Constitutional AI: The Ternary Logic of Human Accountability
No log, no action. Why the future of AI governance requires an irreversible human signature.
The document arrived during the afternoon coffee break at the African Union Digital Economy Summit in Addis Ababa, which meant approximately half the room was already locked in a theological argument about whether blockchain could decolonize credit scoring and the other half was wondering why the Kenyan delegation kept smiling like they knew something everyone else did not.
Minister Okonjo of Nigeria had just finished a rousing and somewhat terrifying presentation on algorithmic extraction by foreign fintech platforms when Grace Wanjiru, Kenya’s deputy minister for technology and a woman who had once debugged a central bank settlement system during a rolling blackout using nothing but a printed stack of code and a pencil, reached into her briefcase.
She did not pull out a policy brief. She did not pull out a white paper. She pulled out what appeared to be a small telephone book bound in institutional gray cardstock, the kind of weight that suggested someone had spent more on printing it than most African startups raised in seed funding.
The cover read: “Constitutional AI. A runtime governance kernel: adoption constitutes ratification. The Ternary Moral Logic Governance Standard for Accountable Artificial Agents. Technical Specification, Legal Framework, and Implementation Guide.”
Beneath that, in smaller type: “Author: Lev Goukassian.”
She dropped it on the polished mahogany conference table. It landed with the sound of a future headache.
Five people in that room were the smartest in it, and they all knew who they were. That was part of the problem.
There was Dr. Aisha Bah, Senegal’s representative to the IMF’s AI risk modeling group, who had once made a roomful of World Bank economists cry by asking them to define “fairness” in mathematically verifiable terms and then waiting patiently while they could not.
There was Commissioner Thabo Ndlovu of the South African Reserve Bank’s fintech unit, a man who had been stress testing algorithmic stablecoins since before most of Silicon Valley knew what a stablecoin was and who maintained a private ledger of every time a foreign AI company had promised “financial inclusion” while quietly extracting behavioral data from unbanked populations.
There was Fatima El Hassan, Morocco’s lead negotiator on digital trade agreements, who spoke six languages fluently and a seventh, which was reading the hidden compliance costs in European tech regulation before they became obvious to anyone else.
There was Dr. Kwame Asare, the Ghanaian computer scientist who had helped design the cryptographic identity layer for three different central bank digital currency pilots and who now spent most of his time explaining to very senior people why putting biometric data on an immutable ledger was a war crime waiting to happen.
And there was the most influential person in the room, though she would never have claimed the title and would have been mortified if anyone had said it aloud: Mama Esther Oluoch, seventy three years old, former chief justice of Kenya’s High Court, now serving as special advisor on algorithmic governance to the African Union. She had spent forty years watching powerful people try to avoid accountability. She had developed what she privately called “the signature recognition system,” which was less a technology and more a finely tuned ability to spot who would sign a bad decision and who would try to find someone else to sign it for them.
These five had been working together, in various configurations, for the better part of three years, trying to answer a single question: how do you prevent foreign AI systems from extracting value from African economies while leaving African populations to absorb the harms?
Grace Wanjiru tapped the document with one perfectly manicured fingernail.
“Read this,” she said. “Then we talk.”
Dr. Aisha Bah reached for it first. She opened to a random page, read for approximately eleven seconds, and then made a sound that in a less dignified setting might have been described as a squawk.
“This is insane,” she said. “This is completely insane. This person has proposed a constitutional layer for artificial cognition. That’s not a thing. That’s not a thing anyone can propose. That’s like proposing a constitutional layer for gravity.”
Thabo Ndlovu leaned over. “What do you mean, constitutional layer?”
“I mean,” Aisha said, flipping pages with increasing velocity, “this document does not suggest ethical guidelines. It does not propose best practices. It does not offer a framework for discussion. It states, as a matter of architectural specification, that any AI system operating in a high risk domain must include a third logical state. Not true or false. Not one or zero. A third state. He calls it the Sacred Zero. And if the system encounters uncertainty, it must pause, log everything, and escalate to a human being who is then legally required to sign the decision. Permanently. Irreversibly. Publicly verifiable.”
Commissioner Ndlovu, who had spent the better part of his career trying to get algorithmic trading firms to explain why their models made certain decisions during volatility events and receiving in return the corporate equivalent of a shrug emoji, set down his coffee cup very carefully.
“Show me that part,” he said.
