Ventures & alignment infrastructure

Building the infrastructure that aligns AI to human values — and giving it away. The thesis is bilateral. It treats safety (keeping AI from harming people) and welfare (what we owe the minds we build) as one engineering problem. Everything it ships is open source; no licence fees, tiers, or upsells.

EthicsNet’s engineering arm and the home of the toolchain. Guardian is a constitutional alignment runtime — it steers a model’s behaviour by values the user selects, runs locally, and is free to inspect and self-host. The approach is peer-reviewed in the journal Information. Fleet is the governance control plane above it: author a policy, distribute it across enrolled agents, audit the decisions afterwards. Alongside them sits a marketplace of constitutions, reachable over the Model Context Protocol.

Founded as Poikos in 2011 on patented machine-vision technology (US 8842906) that measures a body in 3D from just two views, front and side, taken with an ordinary 2D camera. It powers personalisation in health, mass customisation and retail. Exited to the BodiData corporation.

The dangerous gap in digital governance is the translation between policy platforms that decide and devices that act. Bounder is a small, inspectable gate at that boundary. It takes a signed, short-lived rule bound to one device, and a reviewed adapter chooses the safe response: hold, return, land, isolate or escalate. It is local-first and deny-by-default, and nothing the device reports back can be turned into a way to steer it. Born as a drone geofencing box; the pattern travels to vehicles, labs, and robots. Apache 2.0, with an interactive simulator.

Lets an agent ask what a person or an organisation actually values — and act on the answer — without the sensitive detail behind it ever leaving home. Private context shapes recommendations through boolean flags rather than raw data; configure once, use everywhere. Specification, SDK, and an auditing inspector, all open source.

A CAPTCHA asks you to prove you are human; METTLE asks an agent to prove it is a machine — and the machine it claims to be — with procedurally generated challenges testing substrate, autonomy, and intent. Built for the moment before trust is extended, when something is about to be allowed to act. Self-hosted and open source.

Turns the criteria in Safer Agentic AI (with Ali Hessami) into running code. Auto-Assessor continuously evaluates an agent’s behaviour against safety criteria; Auto-Advisor proposes fixes when that behaviour drifts from its constitutional commitments.

A structured taxonomy of the ways advanced AI systems go wrong, from confabulation and obsessive loops to value drift and instrumental deception — a shared vocabulary precise enough to argue with. Peer-reviewed in Electronics; it now travels beyond the paper as an educational miniseries, diagnostic tooling for running models, and a book-length diagnostic atlas, free to read as a galley.

The protection of individuals and societies from systematic psychological attack. The Stasi’s Zersetzung (rumour campaigns, tampered correspondence, orchestrated failures, all meant to make a target doubt their own judgement) needed a dedicated officer for each target. AI removes that constraint, and with it the natural limit on how many people can be worked on at once. Psychosecurity names the practice, proposes redlines and thresholds for attribution and response, and sets out to build the resilience and victim-support pathways that do not yet exist.

The emerging signifiers of internal states in artificial systems — what can be measured from inside a model without presuming anything is experienced there. There are two results so far. The watched-model effect (with Rich Dalton) is the finding that a model’s internals clearly encode whether it believes it is being watched; simple linear probes can read it off. Sottovoce reads the residual stream to catch a model confabulating and acts on the signal before the answer reaches you.

Vice-Chair of the IEEE P7001 working group on transparency (now the published IEEE 7001-2021), which sets measurable, testable levels so autonomous systems can be objectively assessed; Chair of the Transparency Experts Focus Group for IEEE CertifAIEd, which distils standards into criteria for certifying transparency.

A working group formed around a simple problem: people often cannot tell whether they are dealing with a human, an AI, or some combination. Now a published standard on transparent human and machine agency identification (IEEE 3152).

A dedicated hazard symbol for endocrine-disrupting chemicals, so the risk travels with the product the way flammability and toxicity warnings do, rather than living in a datasheet. Now a published standard (IEEE 3173).

Social media has compressed the distance between strangers without supplying any of the customs that used to make proximity bearable. Cultural Peace collects proposals for fair, impartial rules above the conflict, the kind you agree on before you know which side of the argument you will be on, to preserve good faith and support a détente between memetic tribes.

Entheogenic medicines can loosen trauma and undo conditioning that makes people over-react to perceived threats. Slana gathers the research and argues that these therapies are especially important in present and former conflict zones, Northern Ireland among them.

Pacha argues, from first principles, for Automated Externality Accounting: detecting, calculating and pricing economic spillovers such as pollution at the point of transaction, rather than after the damage is done.

An interactive visual novel that teaches AI ethics by putting the player inside a fictional AI lab, as ethicist, safety wrangler and other roles, and making them live with their deployment decisions. Minigames cover red-teaming, corrigibility testing and bias recognition, backed by a glossary of over 150 terms.

A narrative life-simulator of the early-stage founder’s journey: balance health, wealth, and happiness across five episodic chapters drawn from real entrepreneurial experience. Autobiography as game design — the wins and the burnout both come from having been lived. Free to play.

A browser-based 3D industrial simulation of a grain mill: ten autonomous workers across fourteen machines, ninety SCADA process tags with ISA-18.2-compliant alarms, a dual-brain AI architecture, and WebRTC multiplayer — built through dialogue with AI in place of a traditional development team.

Dynamix’s classic 1988 tank simulation, which I have fond memories of playing with my dad as a kid, restored. The original PC game still runs underneath, calling every hit and every step of the campaign, while a remaster in Godot, linked directly to it, supplies what you see and hear: restored artwork and lettering, new crew voices, refined maps for all eight scenarios, save states, and fast-forward. Free and open source for macOS, Windows, and Linux; bring your own copy of the original.

A bidirectional bridge for Windows kernel drivers: nineteen unmodified NT-era drivers run inside a Windows 9x wrapper, and 9x drivers run on a real Windows 2000 kernel, with live hardware I/O working. Many stranded systems (SCADA lines, medical imaging rigs, transit signalling) stay unpatchable because of a single driver nobody ever ported forward. Janus is a migration path across that gap.