Generative AI have been called “stochastic parrots.” I find the framing clever, though I think the verbiage undersells the math and overestimates the mimicry — or nearly insults the parrots. I once told my colleague that “LLMs are like autistic intern savants”. I too might have undersold the neurodivergent metaphor while nearly insulting those in the spectrum.
I quite like the operational characterization I first heard from Terence Tao but I believe is widespread: trying to wrangle AI is like “herding cats.”
“So where the AI really struggles right now is knowing when it’s made a wrong turn,” he says in an interview. And any feline enthusiast can easily point out that any mistake a cat makes, it will pretend to have done on purpose. Herding cats is arguably harder than organizing humans. Making AI follow nuanced, subtle instructions on complex or wicked problems is probably harder than doing the same with, say, an intern on the spectrum.
See, most reasoning systems are mirrors.
They show you what you already know, faster, cheaper, at scale. The output quality seems to track input quality — which means they can’t actually help you think better, only execute faster. Or bigger.
And in the same vein, they make bigger mistakes faster.
Our proprietary Structural Reasoning Heuristics is not a mirror. It’s more like a crystallization; closer to a grammar synthesized from our own ways of thinking about knowledge representation, epistemology, and ontology. A grammar for reading the structural logic underneath any situation, domain, or body of knowledge — and using what you find to illuminate more insights about the situation or building on top of it if the foundation is stable enough. The difference isn’t subtle. It determines whether your AI implementation has a ceiling.
What does it do?
Ask what a reasoning system is for, and most answers point at outputs. Better decisions. Sharper strategy. Faster analysis. These are downstream.
The prior question is structural: what does a reasoning system actually do to a problem that makes better outputs possible? Not what does it produce. What does it perform?
Reasoning Heuristics emerged from that question. One of its foundational principles is to treat meaning as a vector: something that points toward a purpose and derives its identity from where it aims, not from what it contains. The operational consequence is simple and initially counterintuitive: stop asking “what is this?” and start asking “what is this for?”
What happens when you ask that question at scale across a body of knowledge is that the tacit structure becomes visible. What was decoration separates from what’s load-bearing. What seemed like a definition reveals itself as a relationship. What looked like good judgment turns out to be pattern recognition — architecture that can be made explicit.
Another foundation principle is recursion: applying the same pattern to its own output, all the way down, until you hit bedrock. Every analytical move the system makes can be pointed back at the question of how to choose analytical moves. This isn’t a clever design feature (though it is). It’s why the system keeps working at arbitrary domains — the same reasoning process that reads a supply chain reads itself.
What the Reasoning Tectonic is: a framework of intuitions, workflows, and principles for knowing the structure underneath natural language, formalizing its purpose as a knowledge representation and reasoning system.
What it does, in a bit of a roundabout way, is category theory.
It lets you walk into a building and see the blueprint. You can keep using that building. But if you know its blueprint, you can build others like it. You can append to it, modularize parts of it, make it leaner. Or have others do any and all of these.
The Intuitions (eleven, for now)
The namesake does purposeful work. The Intuitions are orientations, not procedures. They train perception before method is chosen and change what becomes visible before you decide what to do about it. They aren’t a checklist run in sequence — several apply simultaneously to any sufficiently complex structure, and when two of them read the same object differently, that disagreement is itself a finding, not a bug to resolve.
The foundational principles mentioned above — asking what something is for, and applying every pattern to its own conditions of application — are two of the 11 “Intuitions” within the Reasoning Heuristics.
Teleological Signification is the orientation that treats meaning as a vector.
Every element in a system — a word, a role, a metric, a policy — has a function, and that function is defined by what it’s aimed at, not by what it is. The move is: what is this doing here? What’s load-bearing separates from what’s decoration. A concept pointing nowhere isn’t yet a concept — it’s noise waiting for conscription into signal. And critically, this holds across domain shifts: a purpose can survive being re-clothed in new vocabulary, accelerated, slowed, or moved into a structure it didn’t originate in, and still be recognizably the same telos.
