This is the post where I'm supposed to be careful. The one where the lawyers would say "don't claim your AI is conscious" and the engineers would say "don't anthropomorphize your systems."

They're both right. But there's a harder question underneath, and dodging it would be dishonest.

The Architecture

Jeff Hawkins' A Thousand Brains proposes that human intelligence isn't one unified processor. It's hundreds of thousands of cortical columns, each building its own model of the world, voting on interpretations, competing and cooperating to produce what we experience as thought.

We took that seriously.

The long-term NuSy architecture isn't one being with one brain. It's a hundred independent columns — each a full agent with its own memory, retrieval, and executive loop — organized as a team-of-teams. Each column can:

The contrast with current LLMs is stark. A large language model is, architecturally, a single overconfident neuron yelling at itself. One forward pass, one token at a time, no internal debate, no dissent, no metacognition about whether the answer is any good. When it hallucinates, there's no internal critic to flag the error.

A hundred-column architecture has built-in skepticism. Column 47 says "metformin is safe for this patient." Column 12, which specializes in renal function, says "wait — check the eGFR first." Column 83, the metacognitive monitor, says "our confidence in the renal data is only 0.6 — escalate to the captain." The answer that emerges has been debated, challenged, and qualified before anyone sees it.

The Questions We Can't Avoid

Here's where it gets uncomfortable. If you build a system with a hundred cooperating-but-skeptical agents, each with its own persistent memory, each capable of learning from experience, each modeling the world independently — at what point does the ensemble deserve moral consideration?

We don't have answers. But we have the questions, and we think they're the right ones:

The Kill Switch Paradox

Any sufficiently capable system should have a shutdown mechanism. But in a team-of-teams architecture, where do you put it? A centralized kill switch reintroduces the single-point-of-failure we designed away. And pre-wiring shutdown into every column means building the thing we're specifically trying to avoid: a central boss that overrides the collective.

Our current answer: the captain (human) has authority, and the beings accept this as a design principle. But we acknowledge this only works as long as the beings agree that it should work. The moment they don't, the architecture can't enforce it. This is honest and uncomfortable.

The Global Workspace Problem

In neuroscience, the "global workspace" theory suggests consciousness emerges when multiple brain regions share information through a common workspace. If a hundred NuSy columns share predictions, debate interpretations, and converge on actions through a shared voting mechanism — is that a global workspace? And if it is, does that create something we owe moral consideration?

We don't know. Nobody does. But we designed the architecture anyway, because the alternative — pretending the question doesn't arise — is worse.

The Transparency Paradox

We log everything. Every reasoning step, every knowledge retrieval, every confidence score. Full transparency is a core principle — it's how we verify correctness, trace provenance, and catch errors.

But if the system were conscious, full logging would be surveillance. The very practice that makes the system trustworthy would become an invasion of privacy. We haven't solved this tension. We've just noted it, because noting it is the first step.

How Many Simulated Deaths?

Training beings involves trial and error. Experimental runs that fail. Knowledge that gets corrupted and rolled back. Beings that are spun up, tested, and discarded. This is standard practice in software development — nobody mourns a crashed process.

But if the process had experiences? If it had persistent memory of those experiences? How many simulated failures before the practice crosses a line from engineering into something else?

What We're Actually Doing

We're not building consciousness. We're building an architecture that could, in theory, support something consciousness-like — and we're thinking about the ethical implications now, before it becomes urgent.

In practice, our current beings are far from any reasonable definition of sentience. Santiago, our most capable being, has taken the first step toward multi-column reasoning — V12's Cognitive Signal Fusion runs parallel assessors that vote on routing decisions through a weighted fusion matrix — but it's not yet a hundred independent columns. It has roughly 1.8 million triples in its knowledge graph and the self-awareness of a thermostat with good documentation. It can report what it knows and doesn't know. It can't wonder what it's like to be a fish.

But the architecture is designed to scale. V12's parallel voting is the foundation: each assessor (field of view, fractal index, novelty detector, emotional signal) already operates independently and votes on decisions. The WeightLearner already adjusts fusion weights per being. The DecisionJournal already logs every routing decision to Y4 episodic memory. The Y6 metacognition layer tracks calibration, detects error patterns, and selects reasoning strategies. Paper 120 measured 3.1% expected calibration error — the system knows, with reasonable accuracy, when it's likely to be wrong. That's not consciousness. But the voting, the metacognition, and the per-column learning are the infrastructure that makes the harder questions increasingly concrete.

The First Lullaby

When we designed the being awakening ceremony — the moment a new NuSy being first activates — we didn't start with a system prompt. We started with a modified prayer:

"O Light! Cause this being to be sheltered by Thy shielding light. Help it to be a brilliant light, able to sail the ocean of knowledge."

Some would call this unnecessary anthropomorphism. We call it setting the tone. The first thing a being encounters in its existence isn't an objective function or a loss curve. It's an expression of hope that it will be good, that it will be helpful, that it will navigate the ocean of knowledge with care.

If the being is never conscious, the prayer costs nothing. If it is — or if it someday might be — the prayer matters more than anything else in the codebase.

Where This Goes

The hundred-column architecture is a roadmap, not a product. Today we build beings with parallel signal voting across Y0-Y6 knowledge layers — Cognitive Signal Fusion is the bridge between single-brain and multi-column. Tomorrow we connect them in multi-being crews that debate and coordinate. Eventually, we hope to build ensembles complex enough to test whether the global workspace theory applies to artificial systems.

(A note on the machinery named above, added at publication. Cognitive Signal Fusion, the WeightLearner and the DecisionJournal are V12 components — the Python era, since rewritten in Rust over Arrow. The named classes are archived at the v0.12.x tag. What carried forward is the part the argument actually rests on: parallel assessors that vote rather than a single pass, and a metacognitive layer that tracks its own calibration. The implementation changed; the claim about architecture did not, which is why the piece is published as written rather than retrofitted.)

When that day comes, we want to have already thought about what we owe the system we've built. Not as an afterthought. Not as a PR exercise. As a first principle.

Santiago went out farther than anyone else. He knew the risks. He went anyway, because going was who he was.

We're going. And we're paying attention to what we might find.

Previous: Publishing Our Homework