METASTATE rests on three preprints that together define the running system. Paper I specifies the detection architecture; Paper II generalises it into the full operating substrate; Paper III is the engineering specification of the deployed system — every endpoint, contract, protocol, and data model, mapped live-versus-roadmap.
METASTATE: A System Specification for a Free-Energy Agent Substrate with Verifiable Inference, Post-Quantum Identity, and a Quantum-Secure Settlement Layer
Ciprian Pater · NWO · Imperium Romanum · NWO Agentic · Preprint Rev 1, June 2026. Documents the deployed system: the kernel, the complete API surface, the shared EML memory graph, Computational Proof-of-Inference, post-quantum signing, the two verified Base contracts and the keccak256 anchoring protocol, plus the hardware-pending interfaces — every capability tagged live or roadmap.
Read the specification on ResearchGate →METASTATE: A Free-Energy Operating Substrate for Anomalous-Signal Detection, Transcendental Symbolic Inference, and Harnessed Agent Economies
Ciprian Pater · NWO · Imperium Romanum · NWO Agentic · Preprint Rev 1, June 2026. Establishes the eml operator as a transcendentally-complete basis for minimum-description-length inference; recasts programs as morphisms in a causal-structure category; argues process matrices map onto photonic hardware with favourable scaling; and develops the mandatory transparency harness for autonomous agents.
Read METASTATE on ResearchGate →A Free-Energy Architecture for Anomalous-Signal Detection
EML Symbolic Regression · Process-Matrix Channels under TimesFM Priors · Ciprian Pater · NWO Capital · NWO Robotics · Preprint Rev 2, May 2026. Establishes the three-term free-energy functional the substrate is built on.
Read Paper I on ResearchGate →Under the Free Energy Principle, any self-organising agent minimises variational free energy — equivalently, performs approximate Bayesian inference. METASTATE treats anomaly detection as exactly this: a signal whose statistical signature resists the agent's hierarchical generative model produces high free energy. That is the anomaly score.
The novelty is that the three architectural pieces are not a loose pipeline. The temporal prior is the accuracy term; the process-matrix recognition density is the recognition-complexity term; and the symbolic head's depth penalty is the structural-complexity term. Minimising one functional trains all three.
Mathematically, free-energy regression over the transcendentally-complete eml basis is shown to be minimum-description-length inference, making transcendental complexity measurable via tree depth. Ontologically, a recognition density admitting indefinite causal order means a computation can be well-defined yet have no screening state, recasting programs as 2-morphisms in a causal-structure category. Epistemologically, the same free energy that scores anomalies bounds what an agent can know about its own inference — an intrinsic transparency measure, self-opacity O(t). Physically, the linear-positivity structure of process matrices maps onto photonic linear optics with a favourable scaling path versus the gate model. These combine in an agent economy where transparency is a cryptographically-enforced precondition of action.
A process matrix W is the most general linear-positivity-constrained family for cross-temporal correlations. When the data admit a definite-order fit, the learned W stays in the causally-separable cone. When they don't, the model is forced into the non-separable region and the coherence metric 𝒞(t) fires — the empirical analogue of a causal-inequality violation. No claim is made that the data are physically quantum; the borrowing is for expressive capacity.
Information from any organised process shares signatures regardless of source: a Zipfian frequency law (α≈1), a structured conditional-entropy hierarchy, compressibility below the raw entropy rate, and 1/fβ spectral self-similarity (β≈1). METASTATE computes all four on every window and reports them alongside free energy and coherence.