The Free-Energy Architecture of Discovery-Capable Intelligence

Abstract

Large language models optimise fluently inside a fixed representational space but do not reliably produce genuine discovery; the field has treated this as a tuning problem. We argue it is an architectural one, and we derive the required architecture from first principles rather than inferring it from current systems. The derivation rests on a single, established formalism used in its exact form: variational free energy, F = complexity − accuracy, which is simultaneously a thermodynamic quantity (U − TS) and a statement about how well an internal generative model fits the world under an Occam penalty. Read this way, intelligence is the minimisation of free energy: the pruning of the space of conceivable representations by the joint constraint of logic (the prior) and the physical world (the likelihood), with perception and action as the two variables that lower the same scalar. We discipline the claim with the Landauer test — the standard by which "information" earns causal status only when an abstract property forces a substrate-independent physical consequence — and we bound it with highly optimised tolerance (the conservation of vulnerability): an optimised system cannot abolish fragility, only relocate it to the edge of what its model covers, so a discovery-capable architecture must make that boundary explicit rather than fail at it silently. A note on rigor and terminology. The paper's physical claims are not rhetorical flourish, and they are not softened into mere analogy: each is held to the Landauer test of §2, which admits an abstract, information-theoretic quantity as physically causal only when it forces a substrate-independent physical consequence. Where a claim meets that bar we state it as physics; where itdoes not, we flag it as correctly applied statistical mechanics of a representational system, not a new physics of information. Entropy is used throughout in its standard sense — representational disorder, measured where stated. This theory makes four architectural demands, and we discharge each against an auditable engine, SEMPL, a constrained-natural-language process engine whose generative model is declared — which lets us demonstrate the free-energy principle in its falsifiable form, a claim neither the neuroscience nor the LLM literature can cleanly make. We give: (i) a single free-energy objective over the engine's repair operation that reproduces, rather than replaces, the programme's prior measured results — the deterministic collapse of a permuted process's ordering entropy to zero becomes an accuracy term driven to zero, and the reduction of a domain's declared structure to its minimal load-bearing residue becomes a complexity-term delta; (ii) program-structural self-maintenance, in which the engine models its own source and surfaces a real module-dependency cycle and a real generality gap in its own ordering machinery; (iii) composable discovery methods that construct process structure (the action face of active inference), bounded and bridge-free; and (iv) a coverage-frontier instrument that flags out-of-model inputs as boundary conditions rather than returning them as confident answers. Around these sits a governed discovery loop — inject → generate → verify → repair → adopt — whose defining property is a machine-checked, byte-identical verify-before-adopt gate; it is the reproducible, formal complement to expert-in-the-loop frontier generators. We then prove the architecture at domain scale. We model a mid-size company's accounting function as ten inter-referable workflow processes sharing a general-ledger backbone; all ten repair to bridge-free free-energy fixpoints, and the family is a genuine network rather than isolated demonstrations. On it we perform the capstone discovery: finding a missing internal control by cross-process analogy — the engine identifies that one payment process discloses money with no independent approval, unlike its siblings, names the exact missing control, verifies in a sandbox that inserting it closes the gap bridge-free, and proposes it for human approval. This is a real audit finding and the concrete embodiment of the conservation-of-vulnerability thesis: the discovered gap is precisely the fragility displaced to where the control is absent. Beyond the demonstration, the architecture speaks directly to the enterprise agentic-AI question: a governed, inter-operable semantic substrate that ingests LLM-generated process representations turns a fleet of task-executing agents into a single auditable, self-maintaining, and self-auditing workflow — with simplification, efficiency, and controlled long-term evolution following from the theory rather than from ad-hoc engineering (nothing is more practical than a good theory). Every result is deterministic and reproducible, and is established with full reproduction artifacts in the companionworking papers rather than in this synthesis. We are explicit about what is not claimed: the verification is machine-checked but confined to a bounded formal representation, not frontier reach; the free-energy account is falsifiable here only because the model is declared; and live-generator quality evaluation is deferred. The contribution is a complete line from a physical principle of intelligence, through the architecture it necessitates, to a working, governed, reproducible proof that the architecture discovers something real.