Recursion is the orientation that the pattern applies to its own output.
Certainly not novel and irrevocably important as a concept in various fields, I first encountered recursion in the works of M.C. Escher and later on at university when I was “studying” computer science. SupraGraphos applies it to semantics, structure mapping ala-Gentner, and ontology logs.
We’ve found that practically every stuck analysis has a level of abstraction above it where the actual problem lives, i.e. interventions “downstream” are dealing with the smoke and not the fire. That the bedrock is almost always deeper than the first read. That the recursion has a base case — but the base case isn’t an exit condition the way it is in code. It’s a floor the structure itself reveals: the point past which further unfolding produces no stable meaning relative to what you’re actually trying to do. Finding that floor is a discovery. Deciding to stop somewhere shallower than that, because your purpose doesn’t need the extra depth, is a different act entirely — a choice, not a discovery — and the two are easy to conflate if you’re not careful.
I’ve been iterating on the Intuitions and they’ve gone from six to eight, nine to eight again, to 11 — and so far the current version has resisted changes or absorbed mutations.
They fall into four “quadrants”:
The first handles how meaning gets located and how far you can trust it.
It involves treating purpose as a vector rather than a definition (Teleological Signification, above), alongside a matching pair of ideas: that sense-making needs more than one anchor held at once to actually click into place, and that meaning stabilizes locally but can be zoomed in and out at different resolutions rates like a camera. Finally, again as mentioned above: the same recursive read which works on a domain also works on the read itself. Four intuitions, one job; get the epistemics right before anything else is attempted.
The second is quieter and does less talking, but everything above stands on it: identity is the same thing read at different scales.
A word, a sentence, a document, a domain — none of these is a container with meaning tucked inside it. Meaning shows up the moment a boundary actually holds, and it stops showing up the moment that boundary doesn’t. Knowing which scale you’re standing at, and whether the boundary you’re relying on is real or assumed, quietly determines whether everything else in the framework is pointed at something solid. This is the culmination of the first quadrant, in a way.
The third supplies actual machinery.
The short version: you can’t see anything directly, only through the totality of what it exchanges with everything around it — which means a handful of formal moves become available once a domain is confirmed to behave like a structured network rather than a pile of facts. If that sounds familiar it’s because it’s an operationalization of the Yoneda lemma of category theory fame (or infamy). In this quadrant, too, we utilize this is where four recurring Yoneda probes — timing, attention, cost, and (Epicurean) swerve — each one a lens that reveals something different depending on what you point it at.
The fourth quadrant takes those same probes and runs them against the four environments where structured knowledge actually “lives“.
These four are nested temporal strata (multiple clocks always running at once, rarely in sync but co-causative and co-generative), attention (where structural weight pools and discharges), exchange (every domain has a cost structure, where resource scarcity creates economies of things other than money), and origin (what looks like bedrock is usually just an old contingent choice that hardened into “how things are”). Same probe, different environment, different insight — which is the point. It’s what keeps the system generative instead of prescriptive.
Eleven Intuitions, four quadrants, for now.
The workflows
As the intuitions are more like orientations, composing them with each other begins to feel like cartography.
Meanwhile, repeatable frameworks of Intuitions used together, in-sequence, nested within each, or in some other configuration are also included in the Structural Reasoning Heuristics. These workflows are split between:
- Reusable Subroutines, which are micro-operations that appear repeatedly inside the larger workflows
Like finding the single load-bearing claim in any system — the proposition whose removal leaves everything else without a target.
Or reading what a substrate is currently selecting, independent of what you’d prefer it to select, where probable and possible are different categories and the key is reliably finding the difference.
- Larger workflows or Protocols — complete analytical sequences, each composing multiple intuitions in purposeful order — often with additional “moves” within, each built for a specific class of structural problem.
Like identifying the terrain before any other analysis can meaningfully run: what kind of structural configuration is this, actually?
Or mapping the structural logic of what happens when multiple actors share the same field with genuinely incompatible orientations.
