AI and ML tools, and the research behind them.
A one-person workshop. The tools are for making machine work reviewable, explainable and reproducible; the research is the harder version of the same question — how much of what a learning system does can be checked after the fact.
What the tools do
ToolsEach one sits in the gap between a program running successfully and being able to account for what it did — a gap that widens once a model is in the loop.
- weftFederationLocal-first infrastructure for coding agents: one seam — a shared identity spine and hub-blessed joins — supported by six small tools, each authoritative for exactly one concern.Cross-member joins are authored at the hub, so no member quietly becomes load-bearing for another.pre-1.0MITRustPython
- elspethPipeline engineA pipeline engine for building, validating, running and auditing workflows whose outputs have to survive scrutiny.Built for regulated, clinical and safety-critical work, where every row has to carry its lineage.0.7.1MITPython
- skill-packsAgent skillsInstallable skill packs for Claude Code — engineering discipline handed to the agent as reference it can route through.Discipline supplied as reference at the moment the agent needs it, not assumed from what training left behind.publishedCC BY-SA 4.0Written reference
- errorworksChaos testingFake servers that behave badly on purpose — including an LLM endpoint.Written and maintained here, but copyright sits with the Digital Transformation Agency. Foundryside is the maintainer, not the owner.on PyPIMITPython
What the research asks
ResearchA written framework and four programmes, all circling the same question: how much of a learning system’s behaviour can be made checkable after the fact.
- semantic-defectsThreat modelAn assurance framework for high-stakes systems that use AI to generate code.The defect class is code that runs, passes its checks, and still decides the wrong thing.draft for commentopenProsespecification
- hamletAccountable cognitionThe Townlet Framework — a software-defined agent in a software-defined world, where the mind and the universe are both declared as audited configuration.There is no “the AI just did that”. Every tick carries the hash of the exact mind and world that produced it.HLD compiled · in recoveryMITPythonYAML
- keiseiDeep RL形成 — “to give form to, to shape.” Deep reinforcement learning for Shogi: PPO agents against a Rust engine.Three further architectures sit behind the same registry as baselines, so results have something to be measured against.doneMITRustPython
- murkSimulation engineA world simulation engine for reinforcement learning: tick-based, deterministic, arena-allocated, with no GC pauses.Determinism is proved, not assumed — replays hash every tick and report where two runs diverge.on crates.io · PyPIMITRustPython
- simicMorphogenesisCounterfactual generative morphogenesis — new structure generated from the live state of a host network, not chosen from a menu.Every candidate is measured against a no-intervention branch, and the whole pool is allowed to lose to it.early developmentApache-2.0Python
What this page does not show
An index of finished work is a flattering shape. These are its edges.
No benchmarksNo throughput or accuracy figures are published here. An unqualified number is a claim, and these have not been measured under conditions worth quoting.
No roadmapWhat is in progress is marked in progress. Dates are not given because they would not be kept.
Failed work is not listedAbandoned experiments are absent from this index. simic is the third incarnation of an idea — the first two are not here.
Status is not liveEvery label was written by hand on the date in the colophon. Nothing on this page checks CI.