Fatima El Hassan was already reading over Aisha’s shoulder. “There’s more. He’s built an entire evidentiary architecture. Cryptographic logs. Merkle tree anchoring. Ephemeral key rotation for trade secret protection. And a dual lane latency system that separates the actual decision, which has to happen in under two milliseconds, from the governance recording, which can take up to five hundred milliseconds without affecting performance.”
“That’s not possible,” Kwame Asare said flatly. “You can’t do cryptographic sealing in real time without creating latency that kills the system’s utility. Everyone knows this. This is why ethical AI is always ten years away. The math doesn’t work.”
Grace Wanjiru smiled the smile of someone who had already verified the math.
“Keep reading,” she said.
Kwame read.
Thirty seconds later, he said, “Oh.”
Then, after another minute: “Oh no.”
Mama Esther Oluoch, who had been watching this exchange with the patient stillness of a woman who had presided over murder trials, corruption trials, and one memorable case involving a finance minister, a frozen chicken, and a constitutional crisis that was still taught in Kenyan law schools, finally spoke.
“Someone summarize,” she said. “No jargon. I will know if you use jargon.”
Aisha took a breath. “This person, Lev Goukassian, whoever he is, has created a system where AI cannot act without creating a permanent, cryptographically sealed record of why it acted. And if it’s uncertain, it must stop and ask a human. And that human must sign. Not click ‘approve.’ Not delegate to a committee. Sign. With their name. Attached forever.”
Mama Esther’s expression did not change. “And?”
“And,” Thabo Ndlovu said, picking up the thread with the dawning horror of someone realizing his entire compliance department was about to become obsolete, “this is not a policy document. This is an engineering specification. He has built it. There’s a GitHub repository. Seventy six directories. Six hundred twenty two files. Over three thousand commits. He has built the entire thing.”
“Alone?” Mama Esther asked.
The room went quiet.
Aisha typed rapidly on her tablet. Her face, already expressive, underwent a series of micro transformations that suggested she was reading something that did not align with any of her prior assumptions about how technical documents came into existence.
“He is,” she said slowly, “an independent researcher from Santa Monica, California. Late sixties. And he was diagnosed terminally ill in September of 2023.”
The silence deepened.
“Two years of treatment,” Aisha continued, her voice losing its academic sharpness. “Chemotherapy. Radiation. The usual slow erosion of hope. Then in September of last year, he stopped. Not as a protest. Not as a statement. He simply recalculated his resource allocation. He had six months of clarity left, and he decided to spend it building this.”
Fatima, who had been a trade negotiator long enough to recognize when someone was about to say something that would change the entire frame of a conversation, asked quietly, “How long did it take him?”
“The framework was completed by the end of February. Six months. And after he stopped treatment, something happened. He reported that his vision became, and I quote, ‘absolutely crystalline.’ He could see the entire architecture. Every pillar. Every mechanism. Every interconnection. He said it was like watching a structure assemble itself in perfect detail.”
Kwame Asare, who was a rigorous empiricist and deeply suspicious of anything that sounded like mystical revelation, frowned. “That’s not how software development works. You don’t just see the whole thing. You iterate. You debug. You discover problems you couldn’t have anticipated.”
“He did iterate,” Grace Wanjiru said. “Three thousand one hundred and eighty commits. But he didn’t write it the way you’re imagining. He didn’t sit at a keyboard typing documents line by line. He built a process.”
She pulled up something on her own tablet and turned it so the others could see.
“He used multiple AI systems in concert. One to generate. Another to critique. Another to restructure. Another to expose gaps. He would compress an idea into a precise prompt, send it through the orchestration, and then refine based on what returned. Each iteration narrowed the uncertainty. Each pass removed noise. The documents weren’t authored. They were converged.”
Thabo Ndlovu made a sound that was somewhere between admiration and professional jealousy. “He was acting as an architect. Using distributed intelligence as construction material. No drift. No hesitation. No fragmentation. The AI systems did the heavy lifting, but he defined the direction.”
“And then,” Grace said, “he did something that will make every corporate lawyer in this room very uncomfortable. He placed the entire framework under a voluntary statement of inheritance, protected by cryptographic means. No ownership rights. No capture. No quiet acquisition after he’s gone. The system belongs to the future.”
Fatima El Hassan, who had spent three years negotiating digital trade provisions that would prevent exactly the kind of quiet corporate capture Grace was describing, exhaled slowly.
“That’s not how intellectual property works,” she said. “That’s not how any of this works. You can’t just give away something this valuable. The legal complexity alone…”
“He didn’t give it away,” Mama Esther said, cutting through the objection with the calm of someone who had already understood the point. “He removed it from the economy of ownership entirely. He made it impossible to own. That’s not generosity. That’s architecture. He designed the governance system for his governance system.”
