Mechanism
The cognition filing introduces uncertainty as a first-class propagated variable that participates in capability determinations, temporal forecasts, and execution synthesis decisions alongside capability and time. Uncertainty, as the term is used in this disclosure, is not a residual or an error term. It is an explicit, bounded epistemic condition that the system computes, maintains, propagates, and consults at every stage of the capability evaluation pipeline. The three dimensions of capability, time, and uncertainty are not evaluated independently. They are evaluated as a joint condition that must be simultaneously satisfied for execution synthesis to proceed.
Uncertainty is represented as a structured bound associated with each capability determination and each temporal forecast. The uncertainty bound encodes the system's assessed confidence in its own evaluation. This is a deliberately narrow object: it is not the confidence of the executing agent, which is governed by the confidence governor disclosed in Chapter 5. The uncertainty bound evaluates whether the system's assessment that the agent can execute is itself reliable. The confidence governor evaluates whether the agent should execute. The uncertainty bound evaluates whether the system's assessment that the agent can execute can be trusted.
What the Uncertainty Bound Describes
The capability evaluation pipeline produces capability determinations and temporal executability forecasts. Each of these outputs carries its own assessed reliability. A capability determination of structurally possible can be held with high confidence or with substantial doubt, depending on how complete and current the substrate's capability envelope is. A temporal executability forecast is expressed as a confidence-bounded window rather than a point estimate: the system does not predict that capability will become available at a single time, it predicts that capability will become available within a window bounded by an earliest and a latest time, with a confidence level derived from the uncertainty model. The uncertainty bound is the structured record of that confidence.
Because the bound travels with the determination it qualifies, a downstream consumer never receives a bare capability or temporal result. It receives the result together with the system's own statement of how reliable that result is. This is what makes uncertainty a propagated variable rather than an internal quantity that is computed and then discarded once a determination has been emitted.
Three Epistemic Categories
The uncertainty model differentiates three categories of epistemic condition, each producing distinct system behaviors. The first is deterministic impossibility: the capability determination has resolved to structurally impossible with negligible uncertainty. The system has high confidence that no executable form of the objective can exist on the evaluated substrate. In this category the system does not defer or retry. It routes, decomposes, or reports non-executability with a deterministic confidence record.
The second category is deferred possibility with confidence bounds: the capability determination has resolved to structurally deferred, and the temporal forecast identifies a future executability window, but the forecast carries non-negligible uncertainty. The system has assessed that execution may become possible but cannot guarantee it. In this category the system may defer execution, but it must propagate the uncertainty bounds to the scheduling subsystem so that contingency plans, such as alternative substrates or decomposition strategies, can be prepared in case the forecasted window does not materialize.
The third category is indeterminate feasibility: the capability determination cannot be resolved with sufficient confidence due to incomplete information about the substrate's capability envelope, the objective's requirements, or the temporal evolution of the execution environment. In this category the system cannot make a reliable routing or deferral decision. It must either gather additional information, invoke a higher-authority evaluation, or report the indeterminate condition to governance infrastructure. The distinction between deterministic impossibility and indeterminate feasibility is structural: the first is a confident no, the second is an acknowledged inability to decide, and the two warrant different responses.
Propagation Through the Pipeline
Uncertainty is propagated through the system's computational pipeline such that downstream decisions inherit the uncertainty associated with their inputs. When a capability determination with non-negligible uncertainty feeds a temporal forecast, the resulting temporal forecast carries both the uncertainty of the capability determination and the additional uncertainty introduced by the temporal projection model. When a temporal forecast with compounded uncertainty feeds an execution synthesis decision, the synthesis decision is conditioned on the aggregate uncertainty and may be withheld if the aggregate uncertainty exceeds a configured threshold.
The defining property of this propagation model is that uncertainty is not silently discarded as it passes through successive computational stages. Uncertainty accumulates and remains visible at every decision point. The pipeline therefore makes progressively more conservative decisions as the basis for those decisions becomes less certain. A determination assembled from a long chain of uncertain inputs arrives at the synthesis gate carrying the compounded uncertainty of that chain, and the gate evaluates against the compounded value rather than against any single stage's local confidence.
The Uncertainty Threshold at Synthesis
Execution synthesis, the process by which the system constructs an executable form of an objective on a substrate, occurs only when three conditions are simultaneously met: the capability determination has resolved to structurally possible or has transitioned from structurally deferred to structurally possible upon arrival of the forecasted temporal window; the temporal executability window is currently open; and the aggregate uncertainty associated with the capability and temporal assessments is below the configured threshold. If any of the three conditions is not met, execution synthesis does not occur.
