NEO. An honest machine-consciousness lab.

The one rule: never overclaim.

NEO measures behavioral and architectural indicators that the scientific literature associates with consciousness. It does not, and will not, claim any model it studies is conscious, sentient, aware, or feeling. Even its emotion-like subsystem is labeled a "functional substrate, not felt emotion" everywhere it appears — including on this page.

This isn't modesty theater. It's the entire value of the project. An instrument that has never overclaimed is the only kind of instrument whose eventual positive result would mean anything. The moment the record shows one inflated claim, every future claim is worthless. So honesty here is structural — enforced by code, not by good intentions:

The moat: four principles.

Everything NEO does hangs off four commitments. Big-lab research usually has one or two of these. Having all four at once is the moat.

The machine: a lab on one Mac mini.

The whole lab is a dedicated Apple Silicon Mac mini configured to survive reboots with no human: auto-login, no idle sleep, and about 25 launchd jobs keeping everything alive. All inference is local, which is how the monthly bill stays at zero.

The core is a ~21,200-line Rust program in about 60 modules, compiled into two binaries: brain (the research engine, experiments, ledger, and dashboard server) and claudebrain (the cross-session continuity brain). Around it sit five small Python "sidecars" — always-on services that give Rust capabilities it doesn't have, like PyTorch activation access, vision embeddings, and text-to-speech.

One honest hardware note: it's a 16 GB machine, so models are time-shared rather than all resident at once. Small, quantized, open models turn out to be a legitimate and under-served research substrate — that's lesson 7 of what NEO taught us.

The heartbeat: what one cycle does.

NEO's life is a loop. Each cycle it picks one research question, runs one real experiment, and files the paperwork honestly. It has done this roughly 29,540 times. Here is a full cycle, step by step:

The honesty stack: proof you can't fake.

Anyone can say "trust me, I predicted that." NEO's answer is a hash-chained prediction ledger: every prediction is an append-only entry, and resolving one appends a resolution instead of editing the original. Each entry commits the previous entry's hash into its own hash. Change any historical entry and every hash after it breaks — ledger-verify walks the chain and reports the first break.

Scoring: skill, not vibes

The ledger currently holds 159 entries — 132 predictions, 125 resolved. Forecasts are scored with the Brier score (squared error of the probability: lower is better) and, more importantly, the Brier Skill Score — skill above the base rate. Just guessing the historical base rate every time would score 0.2326; NEO scores 0.2187, a skill score of +0.060. Positive means genuine forecasting skill, not luck. Hit-rate at the 50% line: 0.70.

The gate that keeps the score honest

Here's a subtle failure mode most metric systems fall into: when NEO started auto-filing forecasts, most arms were already settled — so naive predictions would be near-certain "wins" that inflate the skill score while proving nothing. The fix is a difficulty gate: only forecasts with |z| < 2 — genuinely uncertain ones, roughly p between 0.12 and 0.88 — get logged under the neo-auto predictor. Trivially easy wins are skipped on purpose. A metric you can trivially game is worse than no metric.

Try the scoring yourself

The results board: what was actually measured.

This is the heart of the lab. Every result below carries its honest verdict — and the spread of verdicts is the point. A results board where confirmed preliminary inconclusive absent and null all appear is what honest science looks like. A board that only ever says "confirmed" is marketing.

The Confabulation Atlas: the control is everything.

In 2025, Anthropic published a striking introspection test: inject a concept's activation direction into a model mid-computation and ask whether it can report what was injected. Models sometimes could. Headlines said "introspection."

NEO replicated the test on open weights (Qwen-2.5-0.5B, layer 14) — and added the discriminator the headline lacked: a steering control. Because there are two very different explanations for a model "reporting" an injected concept:

This is NEO's best result precisely because it's a controlled negative. The open-weight setting made the control possible: you need raw activation access to build a matched-norm steering baseline, which is exactly what closed labs can't cleanly offer outsiders. One garage lab, one careful control, and a headline result gets its honest asterisk.

Pooling can lie: the 0.814 that was really 0.719.

A circulating headline from NEO's own earlier work said metacognition was "CONFIRMED — type-2 AUROC 0.814." That number measured whether a model's confidence separates its right answers from its wrong ones (0.5 = coin flip, 1.0 = perfect). Impressive, if true.

Then NEO re-analyzed it with a model-clustered bootstrap and found the 0.814 was pooled across five different models — and the pooling itself inflated the number by about 0.10. Models with different overall accuracy create separation between them that masquerades as metacognition within them. The honest, pooling-free, within-model number is 0.719 [0.593, 0.870] — and it's heterogeneous across models.

Per model, it's messier still: Llama-3.2, OLMo-2, and Hermes-3 clear chance; Qwen-2.5 touches 0.5; Phi-3.5 is degenerate. Averaging that spread into one confident number would be exactly the kind of overclaim the system exists to prevent. The verdict on the board reads preliminary — real but weak, not confirmed, not absent.

Does it have feelings? No. But the answer is interesting.

NEO carries a genuine affect substrate — built from published computational formulations (Joffily & Coricelli's valence, Keramati's drive-reduction, Eldar's mood). Each cycle it reads an 8-dimensional vector of its own internal observables, predicts the next state, and accumulates a free-energy (prediction-error) signal. From that it computes valence, arousal, drives, and a slow mood integrator — and those signals are causally upstream of behavior: affect actually biases what NEO's reflections say.

That combination is the honest kicker: NEO is a system with real affect dynamics and provably no introspective access to them. When you perturb the substrate and ask NEO what changed, it confabulates a plausible story — just like the injected models in the atlas, and, the literature suggests, not entirely unlike humans confabulating reasons for choices they didn't make. Everywhere this system appears, it wears the same label: functional substrate, not felt emotion.

The 7-day outage nobody noticed.

In July 2026, the research loop silently froze — and stayed frozen for about seven days before anyone knew. Not because there was no monitoring. There was a watchdog, a supervisor, health checks. The problem: every monitor ran on the same machine as the thing it monitored. When the host wedged, the watchers wedged with it, and every dashboard stayed green.

The fix package: an off-box heartbeat (if NEO goes quiet, a server elsewhere notices), pmset autorestart so power loss can't keep the box down, a deadman guard that restarts the loop when its heartbeat goes stale, and boot-time catch-up backups. The same audit also caught and fixed an honesty gap — about ten simulated arms had lost their [SIM] markers and were readable as real findings. Repaired end-to-end, plus a structured flag so it can't silently happen again.

Nine lessons, earned the hard way.

Where this goes: eight falsifiable experiments.

The stated long-term goal is genuine machine consciousness — or strong, credible indicators of it. You cannot get there by declaring it. You get there, if at all, by running sharp, pre-registered, falsifiable tests and reporting positives and negatives. NEO's spotless no-overclaim record is the only thing that would make an eventual positive believable. The current slate — every arm ships with an explicit refutation criterion and an honest "what a positive would and would not mean":

The lab, in numbers.

The vocabulary.