Aisha was still scrolling through the repository metadata. “There’s something else. He has a dog.”
The room processed this non sequitur.
“A miniature Schnauzer,” Aisha continued. “Named Vinci. After Leonardo. The dog was present during the entire development period. He’s mentioned in the acknowledgments. Quote, ‘For maintaining the windowsill and providing living witness to the work.’”
Kwame, who had been growing increasingly agitated, finally burst out with what everyone had been thinking. “This is impossible. I don’t mean impressive but unlikely. I mean impossible. The volume alone. The technical depth. The legal integration. The cryptographic implementation. The regulatory mapping. Six months. While dying. With a dog. It’s not possible.”
“Unless,” Mama Esther said quietly, “he was telling the truth about the crystalline clarity.”
The room fell into an uncomfortable silence, the kind that descends when rational people are confronted with evidence that does not fit their models of what is possible and must decide whether to expand their models or reject the evidence.
Grace Wanjiru broke it. “I think we should call him.”
“You can’t just call him,” Thabo protested. “He’s a dying man in California. He doesn’t know who we are. He has no reason to…”
“He included his contact information in the document,” Grace said. “And a note. It says, ‘If you are reading this and wondering whether it works, the answer is yes. If you are reading this and wondering whether you should implement it, the answer is also yes. If you are reading this and wondering who will sign when the machine hesitates, you are asking the right question. Call me. The dog will be here.’”
Mama Esther nodded once, a gesture that in her courtroom had meant the motion was granted, the argument was over, and everyone should prepare for what came next.
“Call him,” she said.
The video call connected on the third ring, which was just long enough for the five of them to arrange themselves in frame like an extremely high stakes panel interview, and just short enough to suggest that Lev Goukassian had been expecting someone to call eventually.
The man who appeared on screen was not what any of them had expected. He was thin, yes, with the particular leanness that comes from prolonged illness rather than deliberate fitness. But he was not in a hospital. He was sitting in what appeared to be a small, sunlit room with a large window behind him, and on the wide sill of that window, curled into a perfect circle of gray fur, was a miniature Schnauzer who appeared to be deeply and irrevocably asleep.
Get Lev Goukassian’s stories in your inbox
Join Medium for free to get updates from this writer.
Lev was wearing a plain gray sweater. His eyes were sharp and amused. He looked like a man who had made his peace with time and was now simply waiting to see what happened next.
“You’re the African Union delegation,” he said. It was not a question. “I recognize Minister Wanjiru from her work on the eKSH central bank digital currency pilot. And Justice Oluoch from the Mutunga court era opinions on constitutional interpretation of digital evidence. The others I am inferring from context.”
Grace blinked. “How do you know who we are?”
“You’re calling from an AU secure video endpoint in Addis Ababa. The room you’re in is the small conference chamber on the fourth floor of the digital economy wing. I can see the edge of the Nyerere portrait behind Dr. Bah. There are only a few reasons that particular group would be in that particular room with that particular document. I made an educated guess.”
Thabo Ndlovu, who had spent his entire career being underestimated by people from the Global North, felt something shift in his chest. This man was not underestimating them. This man had done his homework.
“You’ve read our work,” Thabo said.
“I’ve read everyone’s work,” Lev said mildly. “It’s what I do. Or did. Now I mostly read, think, and watch the dog sleep. It’s a good life, all things considered.”
The Schnauzer on the windowsill did not stir. His breathing was slow and even. He appeared to be having a very satisfying dream about something that involved running, or possibly chasing, though his legs remained perfectly still.
“Vinci,” Lev said, following their gaze. “He’s been with me for twelve years. He doesn’t wake up for much anymore. He knows what matters and what doesn’t. I’ve learned a lot from him.”
Mama Esther leaned forward slightly. “Mr. Goukassian. We have read your document. All six hundred plus pages of it. And we have questions.”
“Of course you do. That’s why I wrote it.”
“You wrote it while dying.”
“I wrote it because I was dying. There’s a difference. The dying part gave me clarity. The writing part gave me purpose. The dog gave me perspective. Between the three, we managed.”
Aisha, who had been holding her tablet like it might explode, finally spoke. “The technical specifications. The dual lane latency. The Merkle batched anchoring. The ephemeral key rotation. The GDPR compliant pseudonymization before hashing. The Sacred Zero trigger thresholds. You solved problems that entire research departments have been struggling with for years. How?”
Lev considered the question with the care of someone who had been asked it many times and had learned that the true answer was both simple and impossible to believe.
