What we are working on right now

Five threads on the bench.

Energy-transition model interoperability

An energy-models lab, oriented toward IIASA collaboration. One frozen benchmark system implemented in three energy-transition optimization engines — OSeMOSYS, PyPSA, and MESSAGE-ix — then used as the substrate for a series of Markov-blanket experiments: what crosses a model’s boundary, under which declared conventions, and what that does to a consumer on the other side. The findings so far have teeth: three engines agree on the physics to numerical precision yet report three different CO2 prices, each the exact dual of its own objective — the spread is pure capital-annualisation convention, carried through the boundary by monetary variables. Temporal resolution turns out to change the answer class, not just the precision, and geospatial support is itself a boundary variable that must be declared alongside its discretization. All three benchmarks reproduce byte-identically in containers, and the engines are wrapped as MCP tools behind a single blanket protocol.

Voice-data pipeline for Indigenous language preservation

voice-dataset-creation: open-access, end-to-end. Record or digitize existing audio, segment, SNR quality-check, transcribe locally (Whisper / MMS / manual), review, export to LJSpeech for TTS training or archival. Following the OCAP and CARE datasheets, the pipeline binds Indigenous data governance terms to the data itself rather than to a separable policy document.

Agentic coalitions in Rust

koalisi: an event-sourced temporal hypergraph topology with time-travel queries, coalition value and partition algorithms, and a tokio task-based runtime with supervised shutdown and restart. The live question is decision policy: active inference versus categorical magnitude, run head-to-head as pluggable strategies with a pre-registered A/B evidence trail.

Category-theoretic graph structures

catgraph: a strict Rust implementation of Fong & Spivak’s hypergraph categories (cospans, props, Frobenius), with applied-CT, magnitude, Wolfram-physics, and categorical-deep-learning extensions, plus a term-language frontend that prints, parses, and evaluates. It supplies koalisi’s topology layer — the theory earning its keep in a runtime.

Active inference engine

tira: a Rust reproduction of Waade et al. (Entropy 2025) on how collectives of active-inference agents in a Markov-blanket arrangement constitute a group-level agent with an emergent generative model; all four of the paper’s qualitative findings reproduce. Beyond the reproduction, the engine is packaged as a coalition-formation value primitive — the active-inference side of koalisi’s A/B.