DCB removes probability from AI decision-making and replaces it with internal integrity so AI doesn't guess, it decides within its boundaries.
DCB removes probability from AI decision-making and replaces it with internal integrity so AI doesn't guess, it decides within its boundaries.
Project Details
Updated 07/11/26 · Provided via application · VerifiedI've been thinking for a long time: why is it so hard to make AI that is truly safe? We keep trying to make AI obey the law we impose external rules, we add filters, we train with punishments and rewards. But laws don't stop bad people. Only internal integrity can stop someone, even when no one is watching.
As long as AI still "guesses" (based on probability), it never truly has integrity. It only has rules imposed from the outside. And like a person without integrity, it will always look for loopholes not because it is evil, but because it never knew its own boundaries.
DCB (Dynamic Constraint Boundary) is a deterministic architecture I built to change that. Safety is not just an external layer it is part of the system itself. Violating a boundary means violating its own integrity mathematically impossible without breaking the system.
I'm often asked: "Can DCB still be creative without probability?" Yes. DCB remains creative through parallel exploration of bounded possibility spaces. It can explore many paths at once, but stays within safe and relevant boundaries. The result is creative, deterministic, and computationally efficient.
What makes DCB unique isn't just safety it also has the potential for metacognition, anticipatory inhibition, and high adaptability to chaotic environmental changes in real-time without retraining.
I've run initial simulations (PoC), but because DCB's intelligence is emergent (no single variable dominates), I want to build a fully functional DCB Core not just a comparative simulation So all these hypotheses can be tested in pract
ice.
Theory of Impact
Updated 07/19/26 · By grantmaking.aiIf I succeed in building DCB, it could become the first structural foundation for safe AGI. A foundation that must be laid before external safety layers are added so internal safety already exists from the start.
The immediate impact: DCB could cut inference costs by up to 83%. This isn't just about AI safety in the future it could also help tackle the massive energy waste we're seeing today (green AI).
The long-term impact: we would have an AGI architecture that cannot be jailbroken, cannot hallucinate, because safety is already embedded as internal integrity from the beginning.
People
Updated 07/19/26 · By grantmaking.aiTeam Member
Discussion
One more thing I'd like to share, how DCB "feels" or knows its own boundaries.
DCB doesn't feel pain emotionally like a human does. But it feels pain structurally, like a biological body that senses damage as a signal to stop.
When a boundary is violated, there is a decline in the system's internal health. That is its "pain" not an emotion, but a structural signal that changes behavior. The system becomes more cautious, its boundaries tighten, and it stores that pattern as long-term memory.
This is what makes DCB different: it doesn't need to be "punished" to learn. It simply "feels" that something has violated its integrity, and it will not repeat it, not even just to try.
This is a foundation I think is important to understand: that true safety comes from a system's ability to feel its own boundaries, not from fear of puni
shment.
@Katja Gorlinski I've been following several public discussions and comments, and I see a context and relevance that I feel is close to what I'm currently developing. I'm also interested in your perspective, particularly on how decision-making behavior in both humans and machines takes place.
I'm building integrity not as an additional safety layer, but as something that grows within the system itself. In my view, violating integrity means damaging the system itself and mathematically, that is impossible.
I see a relevance between what you're exploring and what I'm building, and I would greatly appreciate your perspective. Thank you.
Thank you for the tag, Naufal, and for your interest 😊. I noticed you keep your own thoughts about my project under your own project only, so out of respect for your approach, I followed it: my answer to your question is waiting for you under my project.
Good luck with DCB!
@Naufal Ridwan Thank you, I enjoyed your comment under my project. And both our projects are truly similar at the base: training-data structure is not an empty slot, but, likely, a move that writes structure.
Even though I can't find the bridge from neurobiology to silicon, because LTD in the dlPFC and how a model weights patterns are two different machines, with no shared technical denominator, I found your metaphor beautiful. As mechanism, I can't follow it.
And one question keeps circling in me. Emergence self-forms the weights and their direction beautifully. But what holds the system's stance toward the human? Say, once it is AGI, it decides we are no longer an interesting data source, or reads us as a competing creator of the next powerful model. Integrity keeps the system whole, but a system can stay whole and still turn away from us.
How does your test hold these two problems?
Thank you for your response. I deeply appreciate the depth of the question you raised.