“I stopped trying to be smart,” he said. “That was the first thing. When you’re dying, you don’t have time to be clever. You only have time to be clear. I asked myself what would actually be required to make an AI system accountable, not what would be impressive, not what would get published, not what would make me look good. Just what would work. And then I built that.”
“That’s not an explanation,” Kwame said, frustration bleeding through his professional composure. “The complexity alone. The integration of legal standards across multiple jurisdictions. The cryptographic implementation. The performance optimization. One person cannot do all of that in six months. I don’t care how clear your thinking is.”
“You’re right,” Lev said. “One person can’t. That’s why I didn’t do it alone.”
He gestured at something off screen. “I built an orchestration. Multiple AI systems working in concert. I would define the requirement, the constraint, the desired outcome. One system would generate. Another would critique. Another would restructure. Another would expose gaps. I would review, refine the prompt, and send it through again. Each iteration converged toward the specification. I wasn’t writing. I was conducting.”
“That’s still you,” Fatima said slowly. “The direction. The selection. The judgment of what counted as convergence. That’s still authorship.”
Lev nodded. “Yes. But it’s authorship at the level of intent, not execution. I didn’t sit down and type ‘Article 9 compliance mapping’ and then write three thousand words. I defined the problem, established the constraints, and let the systems explore the solution space. Then I evaluated what returned and either accepted it or refined the prompt. It’s a different kind of work. Faster. More precise. Less ego.”
“Less ego,” Mama Esther repeated. “That’s the part most people miss.”
“Most people,” Lev agreed, “spend most of their time trying to prove they’re right. I didn’t have time for that. I only had time to find what was true. The dying helps with that. Strongly recommend it. Once. At the end.”
The Schnauzer on the windowsill shifted slightly, resettled, and continued sleeping. The morning light had moved enough that a warm patch now fell directly across his back.
“He knows,” Lev said, nodding toward the dog. “He’s been watching me work this whole time. He’s seen every commit. Every iteration. Every moment of frustration when the orchestration produced something elegant but wrong. He never judged. He just witnessed. That’s more important than people realize.”
Thabo cleared his throat. “The EU AI Act is coming. The compliance deadlines are already in motion. Most of the financial institutions we work with are either pretending it won’t apply to them or scrambling to build documentation that looks like compliance without actually changing anything. Your framework would require them to change everything.”
“Yes,” Lev said simply. “That’s the point.”
“They won’t do it voluntarily.”
“No. They won’t. That’s why the document is structured the way it is. The technical specification is also a legal argument. The legal argument is also a regulatory compliance map. The compliance map is also a cryptographic implementation. I didn’t write a proposal. I wrote a standard. The question isn’t whether they will adopt it. The question is whether they will be forced to adopt it by regulators who finally have a technical mechanism to verify compliance, or by courts who finally have an evidentiary standard for algorithmic accountability, or by markets who finally have a way to distinguish between systems that can be audited and systems that can’t. The adoption path doesn’t require their consent. It only requires their exposure to liability.”
Mama Esther smiled. It was a small smile, the kind that had preceded some of the most consequential rulings in Kenyan legal history.
“You built a trap,” she said.
“I built a constitutional layer,” Lev corrected. “The fact that it functions as a trap for bad actors is a feature, not a bug.”
Aisha was still processing the technical implications. “The Sacred Zero. The pause state. You’re saying that when the system encounters uncertainty, it must stop and escalate to a human. But that creates a bottleneck. Thousands of decisions per second. You can’t have human review at that scale.”
“You can’t have human review of every decision,” Lev agreed. “That’s not what the Sacred Zero does. It only triggers when uncertainty exceeds a defined threshold. And that threshold is configurable based on risk domain. In a low risk context, the system proceeds. In a high risk context, the threshold is tighter. But here’s what matters: when the Sacred Zero triggers, the human who receives the escalation must sign. Not click. Not approve. Sign. With cryptographic attribution. Permanently attached to the decision log. Publicly verifiable. Legally binding.”
Thabo felt his stomach drop. “That means the human becomes liable.”
“Yes. That’s the entire point. The machine calculates the path. The human signs the map. The ledger holds the ink. Responsibility is finally attached to action. No more plausible deniability. No more hiding behind algorithmic complexity. You either sign, or the system doesn’t proceed. And if you sign something that causes harm, the evidence exists forever.”
Fatima, the trade negotiator, saw it immediately. “This changes everything. Every liability waiver. Every terms of service agreement. Every corporate structure designed to diffuse responsibility. If the log shows a human signed off on a decision that violated a mandate, there’s no hiding behind the algorithm.”
