Modern intelligence, surveillance, and reconnaissance (ISR) enterprises ingest multi-source indicator streams whose native tempos span four or more orders of magnitude. Overhead imagery (EO/IR, SAR, MSI, HSI) arrives at revisit cadences of minutes to hours; RF and SIGINT detections resolve in seconds to minutes; maritime AIS and acoustic returns update continuously at sub-second tempos; and open-source, supply-chain, and pattern-of-life indicators mature over hours to weeks. A single decision of operational consequence, such as whether a vessel of interest is a combatant or a civilian fisher, or whether an observed buildup presages a near-term operation or reflects routine activity, typically turns on whether causally related evidence drawn from several of these phenomenologies can be assembled, reasoned about, and surfaced fast enough and cheaply enough to matter.
Classical early- and late-fusion pipelines assume temporally aligned inputs and emit single snapshot outputs, useful when cadences match and brittle when they differ by orders of magnitude or when a depended-on sensor is denied. Reinforcement- and active-learning policies can select next observations but treat operational constraints as reward-shaped scalars the optimizer can trade away, and require retraining cycles measured in weeks when priority intelligence requirements or adversary tactics shift in hours. Neuro-symbolic and causal methods contribute structure and interpretability, yet in practice operate over hand-authored graphs and synchronous data, with post-hoc explanations rather than forward-looking traces. Agentic large-language-model orchestrators produce narrative rationalizations whose internal state cannot be audited, and whose plans cannot be inspected for constraint feasibility before execution. What is missing across all of these approaches is a structured, auditable, constraint-feasible substrate whose rationale for each recommendation is transparent, reproducible, and directly inspectable from internal state.
This topic seeks novel approaches to artificial metacognition: an orchestration substrate that reasons about its own reasoning while reasoning about the world. A metacognitive substrate maintains a structured model of evidence-to-outcome relationships across heterogeneous modalities and tempos; tracks its own uncertainty, prediction error, and representational coherence; computes expected information gain for candidate collection actions under hard operational constraints; returns option-sets rather than single collapsed policies; and, when denial, deception, or distribution shift is detected, locally refines its internal model rather than triggering a full retraining cycle. The resulting capability is the computational backbone of a human-machine formation: the autonomous system acts at machine speed inside a bounded operating envelope, while the human commander sets objectives, constraints, and rules of engagement before the mission and reviews the auditable decision record after.
Candidate substrate families for the metacognitive layer include hyperdimensional computing and vector-symbolic architectures; neuro-symbolic architectures that couple learned encoders with symbolic or probabilistic inference; and graph-based reasoning over typed knowledge graphs, causal graphs, and structured inference engines. Hybrids across these approaches are encouraged. This topic is deliberately substrate-agnostic within that class. Offerors are expected to justify their choice on quantitative grounds. Any viable substrate shall provide:
• Compositional, inspectable operators (algebraic, logical, graph-based, or probabilistic) that support real-time assembly and decomposition of multi-source evidence structures, rather than end-to-end learned embeddings whose compositional semantics cannot be examined.
• Bounded degradation under partial, corrupted, delayed, or adversarially manipulated input, quantified against a mainstream deep-learning baseline.
• Transparent, reproducible reasoning operations that yield auditable explanations directly from internal system state, rather than post-hoc narrative produced by a prompted language model.
• Efficient operation under sparse and few-shot regimes, with measurable performance at small label budgets on novel scenario classes.
• A credible path to tactical-edge deployment on Size-Weight-and-Power-constrained platforms.
Recent work across these areas has independently demonstrated subsets of these properties, including graph-structured anomaly reasoning, compositional multi-modal scene understanding, and interpretable edge AI with online adaptation.
Converging them into a single autonomous metacognitive orchestration substrate is the objective of this topic. Approaches that rely on monolithic end-to-end deep networks without compositional or causal structure, or on narrative-only explanations produced by prompted language models, fall outside the scope of this topic.
The program seeks novel approaches to key technical challenges, including:
• Representing and reasoning over time-lagged causal dependencies across multi-modal sensor streams whose native tempos span four or more orders of magnitude, under missing, late, corrupted, and adversarially manipulated inputs.
• Computing expected information gain over candidate collection actions while treating operational constraints (emissions, bandwidth, power, time-to-decision, access) as hard-binding semantic variables rather than reward-shaped scalars, and producing Pareto-optimal option-sets rather than single collapsed policies.
• Self-monitoring the system’s own confidence, prediction error, and representational coherence; detecting adversarial denial, deception, and distribution shift as learning signals; and triggering targeted local refinement of the internal model without offline retraining.
• Generating forward-looking “why / why-not” traces, reproducible from internal system state, that link each recommendation to its supporting evidence chain, the constraints it satisfies, the alternatives it dominated, and the expected futures it forecasts, supporting downstream review by a subject-matter expert (SME) or within a human-machine formation.
• Probabilistic root-cause analysis of observed outcomes, tracing back through the internal representation to identify contributing evidence across domains and timescales, with ranked alternative hypotheses and quantified confidence.