This places the propagated uncertainty bound on equal footing with capability and time as a gating condition. A substrate can possess the structural capability to execute an objective and present a temporally viable window, and synthesis can still be correctly withheld because the uncertainty associated with one or both of those assessments exceeds the threshold at which synthesis is warranted. The withholding is not an error. It is the structurally correct outcome of the capability-time-uncertainty evaluation.
The Uncertainty Ledger
The system maintains an uncertainty ledger: a structured record of the uncertainty state associated with each active capability determination, temporal forecast, and execution synthesis decision. The uncertainty ledger is persisted in the agent's lineage and is available to governance auditors. It enables retrospective analysis of decision quality. An auditor can ask whether the system made appropriate routing decisions given the uncertainty bounds available at the time, whether uncertainty thresholds triggered appropriate contingency behavior, and whether the system correctly distinguished deterministic impossibility from indeterminate feasibility.
Because the ledger records the uncertainty state at each stage rather than only the final outcome, it preserves the information needed to trace a withheld or deferred synthesis back to the stage whose uncertainty dominated the aggregate. This supports both post-incident review and proactive refinement of the projection and evaluation models that produce the bounds.
Recalibration of Forecast Uncertainty
The uncertainty bounds attached to temporal forecasts are not fixed. The system continuously compares forecasted temporal executability windows against actual executability observations, computing a forecast accuracy metric for each substrate and each capability dimension. When forecast accuracy for a given substrate-dimension pair falls below a configured threshold, the system recalibrates its temporal projection model for that pair by adjusting trend parameters, increasing uncertainty bounds, or replacing the projection model with a more conservative one.
Recalibration keeps the propagated uncertainty honest over time. As the operational environment changes in ways that invalidate the assumptions underlying the original projections, the bounds widen to reflect the reduced reliability, and the wider bounds then propagate forward into capability and synthesis decisions. An agent with elevated uncertainty sensitivity, drawn from the affective state field described in Chapter 2, applies wider uncertainty margins to temporal executability forecasts, effectively shortening the acceptable forecast horizon and favoring substrates with immediate executability over substrates requiring temporal deferral.
Distinction From Confidence Scores and Error Terms
Conventional systems treat the reliability of a feasibility assessment as an implicit quantity, computed internally and collapsed before a result is exposed, or as an undifferentiated failure when an assessment proves wrong. The present disclosure separates two questions that such systems conflate. The confidence governor of Chapter 5 answers whether the agent should execute. The uncertainty bound of this chapter answers whether the system's assessment that the agent can execute is itself reliable. Keeping these distinct allows the system to recognize that a confident decision to proceed can rest on an unreliable feasibility assessment, and to withhold synthesis on that ground alone.
The disclosure also separates a confident negative from an inability to decide. A system that reports only an undifferentiated failure cannot distinguish deterministic impossibility, where no future configuration of the substrate will satisfy the requirements, from indeterminate feasibility, where the system simply lacks the information to resolve the question. Because the uncertainty model classifies these conditions explicitly and propagates the classification forward, the agent that receives a non-synthesis determination receives useful structured information about why synthesis was withheld, enabling informed rerouting, deferral, decomposition, or objective revision rather than blind retry.
Disclosure Scope
Uncertainty as a first-class propagated variable, comprising the structured uncertainty bound associated with each capability determination and temporal forecast, the three epistemic categories of deterministic impossibility, deferred possibility with confidence bounds, and indeterminate feasibility, the propagation model in which uncertainty accumulates and remains visible as it passes from capability determination to temporal forecast to execution synthesis, the uncertainty threshold as a joint gating condition at synthesis alongside capability and time, the uncertainty ledger persisted in the agent's lineage, and the recalibration of forecast uncertainty against observed accuracy, is disclosed in the cognition filing (U.S. Application No. 19/647,395 and its international counterpart) at Chapter 6, with the uncertainty model described principally in Section 6.6 and its operation within capability-native computation, temporal executability forecasting, and execution synthesis described in Sections 6.4, 6.5, and 6.7. This article describes that disclosed mechanism. The scope extends to embodiments in which the uncertainty bound qualifies physical and biological capability assessments, where the disclosure notes that the bounds associated with physical capability dimensions are typically wider than those associated with computational dimensions, reflecting the greater epistemic uncertainty inherent in physical systems.