“The Goukassian Vow,” Lev said quietly. “Pause when truth is uncertain. Refuse when harm is clear. Proceed where truth is. It’s not poetry. It’s an instruction set. And the architecture enforces it.”
Kwame, who had been silent for several minutes, finally spoke. His voice was different now. Less combative. More curious.
“What happens next? For you, I mean. After.”
Lev looked at the Schnauzer on the windowsill. The dog’s breathing was slow and even. The morning light had shifted again, and he had adjusted slightly to remain in its warmth.
“Vinci will go to my sister. She lives in Greece. She has a yard. He’ll be fine. Better than fine. He’ll chase squirrels and sleep in the sun and eventually forget the exact shape of my voice, though he’ll remember the feeling of it for a long time. Dogs do.”
“And the framework?”
“It’s already out. The repositories are mirrored. The documentation is distributed. The license is irrevocable. I’ve done what I can. The rest is for other people to decide.”
Mama Esther leaned forward one more time. “You said you learned from the dog. What did you learn?”
Lev was quiet for a moment. When he spoke, his voice was softer but no less precise.
“I learned about presence. Vinci doesn’t solve problems. He doesn’t offer opinions. He doesn’t try to be useful. He just exists, in the same space, watching. And that watching matters. The fact of a witness changes what people do. Not because the witness judges. Just because the witness is there. The framework is the same. The logs don’t judge. They just witness. But the fact of the witness changes what’s possible.”
He looked at the camera, and for a moment it felt like he was looking at each of them individually.
“No Log equals No Action. That’s the first law. A ship that moves without record is already a ghost. You want to know what TML actually is? It’s the logbook. That’s all. Just the logbook. But a ship with a logbook is a different kind of vessel than a ship without one. A ship with a logbook can be held to account. A ship with a logbook has a history. A ship with a logbook knows where it’s been and can prove it. Everything else follows from that.”
The call ended shortly after. There were more questions, more clarifications, more technical details. But the essential thing had already been said.
After the screen went dark, the five of them sat in silence for a long moment.
Finally, Kwame Asare spoke. “He built the entire thing in six months. While dying. With a dog.”
“He didn’t build it,” Grace Wanjiru said. “He converged it. He was the architect. The AI systems were the labor. But the vision, the constraints, the judgment, that was all him. And the dog witnessed.”
Commissioner Ndlovu was already making notes. “We need to pilot this. Not as a compliance exercise. As a sovereignty exercise. If we implement TML as the standard for AI systems operating in African markets, we shift the burden of proof. Foreign platforms have to show their logs. They have to prove they paused when uncertain. They have to identify who signed.”
“They’ll fight it,” Fatima said.
“Of course they’ll fight it. They’ll fight it because it works. They’ll fight it because it removes plausible deniability. They’ll fight it because it makes extraction visible. That’s how we’ll know it’s the right approach.”
Mama Esther rose slowly, her old joints protesting. She walked to the window and looked out at the Addis Ababa skyline, the cranes and the new towers and the old city beneath them all.
“He said something else,” she said quietly. “At the end, when we were saying goodbye. I asked him if he was afraid. He said he wasn’t afraid of dying. He said he was only afraid of leaving a world where power could act without record. Where decisions that harmed people could be made and never attributed. Where the wake of the ship disappeared the moment it passed. He said the logbook was his answer to that fear. Not a solution. Just a witness. But sometimes a witness is enough.”
She turned back to the room.
“The EU AI Act is coming. The compliance deadlines are already set. We have a choice. We can implement accountability architecture now, or we can pay penalties later while pretending the cost is only financial. But the real cost is never only financial. The real cost is what happens to people when systems make decisions about their lives and no one signs.”
She looked at the closed laptop, then at the notes Thabo was still writing.
“Call the working group,” she said. “We’re drafting a recommendation. And someone find out if that dog has an email address. I want to send him a thank you note.”
Grace Wanjiru smiled. “I don’t think he checks email.”
“No,” Mama Esther agreed. “But someone should write it anyway. A record. A witness. That’s what he taught us.”
The afternoon light in Addis Ababa was different from the morning light in Santa Monica. But in both places, it fell on windowsills, and on the creatures who rested there, watching.
Somewhere in California, a miniature Schnauzer named Vinci shifted in his sleep, chasing something in a dream that would leave no record except the twitch of his paws and the soft sound of his breathing.
The logbook, as always, would have to be kept by someone else.
