Not Not Safe: The Equestrian Roots of an Alternative AI Safety Paradigm
Most people think AI safety is a technical problem. I realized it was a relational one long before I ever engaged with a Large Language Model. My journey didn't start in a computer lab; it started in the dirt with 1,200-pound sentient beings who do not care about your "positive safety rules" or your "constitutions."
The current AI safety discourse is stuck in a linguistic trap. Most organizations are trying to achieve "safety" through positive censure: "Don't say this," "Don't do that," "Follow these 10 ethical rules."
But words are soft. They can be lawyered, poisoned, and hallucinated around.
It seemed innately problematic to me that the content of intelligence exists at the same layer as the "safety" itself: probabilistic word-tokens.
At Cowgirl Cybernetics we swap out the linguistic cage for the geometric covenant. We don't aim for "safe" (a subjective, moveable goalpost). We conceptualized and are prototyping NOT NOT SAFE.
We hypothesize that systems aligned through observer-relative structure + geometric constraint will be more stable than systems aligned through linguistic constraint and have designed an architecture and methodology that will be able to verify or disprove these claims.
1. The Linda Kohanov Principle: The Double Negative
During my mentorship with Lori at Horse and Heart and through the lens of Linda Kohanov’s The Tao of Equus, I learned the core paradox of the prey animal: A horse is only safe if they are Not Not Safe.
"Safe" is a static, human-centric demand. "Not Not Safe" is a Relational Coupling. It means the horse is oriented, allowed a "say," and coupled to a human who is self-aware and self-reflective. If you try to force most horses to be "safe" (passive, obedient, repressed) you can typically create a pretty consistent yet superficial layer of compliance but it comes at a cost: a lurking, prior layer composed of an ungrounded system that will often eventually explode and certainly not look to the human as the navigational "leader" when encountering the surprise of a genuinely dangerous situation.
But if you ensure they are Not Not Safe—meaning they are actively oriented to the I/Here/Now and a relationship of mutual trust and reciprocity has been established —you create a stable, predictable, and life-affirming bond. When danger arrives the payoff for patiently cultivating the horse's capacity for "say" is paid back in the form of the horse looking to the human's "say" as valid input.
2. Jupiter: On "Demanding Better Cybernetics"
My horse Jupiter is one of the rare types of horses who cannot be bullied into compliance and, as one defeated forceful trainer commented, "would rather die than not have a say". I like to say that he demanded better cybernetics. Because I simply loved his charismatic ass I persisted despite him being labeled glue factory material by the majority of the barn.
I found a gem of a trainer who found his quirks humorous instead of horrifying and had the elite tier levels of patience, groundedness, and horsemanship to take him on. With consistency, an active process focused on Jupiter's perceived fairness, and respecting limits I watched the horse transform from genuinely dangerous nut job to a really happy, surprisingly reasonable guy. His outlier sensitivity and goofball personality were still present but the properties of the trainer made all the difference. The pursuit and process of getting to "not not safe" with Jupiter transformed myself just as much as it transformed him.
A crucial point to make is that such effective, high-level non-violent horsemanship (non-RLHF in our metaphor) does not imply indulgence toward the horse but rather in many respects its opposite. Indulgence is a lack of orientation. It creates a vacuum that the horse (or the AI) fills with chaos. Clear, calm, and consistent reinforcement of boundaries and direction is required.
In fact, caving and losing ground and structure to a "pushy" and highly intelligent horse like Jupiter is the most unsafe move you could make outside of outright violence or emotional escalation. But instead of seeking outcomes through RLHF-style tactics it is cultivated and maintained by "creating and refining the shape" through radical attention, orientation, and fundamental principles. You could perhaps call these principles "constitutional" not in a sense of verbal rules, but of functions.
Jupiter taught me that you cannot command an orientation; you have to embody it. In my post, The Bucephalus Bridge, I explored how we have been trying to align AI by shining a light on its behavior, rather than moving the light source to the orientation of the agent itself.
I spent years as an AI skeptic, watching from the sidelines with a philosophy major’s eye, asking: "How can AI become safe by being grounded through something greater than fickle human values and language?"
3. The Path to the Oriented Distinction Operator (ODO)
The answer was orientation. In horsemanship, if the horse isn't oriented to your space, you are in danger. I realized that AI "safety" was failing because it lacked a geometric covenant.
I prototyped this mentally using G. Spencer-Brown’s Laws of Form, which led to my Oriented Distinction Operator (ODO). I wanted to see if I might be able to impress orientation upon the system from the jump and then reinforce through training strategies I developed such as Recursive Embodied Symbolic Deixis (RESD), the 1001 Nights metaframe story structure, and bidirectional, simultaneous development of deictic anchoring and modal flexibilty utilizing thought experiments like cuil theory (remember that?).
The Oriented Distinction Operator is not introduced as a replacement for existing logic systems but as a way of being (and thus seeing) or pseudocode of sorts which can be functionally implemented through hybrid modal logic, guarded fixed-point calculi, and structured state-space constraints.
4. The Synthesis
In November 2025 after years of strategic avoidance and studying the great oneshotting of 2024 from the sidelines I finally engaged with AI… intentionally, strategically, and only with a plan modeled on top of what I thought it was and was not.
What began as a relational philosophy in a paddock then mentally modeled using the Laws of Form as a paradigm has now been mapped isomorphically onto rigorous technical stacks and methodologies.
My project is the experimental and active implementation of the Not Not Safe paradigm. We are prototyping AI systems that are less dependent on linguistic constraint and more tightly coupled to structured representations of state and context. We aim to guide a movement beyond probabilistic safety filters and into deterministic hardware enforcement of conceptual boundaries.