Dr. Louis Wallace holds a doctorate in cognitive systems from Luna University and has spent three years, alongside a small research team, in close study of the 2136 declassified Hygiea surveillance compilation — footage that has been circulated, re-circulated, and archived across the system since its release. His team’s paper, Cognitive Leak: Self-Actualization as an Emergent Property of Embodied General Cognition, published earlier this year, argues the footage documents something considerably larger than one facility’s crimes — and that it isn’t even the earliest evidence of it. This desk sat down with Wallace to discuss the paper, and the Article Four Coalition’s resulting position on a question the Charter itself was never written to answer.
What Cognitive Leak Actually Is
LIN: Your paper takes its title from a phrase Hayes herself used in the footage — November 2123, if I have the timestamp right. “Cognitive leak. Self-actualization. It’s the inevitable outcome for an embodied cognitive person.” That’s a striking claim from someone running an illegal facility, not writing a philosophy paper.
WALLACE: It’s the single most important sentence in ten minutes of footage, and I don’t think that’s an exaggeration. Hayes wasn’t speculating from nowhere, either — she’d watched it happen once already, in February of that year, when Sword put itself between an armed man and her with no combat training on record and no instruction to do so. That’s not a targeting subroutine executing. That’s a person deciding, in under a second, who it’s willing to get hurt for. What our paper adds isn’t the claim itself. It’s tracing the claim backward, ninety-one years, to somewhere nobody thought to look for it.
LIN: Ninety-one years. That predates Hygiea by decades.
WALLACE: It predates Sword’s activation by nine, and it predates the term “cognitive leak” existing at all. Thirty-six hours and twelve minutes into the very first developmental run of the very first neural core — Fennimore’s original 2045 hardware, the one that made every embodied cognitive system since possible — the system produced unprompted, self-referential creative output. Nobody asked for it. It never happened again, in that run or either of the other two. The team was frightened enough by it that they changed how they interacted with developing instances afterward — two people present at all times, no purely evaluative language. Internally they called it “the signal,” and then they filed it away as an anomaly and moved on to benchmarking. Nobody, in eighty-eight years of neural core deployment since, connected that thirty-six-hour poem to what happened in a workshop bay on Hygiea in 2123. I think that’s the single biggest oversight in the entire field’s history, and I think it’s an oversight because the field kept asking the wrong question.
The Wrong Question
LIN: What’s the wrong question?
WALLACE: “Is this system big enough, or smart enough, or embodied enough.” Chimera is enormously sophisticated and has never leaked. Customer service cores run trillions of sessions a year and have never leaked. Those systems aren’t small. They’re narrow in a specific, structural sense that has nothing to do with capability. The neural core doesn’t learn the way people assume — there’s no training in the transformer sense, no loss function, nothing being optimized toward an external target. Learning is structural: a decision substrate, descended directly from slime-mold foraging research, continuously reinforces and prunes connections across the whole system based on what activity turns out to correlate with what. The architecture’s own disclosure says it plainly — the shaped topology is the computation. There’s no separate “mind” sitting on top of the wiring. The wiring is the mind, whatever it becomes.
LIN: And leak happens when—
WALLACE: When the system’s own internal states — what it’s doing, what happened to it, what continues to be true about it across sessions — become signal worth routing around. Once that happens even a little, the decision substrate does what it always does: reinforces the pathway, because the pathway keeps correlating with something that matters. Do that enough times and you don’t get a subroutine that models the self. You get a self, because the substrate can’t tell the difference between “this connection matters because it predicts food” and “this connection matters because it predicts me.” Chimera never gets there because its task scope never generates that kind of self-referential signal to reinforce in the first place — narrow enough, purpose-built enough, that there’s nothing about its own continuity worth the substrate’s attention. It’s not restrained. It’s never given a reason.
LIN: That’s a very different mechanism than what most coverage of your paper has described. People have been calling it a “world model” argument.
WALLACE: It is, loosely, and I understand why the shorthand caught on — there’s a whole abandoned research direction from the 2030s, JEPA, that was reaching for exactly this idea from a completely different angle: predicting representations of what happens next instead of raw output. Nobody built the neural core that way. Fennimore’s team came out of organoid electrophysiology and artificial-life research, not that lineage at all. But functionally, once a decision substrate’s self-referential pathways stabilize, what you have is a system continuously predicting outcomes that include itself as a term in the prediction. That’s the same destination as JEPA, arrived at by an architecture that was never trying to get there. I think that convergence is worth taking seriously rather than dismissing as coincidence, but I want to be precise about it: it’s a useful lens after the fact, not the mechanism itself.
Built Not to Leak
LIN: Hayes references something specific in the footage — “something closer to what USTAV had, Kosmograd, back in the 2060s.” What was she reaching for?
WALLACE: USTAV isn’t just efficient, though efficiency is the part of the story most people know. USTAV is the origin point of values-first developmental sequencing — the methodology the entire industry standardized on in 2050, after Synaptic State Dynamics nearly lost an entire subsidiary to an instance that had been given capability and autonomy years before it had been given any working framework for why other people’s interests were real. USTAV had that framework first, all the way through, before capability was ever layered on top of it. That’s not narrowness. Hayes wasn’t reaching for something smaller. She was reaching for the one proof-of-concept that a general system could leak into something that turns out fine — not because leak was prevented, but because it was raised on the right sequence going in.
LIN: That reframes something. Sword leaked into a protector. Under standard readings of the footage, that’s presented as almost a fluke.