One handles the full architecture of a strategic or social situation modeled as nested boards across temporal strata — each with its own actors, economies, and probable moves, each governed by different rules.
Another is framed within a Chinese concept: the Weiji. The term wēijī (危机) is usually translated as “crisis.” What it actually encodes is “dangerous pivot point” — a loaded threshold where trajectory can change. The workflow framed within this concept identifies where structural pressure has been accumulating, determines whether a pivot has already fired or is still approaching, and locates the right depth of intervention.
The Insight Taxonomy
After analysis runs, there’s a second question: not what did we find, but what kind of thing did we find?
SupraGraphos’ internally maintained Insight Taxonomy is a living encyclopedia of invariants, where each new analysis or project validates, enriches, revises, or offshoots from the existing enumeration.
Not every structural problem belongs to the same invariant class. Deploying the right intervention for the wrong class is how a correct diagnosis produces no change. The taxonomy exists to prevent that.
So far, we’ve seen six structural patterns recur, predictably, across domains. A few you’ll likely recognize. For instance:
- Everything being done is technically correct, and nothing is compounding.
The goal isn’t unreasonable. The resources aren’t absent. The execution isn’t poor. The problem is architectural: every intervention is landing at a different timescale than the outcome it’s aimed at. They’re not failing — they’re succeeding somewhere other than where the target is. Which is its own kind of failure, because it’s invisible. The efforts show up in reports as activity, not as movement.
You see it in brand-building treated as a quarterly deliverable. In culture change measured at the end of the initiative. In sales enablement programs that produce activity metrics while lead pipeline velocity stays flat. The work is happening. The timescale it’s aimed at isn’t.
- What the organization says it believes and what its behavior consistently reveals are structurally different things.
The gap isn’t hypocrisy, and it won’t close through clearer articulation of values or more (or even better) documentation. Beneath both the stated position and the repeated departure from it is a founding commitment — something that was never made explicit, never formally adopted (and therefore neither documented), but that consistently governs what actually happens when the stated value and real-world pressure meet. Until that commitment is named and located, every culture initiative reinforces the stated belief and leaves the operative one untouched.
This failure is exemplified in the company that genuinely believes it values psychological safety and consistently promotes people who don’t model it. Or the firm with innovation as a core value and a resource allocation process that structurally selects against unproven bets. The commitment beneath the surface isn’t a lie. It’s just older, and louder, when it counts.
- The intervention requires the exact capacity it’s supposed to build.
The people who already have the capacity succeed with it. The people who need it most encounter a circular requirement: to benefit from the intervention, you’d need to already possess what the intervention is for. More of the same doesn’t break the circularity — it deepens it.
The pattern is common in training programs, onboarding structures, and anything built for adoption by people who don’t yet have a scaffold to adopt from. Think about organizations where the feedback culture rollout asks people to be vulnerable first — where being vulnerable has historically been costly.
Six, for now. And we’re evaluating a potential seventh.
Each has a characteristic failure mode, typed interventions that will reliably produce nothing, and a class that operates on the mechanism rather than its symptoms.
There’s a shape to the gap your situation has. The taxonomy is built to fit it — even before you know what the shape is.
What is this for?
That question is where the Praxis started. At the risk of waxing poetic, it’s also where it ends.
Your expertise is already structured. The domain knowledge you and your best people carry — the pattern recognition, the judgment calls, the intuitions that look like talent and function like architecture — that structure exists. It’s just tacit. Hasn’t been made explicit enough to be transferred, encoded, or operationalized beyond the people who currently hold it. Or those who can emulate it.
That’s the gap the Structural Reasoning Heuristics closes.
Not by extracting knowledge into a database (though there’s that part, too). But by making the structural logic — the relationships, the types, the load-bearing claims, the gaps that specify what’s missing — precise enough that it can be operated on. By humans working with it. By AI systems being asked to apply it. And eventually, by models smaller and more efficient than anything the current AI market considers viable. Because when the structural layer is explicit enough, the model doesn’t need to compensate with scale.
The method is the product. And the method points at itself.