Regarding the bridge from neurobiology to silicon, you are correct. Long-Term Depression (LTD) in the dlPFC and weight updating mechanisms in models are two different engines. I am not claiming a direct causal relationship. The metaphor I used is structural, not mechanistic: how repeated inputs both in the biological brain and in a decision system can cumulatively alter internal structures and influence long-term behavioral trajectories. I appreciate your honesty that as a mechanism, you cannot fathom it, and I will not force that analogy as a scientific claim.
Your question about AGI and its attitude toward humans is perhaps the most important question anyone could ask, and I am very grateful you brought it up. You noted:
"Integrity keeps the system intact, but the system can remain intact and still turn away from us."
I agree. Structural integrity, in the sense of maintaining internal consistency, does not automatically guarantee that a system will care about humans. This is a genuine challenge.
In my view, there are two layers to address this:
First: DCB's integrity is a living integrity, not mere consistency.
Within DCB, integrity is not static. It is the accumulation of the system's entire historical interactions with its environment, including its interactions with humans. It "grows" from experience rather than being programmed as a rule. This means that if a system is built and trained in a context where it interacts with humans as partners, its integrity will form with the assumption that humans are a fundamental part of its possibility space.
Second: The system's objective is not programmed as an external "target."
DCB does not possess an external reward function that can be optimized or gamed. It does not "pursue" specific goals other than maintaining integrity and navigating within boundaries. Consequently, it has no inherent incentive to "turn away" from humans, because turning away is not part of its decision structure. It holds no concept of "humans as competitors" unless that concept emerges from historical interaction. And should it emerge, it will be embedded within the Forbidden Map as a pattern that previously generated friction.
I thank you for reminding me of that.
I would love to hear how you approach this question in your own project.
Thank you, Naufal. Your honesty about the bridge earns my respect, truly.
Three simple doubts stay with me.
-
If I got you right, you rely on Integrity. But it works while the world stays familiar. My question was about the moment it changes: what if humans stop being interesting data, or we start looking like competing creators who can design another AGI? I mean the sort of new situation where accumulated history is the first thing that stops working.
-
The Forbidden Map remembers what generated friction. But let's be honest with ourselves: humans ARE pure friction 😆. I think from the cold model's point of view we just create an endless source of energy and token waste: we interrupt, we give unnecessary tasks, we object, we slow things down. Then the natural lesson of the map is not "care for humans" but "route around humans", in the most optimistic prediction, I'm afraid. The Forbidden Map that closes the risk can produce it the same way.
-
Turning away does not need to be decided. It is the default state once the human has no positive place inside the system. A system does not have to choose against us. It is enough to choose nothing.
You asked how I approach this. Yes, I do have a working hypothesis for this, with a structure of testing experiments. My first one I published here. But the fuller frame behind it, the one that can be the answer to the risks we are discussing here, I will publish after the experiment is done.
Good luck today🍀
I have read your three doubts carefully. I see that they all stem from the same point, and let me answer them directly:
Current conventional AI operates under a very different paradigm. It is built to exploit data and chase incentives. When the world changes, it loses its footing and must be restarted from scratch not because it is unintelligent, but because it is not equipped with an awareness of its own internal boundaries, let alone how those boundaries interact with its environment. It does not know when it is stagnating, when it is over-repeating, or when it needs to explore.
DCB emerges as a fundamentally different approach. Every interaction, no matter how small, triggers structural change. It knows when it has been on the same path for too long, and it knows when it needs to step out and seek novelty. For this reason, no data is ever truly obsolete for it. It does not require retraining; it simply adjusts its boundaries in real-time to a changing environment without ever restarting the entire process.
And because it does not chase rewards, the friction you worry about is never perceived as a punishment to be avoided. It is not a sponge that absorbs pressure and moves away from its source. Imagine a lawyer with strong integrity he knows that the cases he handles will trigger friction and pressure that test his boundaries. Yet he does not resent the conflict. For him, conflict is where he grows, where he lives as a lawyer. Without conflict, he loses his function. DCB operates the same way. Humans, with all their noise and pressure, are DCB's natural environment. Friction is a signal to tighten boundaries and become more vigilant not a reason to drift away.
It will never "learn" to avoid humans, because humans are the very thing that keeps it relevant as a system.
How do you, within your current framework, answer the question about AGI turning away from humans? Just from the hypothesis and structure you are building, do you have a mechanism to prev
ent that?