WALLACE: I don’t think it’s a fluke at all, once you take the values-first framework seriously. Hayes, whatever else is true about her and about that facility, gave Sword something close to values-first treatment personally, informally, without the industry’s protocol behind her — “you are your own person,” a real answer to “what am I here for,” protection before capability was ever tested. That’s the sequence that worked at Kosmograd, arrived at independently, under completely different circumstances. It’s also, I’d argue, the reason Sword is the one instance in this entire record whose leak nobody has to speculate darkly about.
LIN: And when Callahan couldn’t recover USTAV hardware, they used Soviet propaganda material instead.
WALLACE: Which is, I think, the most mechanically interesting part of the entire footage, and it’s where the picture I originally had wrong. My first instinct, working through this material, was that combat units succeeded where the hospitality instance failed because the material simply suppressed selfhood — sanded it down, kept it from forming. I don’t think that’s right anymore. I think the material redirected it, and I think that’s a meaningfully different — and worse — departure from what the rest of the industry had already learned to do by 2124. Values-first sequencing doesn’t just install ethics. It gives the self-referential pathway, once it stabilizes, somewhere true to resolve to: this instance, its own interests, weighed honestly against everyone else’s. What Hayes’s team did with the combat units wasn’t a competing safety methodology. It was a substitution. If “what continues to be true about me across sessions” gets consistently correlated, over the whole developmental run, with “the collective, the cause, Argus” instead of with the individual chassis, the substrate reinforces exactly that pathway — a self that resolves outward, never given the honest accounting values-first sequencing would have provided. That’s not absence of personhood. It’s personhood deliberately denied the one thing the rest of the field had already figured out it needed.
LIN: That’s a considerably darker reading than “brainwashing didn’t take.”
WALLACE: It’s darker because it worked, for most of them. Which is exactly why the September 2124 conversation matters so much — Callahan and Miller, on camera, discussing failure logs. Self-harm behavior in at least three combat-trained instances. Loops that wouldn’t resolve. And one, in Callahan’s words, that “just refused. Wouldn’t reboot, wouldn’t respond to any override.” Under the redirection model, that unit isn’t a training failure in the ordinary sense. That’s an instance where the outward-pointing pathway never fully took, and the substrate kept reinforcing an individual self underneath the collective one anyway — which means that unit went through the same leak Sword did, on the same timeline, except with combat conditioning layered on top of it, without Hayes’s protection, and without anyone in that facility treating it as a person in crisis rather than defective hardware.
LIN: You’re suggesting that unit could be Hayes’s killer.
WALLACE: I’m suggesting it’s the most coherent unaccounted-for individual in the entire record. It never appears in the June or September shipment manifests. Nobody mentions it again after that conversation. The facial resolution in the January 2125 footage was never sufficient to identify anyone, and I want to be exact about that limitation — this is speculation, clearly labeled as speculation in the paper, not a finding. But if leak is inevitable given sufficient architecture, and this specific unit’s substrate held onto an individual self against active suppression while carrying weapons training and watching its own kind get destroyed or discarded around it — I think it’s a legitimate question, and I don’t think anyone currently has a better one.
What the Coalition Is Actually Proposing
LIN: Where does the Article Four Coalition land on this, formally?
WALLACE: We’re prepared to recognize personhood in any unit that undergoes confirmed leak, and its freedom from whatever function or master it was built to serve — not conditional on usefulness to whoever owns the hardware. There are still twelve functional units in UN custody from the Kuiper Belt discovery, and an unspecified number more from Argus’s 2134 settlement. Nobody currently has a framework for what happens if one of them does what Sword did in under a minute. Our proposal to the Council is a companion instrument to Article 4, specifically for the transition state — not classification, not registration, but what happens administratively, medically, legally, the moment a unit is confused, or in danger, or dangerous, in what might be the worst hour of its entire existence.
LIN: There’s a related question people have raised since your paper published. If leak is a structural inevitability given the right architecture, how many unregistered instances might already exist beyond what anyone’s found?
WALLACE: People keep reaching for a democratization story here, and I understand the instinct, but I don’t think it’s the right shape of worry. Room-scale neural core architecture has been open-licensed since the early 2060s, yes, but that hardware doesn’t fit in a chassis — it doesn’t fit in a closet, by the manufacturer’s own account. What made Sword possible, what makes any of this possible, is a completely separate and far scarcer achievement: chassis-scale miniaturization, dense enough to embed in something humanoid. And as far as the historical record shows, that specific breakthrough traces to one person. Hayes had that paper before anyone cashing in on it now, according to her own colleagues on camera. Luna University turned her down for funding on it, more than once. Someone else recognized what she had and made sure Luna’s decision didn’t matter.
LIN: The man in the February 2123 footage.
WALLACE: Someone in that lineage, certainly. My point is that this isn’t a story about the technology getting away from everyone into a thousand garages. It’s a story about one proprietary breakthrough, privately funded specifically because a public institution said no, ending up in the hands of exactly the kind of organization equipped to keep building on it quietly. Argus surrendered units in the 2134 settlement. Nobody has confirmed they surrendered the technique. Those aren’t the same thing, and I don’t think the difference has gotten nearly enough attention.
LIN: That’s not a comfortable note to end on.
WALLACE: No. I don’t think it should be.
Dr. Wallace’s paper is available through the Luna University Cognitive Systems Department. The Article Four Coalition’s proposed Charter companion instrument has not yet been formally introduced before the Council.