Extending the BDI Abstract Interpreter for Stochastic Sensors

HADD: Extending the BDI Abstract Interpreter for Stochastic Sensors presents a conservative architectural extension of the classical Belief–Desire–Intention abstract interpreter proposed by Rao and Georgeff in “BDI Agents: From Theory to Practice” (ICMAS 1995). The original BDI interpreter established one of the most influential computational models for rational autonomous agents, organizing agent behavior around beliefs, desires, intentions, deliberation, commitment and execution in dynamic environments. However, as Rao and Georgeff explicitly acknowledged, their model relies on an implementation-level assumption regarding the reliability of sensors: beliefs are expected to provide information about the likely state of the environment, while assumptions such as sensor accuracy may be violated in practice. This work identifies that sensor reliability assumption as a critical architectural gap when modern autonomous agents use large language models as perceptual sensors. Unlike classical deterministic sensors, LLM-based sensors produce stochastic, fluent and plausible outputs that may contain unsupported claims, contradictions or hallucinated facts. In such systems, hallucination is not merely a generation error at the language level, but a structural risk: if an unverified LLM output is allowed to enter the belief store, it can contaminate downstream deliberation, intention formation and action execution. HADD — Hybrid Agents with Deterministic Decisions — addresses this issue by extending the Rao–Georgeff BDI loop without replacing its deliberative core. The proposed architecture introduces three precisely located components: an Epistemic Verification and Routing Gate Φ between perception and belief commitment, an EMRE epistemic calibration adapter for tenant-specific sensor reliability, and a Double Tribunal for epistemic and operational verification before execution. These components transform the original assumption of sensor accuracy into an explicit, auditable verification process. The central claim of the paper is that HADD does not attempt to make LLMs infallible. Instead, it converts hallucinated or unsupported perceptual outputs into detectable, traceable and auditable epistemic errors before they can become operational beliefs or executable intentions. In this sense, HADD protects the deterministic downstream behavior of the BDI interpreter from stochastic perception. The architecture preserves the core BDI invariants while strengthening the perception–belief and intention–execution boundaries. The paper formalizes this extension through a set of invariants, theorem statements and verification mechanisms. In particular, the Learning Boundary Theorem (T9) argues that EMRE calibrates confidence without modifying the deterministic BDI decision layer, while the Total Coverage Theorem (T10) states that, under a complete verifier and tribunal profile, every executed action is supported by epistemic grounding and operational legality. HADD is also extended to holonic multi-agent systems through the MINERVA runtime, where cognitive holons generate candidate outputs and reactive holons verify them semantically, executably or cryptographically. This cognitive–reactive pairing is used to localize hallucinations and convert them into structured audit records containing error type, segment, detector, layer, confidence and trace identifier. The work is validated through MINERVA v0.2, a production holonic multi-agent runtime for legal and financial document intelligence. The implementation processes multiple desire types, including document auditing, consistency validation, risk detection, classification, clause search, sensitive-data detection, summarization, translation and spreadsheet analysis. Empirical results report an average ReasoningVerifier score of 0.9430 across twelve desire types, with strict verification profiles for critical document operations and compliance checks such as J7 DORA and J8 Fairness. Overall, this technical report contributes a formal and operational bridge between classical symbolic BDI agents and contemporary LLM-based agent systems. Its main contribution is not the replacement of BDI, but the modernization of its perception boundary for a new class of stochastic, generative sensors. HADD reframes hallucination as a structural property of language-model-mediated perception and introduces an auditable mechanism to prevent such hallucinations from silently propagating into beliefs, intentions and actions.