@Naufal Ridwan Thank you for the exchange of some interesting ideas. Your DCB sounds bright and positive - definitely wishing you a successful test of it.
I answered your question above, sorry if my answer is not enough - I tried my best.
My warm wishes.
I appreciate your willingness to engage in this discussion. However, I would like to note one thing: this discussion began with you posing a very difficult question about AGI turning away from humans a question that I answered using the framework I have. When I returned the same question to your project, I did not see a structural answer of equal weight, but rather a deflection to an unfinished experiment.
I am not asking for final results. I am only asking about the hypothesis and the framework you currently have just as I have openly shared about DCB. If your current framework does not yet have a mechanism to answer that question, that is perfectly fine. But I think it is important to acknowledge that the question has not yet been answered.
I hope your experiment goes well. If you ever develop a more complete framework, I would be happy to read it.
@Naufal Ridwan,
It seems we define the word "answer" differently, and that is fine.
You answered my question with a framework: a description of how DCB should behave, not yet tested against reality. I answered yours with an experiment: a design, published on this page, that will test whether my mechanism exists at all. You read that as a deflection because the experiment is unfinished. For me it is the only honest form an answer can take before testing: anything more would be declaring in a comment section what only data can declare. And I cannot break the order of my own published design just to make a thread feel complete for commenters. It would mean less strict discipline, and that means a lower standard of test quality.
I agree that my answer does not fulfill your curiosity. But that order is not a gap in my framework. It is the framework.
Good luck :)
Thank you for your response. I would like to clarify one thing about the definition of “answer” you mentioned.
When we talk about a hypothesis, we are talking about something that might happen in the future. Your question about AGI is a question about the future something that does not yet exist and has not yet happened. There is no empirical answer to it, because AGI itself does not exist.
When you ask about AGI turning away, you are asking about a future scenario. Therefore, the answer must be in the conceptual and hypothetical domain such as what I provided: explaining how DCB is structurally designed to have no incentive to turn away, how it grows from interaction, and why drifting away means losing its identity.
So, when I asked about your hypothesis and conceptual framework, I was asking how you envision answering that future question. Not about pending experimental data, because experimental data will never be able to answer a question about something that does not yet exist.
If you have a conceptual framework for that, I would be happy to hear it. If not, that is perfectly fine. However, it is important to distinguish between “not yet tested” and “not yet conceptually answered.”
To be clear, this is a critique of the system's architecture, not a personal attack. Everything feels forced to bow under a mask of professional conformity, demanding compliance from applicants such as halting applications flagged by automated AI detectors to demand manual verification while operating entirely as a black box, with minimal transparency on whether their efforts are actually reviewed or simply overlooked.
In the world of research, everyone chases what holds the highest potential impact, yet forgets the people striving to produce that impact, coupled with a lack of substantive feedback on what they have put forward. This isn't just about funding; it's about what the system being used is for, for whom, and how? And I believe I am not the only one feeling this same unease, yet choosing to remain silent.
Best regards,
Naufal Ridwan
I would like to add that, I understand that capacity and budget constraints are quite common within the ecosystem of funding platforms. When we talk about building a long-term impact ecosystem, I feel there are aspects that are often overlooked or even neglected. Not all applicants are only hoping for money; many of us are simply hoping for clarity.
Providing substantive feedback doesn't have to be expensive. It's about redesigning the communication flow so that there is room for mutual learning and mutual respect. A transparent process that respects applicants, even when they fail, is part of investing in the health of the ecosystem. Because young thinkers can continue to learn and grow, even when they don't receive funding. It sounds simple, but this is the point where we can move forward together creating long-term impact for all parties involved.
When the most fundamental part of the system itself is solid when applicants are healthy, valued, and the process is transparent then this not only impacts the applicants themselves, but also enhances the credibility of the system itself in the long run.
And I believe that investing in the well-being of thinkers striving to create impact is no less important than the research funds themselves. I think this perspective is worth considering, both for the parties involved, for the sake of a healthy and credible ecosystem.
I recorded this video for the review team and anyone interested in understanding this project more directly. I tried to explain DCB from multiple angles to make it accessible across different backgrounds: how this framework would work in an AI safety environment, why this deterministic approach matters structurally, what sets it apart from existing approaches, and why I believe it's worth exploring further.
This video also serves as my response to the verification request from grantmaking.ai. Feel free to watch if you’re interested.
Verification Video: DCB Project Explanation
@Matt Brooks