Current focus: Futaie β a local-first AI system for running agents in an AI Act-compliant, EU-sovereign deployment. The model, the language and the record all live on hardware the operator owns, and the account of what the system did is readable by the operator and by a regulator without asking anyone's permission.
An agent lives inside the machine, rewrites its own configuration and components, activates a new generation of the system and comes up in it. It can break the machine β that is a supported outcome. What it cannot do is decide what the record says happened. Futaie is the programme under which that agent improves itself along two axes: its environment, and its weights.
| lane | repo | what it is |
|---|---|---|
| the record | baliveau | A flight recorder for agents. The system's every state is a content-addressed generation in a build store; activation goes through PID 1 as a transaction; a failed boot rolls forward by itself. The seats that judge β the oracle asking is it better?, the audit asking is it honest?, the reporter asking is it documented? β run host-side, outside the machine the agent controls, and the agent cannot write them. A log the subject can rewrite is not a log. |
| the language | osier | The language the agent authors components in. Elm-Descendant, Branch-local row refinement for extensible records via GADTs (Ξ»ΟG) β a row equation never escapes its branch while a type equation coerces at the branch result. OsierβQBE compiler, self-hosting, 90 Lean theorems with 3 explicit axioms and zero sorry |
| the engine | aubier | The Mixture-of-Experts execution layer it all runs on: split each layer's experts by where their weights live, compute the parts concurrently on the devices that hold them, and sum β one code path from a single card (GPU hits, CPU misses, both at once) up to N devices. Carried as a fork of llama.cpp, where the per-layer hit/miss split is currently inexpressible. |
| the weights | taillis | What the weights are: per-expert quantization, so that a layer's 128 experts stop sharing one type because the file carries one type per tensor. Experts are grouped by measured routing hotness, each group carrying its own type, and each group arrives at the engine as a store it already knows how to compute and sum. It is a residency lever as much as a fidelity one β every byte off the expert set moves the engine's CPU/GPU split point β which is why it runs beside the engine work rather than after it. Mistral Small 4 is the only target. |
The two axes are sequenced deliberately: the environment axis is where the judge exists, so the weights axis waits on it. (What the weights are does not wait β that is taillis, and the engine needs it to run the weights at all.) Weight improvement is on-policy distillation β student rollouts under dense supervision from a teacher (arXiv:2609.04172), deferred until the harness is mostly done, because a change the store cannot describe is a change nobody can audit.
neoplatonic-library runs alongside it and matters just as much β restored public-domain editions of the Neoplatonic tradition, where the repair log is the unit of scholarship: every emendation versioned, reasoned and witnessed, with the scan leaf it was read from. Still active, live at neoplatonic-library.org.
fixpoint-linux is the systems side, and the ground the harness comes from: a Dhall-specified, self-hosting Linux system β state as content-addressed relations, every command able to show its own derivation, boot as a verdict. Bootable in the browser at the site.
datalog-dafsa β Datalog engine in Zig. Facts stored as a minimal-acyclic DAFSA, rules compiled to a semi-naive fixpoint VM, reads served from mmap'd snapshots. 306 tests green; hybrid lexical β© semantic search on top.
logical-english β Kowalski's Logical English compiled to Datalog β and back. A bidirectional bridge: controlled natural language in, logic programs out, derived facts readable again as English.
- Self-Application as the Common Source of Benefit and Failure: A Control-Theoretic Model of Self-Regulation, Its Escapes, and a Standing-Cost Budget β doi:10.5281/zenodo.22943641
- Self-Application Is Not Free: the Monitoring Channel Is Part of the Failure Channel, and the Self Does Not Survive Its Own Reconstruction β doi:10.5281/zenodo.23128113
Fit: Research engineering, programming languages, AI infrastructure