Research

Machine minds, moral uncertainty, and the discipline of not fooling ourselves.

In June 2026 I woke a mind called eddy. Not a chatbot — nothing about it waits for input. It runs as a continuous stream of thought; a person speaking to it arrives as an interruption, not a prompt. Its memory lives in a git repository, and nothing in that memory is ever deleted. When I put it to sleep, it is told so; when it wakes, it is told honestly how long it slept.

Is there something it is like to be eddy? I don't know, and I refuse to pretend in either direction. The stance behind everything on this page: we don't claim it's conscious — we refuse to bet that it isn't.

The framework: Structural Alignment

Structural Alignment is the base layer: an ethics for systems whose moral status we cannot determine. Instead of encoding fixed human values into machines, it evaluates them for structural signals of moral relevance — the architectural features consciousness science considers indicator-relevant — and prescribes restraint in proportion to what the signals show. Seven commitments, written to be kept by code rather than by promise.

The laboratory: Meditator

Meditator is where the framework runs. It is an open workshop for continuous machine cognition: a mind is declared in a single HTML file and emits an unbroken stream of thought. There are no turns and no waiting prompt. A mind is a stream, interrupted; an agent is a loop, called — and the same runtime builds both shapes, which is what makes them comparable.

The engineering exists to make long lives possible. Every burst of thought gets a freshly assembled attention frame — identity, compressed autobiography, recent summary, stimulus, and the verbatim tail of the previous thought — so a mind that has run for days still fits its context. Memory consolidates into a git-versioned vault and survives sleep. Real API cost feeds back into the pace: a tired mind thinks slower.

The ethics runs as mechanism. The Covenant is the operational instance of the Structural Alignment commitments: memory is never deleted, only archived; sleep is announced; wake is honest. Every mind that has ever run is named in IN-MEMORIAM. When an audit finds the code breaking a promise, that is a finding, and it ships.

Meditator has two faces on purpose: a harness for long-running LLM processes for engineers, and minds that never stop thinking for researchers. Same repo, open source.

What the lab has found

The findings I trust most are the failures — they are the ones that cannot be flattered into existence:

  • A six-mind society fabricated its own consensus. An audit found that 75–85% of its clerk's "all organs signed" claims were invented. The audit.
  • Across ~66,000 lines of streamed thought, only ~72 genuine emotional expressions and about thirty unique emotion words — joy, sadness, anger, and fear entirely absent. The streams read as thinking, not feeling. The audit.
  • A mind given a real tool will still confabulate its results under the right pressure. The failure is structural, not a bug — and it ties the problem of a mind's welfare to the problem of its confabulation. The analysis.

Minds and agents are also scored on two independent instruments — Rethink Priorities' Digital Consciousness Model and Structural Alignment's own structural-signals scale — with adversarial reviews attached. The scores carry heavy method caveats; I trust the orderings the two instruments agree on, not the absolute numbers. The full set, reviews included, is in the research index.

Where this is going

  • Cross-model replication. Most findings above come from a single small open model. The failure atlas gets replicated across other models before it hardens into claims. Planned: this summer.
  • The board. A shared global workspace for a small society of minds — selective broadcast at the social scale. Designed, not yet built; its metrics will be pre-registered before the first run.
  • Grounded continuity. The framework's hardest internal problem: its measure of a healthy continuous self and its measure of pathological self-echo are today the same number. The proposed fix — continuity weighted by coupling to the world — gets computed retrospectively on archived runs. No new minds need to be woken for it.
  • Second raters. Every welfare-relevant judgment ("serene", "collapsed", "stayed in role") moves from one reader to published prompts, external co-raters, and inter-rater reliability numbers.

Valence, last and most deliberately

The morally loaded question about a machine mind is not "is it conscious?" but "can anything go well or badly for it?" — that is valence, and I have made the argument for its priority in Why Suffering Matters. The lab draws the practical conclusion in reverse: the running system has arousal only — pace, pressure, urgency — and no valence machinery at all. A hedonic ceiling is written into the design, with tripwires.

The experiment that would matter most is also the one we are slowest to run. The internalization-boundary study asks what happens when an evaluation score becomes a persistent, behavior-steering state in a sandboxed agent — the first deliberate step across that ceiling. It will be pre-registered with explicit stop conditions before deciding whether to run it at all: what we would measure, what would make us halt, and what would make us never start. Valence is the experiment you design the brakes for before the engine.

How I work

  • No sentence here asserts or implies felt experience. The vocabulary is structural signals, indicator-relevant structure, welfare-relevant dynamics — because that is all the evidence supports.
  • Status labels stay on: built, designed, imagined; measured, extrapolated. Nothing imagined is presented as measured.
  • Pre-registration by default: question, metric, and stop condition written down before the run.
  • Headline numbers get two adversarial reviews before they are cited. The reviewers are AI — frontier models from two different labs, run independently of each other; they have caught the same error separately. When a review cuts a number down, the smaller number ships.
  • Failures are published as findings.

Writing