Bounded LLM Reasoning: A Neuro-Symbolic Blackboard Architecture for Production Document Analysis

We present a neuro-symbolic blackboard architecture for production document reasoning that achieves perfect ledger reconstruction accuracy at roughly one-twenty-ninth the LLM token cost of an orchestrated multi-LLM baseline. Originating from production work on TaxBridge — a UK Making Tax Digital compliance engine — the architecture inverts the dominant pattern in recent LLM multi-agent systems: rather than coordinating LLMs through structured frameworks, we deploy LLMs as bounded heuristics that fire only when symbolic Knowledge Sources cannot resolve ambiguity. The architecture is parameterised across four declarative axes — evidence hierarchy, evidence polarity, plugin registration, and focus-of-attention scheduling — and includes a Macro-Micro-Macro opportunistic refinement pattern that exploits monotonic tick ordering to combine global LLM normalisation with zero-cost symbolic classification. We evaluate on two corpora: TaxBridge-Bench (a held-out set of 30 adversarial synthetic workbooks with offset tables, native formulas, and visual metadata) and a 45-case panel corpus reflecting real-world sole-trader bookkeeping diversity. Against the system it replaced in production — a four-role LLM debate panel — the blackboard achieves 100% ledger accuracy versus 0%, at 26× lower latency and 29× lower token consumption. On the panel corpus, the Macro-Micro-Macro pattern fires on 38% of cases (the messier, bank-export-style workbooks) and the system classifies 99.6% of transactions. Memory management is implemented as a Knowledge Source, enabling deterministic compaction that respects level-schema persistence policies — a contrast with LLM-driven cleaner agents inappropriate for regulated domains. We argue that this inversion — symbolic-first with LLM fallback — is the economically rational and regulatorily defensible architecture for high-stakes document reasoning.