar surface. A lander maintains parallel descent profiles and an abort-to-orbit branch; as the navigation solution and hazard map resolve during descent, ineligible profiles are pruned, the surviving profile is promoted through the gate, and the abort branch remains contained and ready. Confidence-gated dispatch holds the vehicle in a non-executing planning mode if the state estimate degrades.
  • Deep-space cruise and flybys. The engine maintains trajectory-correction-maneuver options and a powered-flyby branch alongside a safe-mode coast branch, ranking them as the orbit-determination solution improves, with promotion thresholds set high to reflect the irreversibility of deep-space maneuvers.
  • Proximity operations and rendezvous. Candidate approach corridors and collision-avoidance branches are held in parallel; a conjunction warning elevates a contained avoidance branch to eligible and, on validation, promotes it, rather than triggering a from-scratch replan.
  • Robotic precursor versus crewed. Personality and affect modulation tune speculative breadth and promotion thresholds to the risk posture of the mission class over the same invariant engine.
  • Constellation and multi-vehicle. Executive aggregation reconciles per-vehicle planning graphs into a coherent group plan, surfacing conjunction and shared-resource conflicts structurally.

Across the described embodiments the same structure holds: the engine forecasts, ranks, classifies, and prunes within a containment boundary, and a maneuver reaches the thrusters by passing the promotion interface, which records the committed transition in lineage.

8. Adoption Pathway

Adopting the Forecasting Engine does not require replacing a spacecraft's flight software, guidance algorithms, or ground systems. The engine sits beneath them as a planning substrate. Existing flight dynamics tools, fault-management systems, and ground planning suites remain the surfaces through which engineers express constraints, observations, and abort criteria; what changes is that those surfaces read from and write to a planning graph with the containment, classification, and promotion properties described above, rather than to ad hoc per-tool state.

A pragmatic sequence begins with a non-safety-critical planning function on a robotic mission, where the contingency space is concrete and a self-generated, governance-promoted plan can be flown with ground confirmation still in the loop. Instrumenting that function with the forecasting substrate produces an after-the-fact reconstruction of the anomaly-response decision space that is generated by the architecture rather than inferred from telemetry: which branches were considered, why each was eligible or pruned, which observations supported promotion, and under what authority. The safety case for extending the substrate to higher-criticality functions, and eventually to crewed missions under light-time delay, follows from that record and from the disclosed containment-integrity and restoration guarantees.

9. Disclosure Scope

This article describes an application of the Forecasting Engine disclosed in United States Patent Application 19/647,395. The mission classes, regulatory context, and deployment scenarios discussed here are illustrative application framing and are external to the patent. The technical mechanisms relied upon, the planning graph and its six-phase forecasting cycle, branch classification into eligible, introspective, delegable, and pruned states, the containment layer with immutable speculative markers and the delusion boundary, confidence-gated dispatch, personality-modulated and affect-modulated branch expansion, the governed promotion interface, and executive-graph aggregation, are disclosed in United States Patent Application 19/647,395. Where this article describes orbital-mechanics behavior, abort logic, or mission-specific constraints, those are domain implementations of the disclosed architecture and are not themselves claimed by the cited application except insofar as they instantiate the disclosed mechanisms.