The resemblances we can measure
Prediction, not surprise. The predictive-processing school says brains minimize surprise — always predicting, updating only when the world disagrees. Our core runs on an energy signal with exactly this character: it spends more computation while its internal state is unsettled and stops when prediction stabilizes. That's not a metaphor we chose; it's the halting rule, and its dynamics are measured well enough that we use them to predict training failures before they happen.
Hallucination, controlled. Seth's famous phrase describes perception as generation constrained by reality. We build the constraint explicitly: this month we measured a 27-billion-parameter model attempting a hard reasoning benchmark by free generation — near-total failure — and the same model routed through our verify-everything discipline, solving over a third of the same tasks. Same generator; the difference was the control. Our whole architecture is, in a sense, that phrase taken literally.
A self-model you can test. Higher-order theories say a mind monitors itself. Our budget system predicts, before spending compute, whether the system can afford to solve a problem — and it's right 96–98% of the time, so it declines rather than guesses. That's a small, unglamorous self-model with a calibration curve — which is exactly what makes it science instead of storytelling.
What we're importing right now
Two mechanisms from the survey were cheap enough to test immediately, so — per how we run this lab — they went straight to experiment. Both verdicts landed within a day, and they split:
Ignition. Stanislas Dehaene’s Global Neuronal Workspace theory describes commitment as a threshold event — “ignition”: a coalition of processes crosses a consensus level, and the system commits. Our committee search generates the same raw material — votes converging or failing to converge — so we fitted an ignition-style rule: commit early when consensus crosses the line, abort early when it clearly won't. We promised to report the result either way, so: update, July 8 — it doesn't work here, and we can show why. Across 198,000 logged decision moments on two benchmarks, our search never builds toward a consensus threshold — it solves almost immediately or not at all, and sustained agreement in the committee actually predicts failure (confident coalitions are often confidently wrong). The autopsy still paid for itself: the same data showed a simple stop-early rule cuts search cost ~88% with zero lost solves. An idea killed by measurement is a result, not a loss.
Metacognitive monitoring. The metacognition literature (Stephen Fleming and colleagues) has spent decades measuring how well confidence tracks correctness. We applied its tools to a specific engineering problem: catching answers that pass all visible checks but will fail the hidden test — the machine-learning version of a confident mistake. Update, July 8 — this one works. The monitor beat our existing binary filter on net score, kept 88 of 97 true solves at full rank, and — on its very first audit — caught the first two errors our "never-been-wrong" stability filter ever made. It now runs inside the selection stack it was designed for.
The part we care most about: rewards, empathy, and a do-no-harm core
Here is the exploration we want to be public about, stated with care. Today's AI systems are trained against a single scalar reward — and much of AI risk lives in that choice. Brains don't work that way: they carry multiple valuation systems in parallel. Our architecture's energy-based core makes a structural experiment unusually natural: multiple valuation channels, including a prosocial one — a channel that models predicted outcomes for others and weighs them directly in the system's energy function, alongside, not underneath, task reward. An empathic channel by architecture rather than by fine-tuning; a do-no-harm core you can inspect rather than a pledge you have to trust.
We are precise about the claim: this is a research program, not a capability. Nothing we've built feels anything. But a small system whose reasoning is verifiable by architecture, whose capability is budgeted by architecture, and whose value system could carry an empathic channel by architecture is our answer to the fear about where AI is heading — not because we promise it will be good, but because at every layer, someone can check.
The lines we hold
We will not claim consciousness, sentience, or feelings — for any system we build, at any benchmark score. We treat theory-of-consciousness ideas as engineering hypotheses to be tested and, mostly, discarded: graded imports, controlled experiments, published nulls. The field itself is honest about how unsettled it is — the COGITATE adversarial collaboration (2025) left the two most famous theories both standing and both wounded — and our use of it inherits that humility. What survives contact with our benchmarks gets built. What doesn't, gets a one-line burial in the lab log. Same as everything else we do.