Published on: July 8, 2026
Ready to accelerate your breakthrough? Send yourself an Un-Robocall™ • Get transcript when logged in
Send Strategic Nudge (30 seconds)Green in-lane · amber a little out · red drift. Every panel is a real commit, byte-identical on recompute. Tap any panel to open its shareable receipt.
Before the claim, the artifact. Run this one command on your own machine — not because we said so, because you can:
![tolerance panel for commit 21ebef0 — chore(blog): attach commit tolerance panel as OG image [panel-attached]](/blog/post-panels/2026-07-08-the-warren-harding-test/21ebef0ab38550be9dfcbbf21a14646fe5ab456c.png)
npx thetacog-mcp attest-demo
Here's exactly what that does: it hires no model, asks no chatbot for its opinion of itself, and returns a signed, recomputable coordinate — where a piece of work actually landed on a fixed 144-cell lattice, not where its own output claims it landed. Run it twice and you get the same answer twice. Nothing below asks you to trust the next paragraph. It asks you to check it against a number you can reproduce, yourself, in about ninety seconds.
Here is the claim, in the form you're allowed to swing at: any AI output you cannot trace to a specific physical execution was hired for the exact reason Warren Harding was — it looked the part. In 1920, a syndicate of Republican operatives went looking for a candidate and found a senator with a magnificent voice, a jaw you could set a watch by, and almost no legislative record worth discussing. They ran him anyway, because — as one of them reportedly put it — he "looked like a President." He won in a landslide. His administration is now the textbook case, taught in Malcolm Gladwell's Blink as the Warren Harding Error, of what happens when a hiring decision optimizes for the presentation layer and never checks the substrate underneath it.
That is not a 1920s problem you get to file away. It is the exact failure mode sitting in your own stack the moment a chatbot produces a fluent, confident, structurally beautiful answer with nothing underneath it but more fluent text. If Warren Harding had shipped an API, the changelog would read: "improved response confidence; no change to underlying competence." You're about to watch that gap up close — through a Wall Street trading floor, a Groupon stock purchase, and a room full of my ex-quant colleagues laughing at me in a way that, in retrospect, they'd fully earned the right to.
You've met a Warren Harding. Everyone has. The candidate, the vendor, the hire, the pitch deck that felt so structurally sound you skipped the diligence — the moment a presentation was so complete that checking the substance underneath felt almost rude. That feeling is not a character flaw. It's the oldest cognitive shortcut there is, and it has a name older than any of us: Alfred Korzybski's axiom, "the map is not the territory." A representation of a thing is not the thing. A résumé is not the work. A confidence score is not competence.
The 20th century's most expensive demonstration of this was Robert McNamara at the Department of Defense, running the Vietnam War off body counts and kill ratios — a map so elegant it had its own name, "systems analysis" — while the actual territory (the war it was supposedly describing) did something the spreadsheet had no column for. The 2008 financial crisis was the same error wearing a nicer suit: PhDs with genuinely dazzling models pricing mortgage-backed securities off correlation assumptions that were, structurally, a map of a territory that no longer existed. Every one of these disasters has the same shape: an institution mistook a representation for the thing represented, and nobody in the room had a job whose entire function was to keep asking "compared to what, exactly, in reality?"
Here's the part that should actually worry you if you deploy AI for a living: an LLM is not a flawed version of this pattern. It is the pattern, industrialized, and for the first time actively generating the territory it's supposed to only be describing — a fluent paragraph confident enough that the operator reading it stops checking whether anything downstream matches it. It is, definitionally, a map-generation engine with no territory attached — extraordinarily good at producing the representation of a correct answer, structurally incapable of knowing whether that representation touches anything real, because nothing forces it to. You are not worried about AI lying to you. You should be worried that "lying" implies it knows the difference.
I have to tell on myself here, because it's the exact contribution this section owes you: I worked on Wall Street early in my career, and my colleagues nicknamed me the quant — not as a compliment. I fell, hard, for the Groupon hype. I had the narrative bandwidth to feel the momentum of a story and none of the disciplined grid yet to check whether the story was standing on anything. They were merciless about it, and — this is the part that took me a decade to admit — they were right to be. A trading floor runs on a brutal, correct instinct: a beautiful thesis with no P&L behind it is worth exactly nothing, and the floor will tell you so, loudly, in front of everyone.
But notice the shape of the insult, because it's the shape of the whole thesis: quants get laughed at not because they're wrong about the model, but because the model is the only thing they checked. A trading floor's contempt for "the quant" and Malcolm Gladwell's contempt for "the Warren Harding hire" are the same joke, told about two different failure directions. The quant has all grid and no grip — a beautiful, airless model of a market that keeps declining to cooperate. Harding has all grip and no grid — a magnificent presence with nothing structural underneath it to hold weight. Neither one, alone, survives contact with the territory. That's not a coincidence. That's the whole design constraint — and it's exactly why the engineering team and the risk team at your own company, kept in separate reporting lines, keep producing separate disasters instead of one shared correction.
Here's the thing I've avoided saying in public, because it sounds more isolating than useful, and I'm going to say it anyway because avoiding it stopped being the safer move: the reason I keep building a 144-cell lattice that walks breadth-first across a connectivity grid instead of drilling one clean line to a single answer is that it's a physical model of how I actually think. I don't process problems as a sharp, single-point drill. I triangulate — I hold a dozen slippery, half-formed, wildly distant concepts in the air at once and slowly, disciplinedly, grid them into place until they tie into a bag that holds. For a long time I called that a liability. People experienced it as too much — too sensitive, too all-over-the-place, not focused enough to be taken seriously in a room that rewards the sharp single line.
But raw intellect — the sharp knife — is an excellent tool for one job: cutting a clean line through a problem you've already correctly identified. It is a catastrophic tool for the job that actually matters most, which is figuring out whether you're cutting into the right problem at all. That's not a knife's job. That's a net's job — a disciplined grid wide enough to catch the thing before you've committed the whole blade to it, which is exactly why the boogeyman is invisible to whoever's holding the knife: they're looking at the cut, not the room. Every master of this kind of triangulation was, on the way there, a disaster: too sensitive, too scattered, too slow to commit to a single confident line. The isolation wasn't a sign I was doing it wrong. It was the friction of doing something a sharp-knife room has no slot to reward yet — and, it turns out, the exact thing that room can't hire its way into.
Let's be honest about the limit, since the whole pitch collapses the moment we oversell it. I am not claiming to be the smartest person in a given room, and this framework does not claim to know whether any given answer is correct. Rice's theorem — a real, load-bearing, decades-old result in computability theory — says that almost no non-trivial property of a program's behavior can be decided in general, for every possible input, in advance. Applied here: no system, mine included, can tell you with certainty that a given AI output is "good." Anyone selling you that number has quietly sold you a solved halting problem, and you should ask them how.
What breadth-first triangulation and a physical lattice buy you is narrower, and I'd rather undersell it than have you discover the ceiling later: not "is this correct," but "is this on-domain" — did the work land inside the territory it was authorized to touch, or did it wander off into a territory nobody signed off on. That's a real, measurable, falsifiable question with a denominator — the same fence explored at more length in Undecidability Is the Asset. A high probability is a guess with good manners; it is not the same claim as a physical receipt, and only one of those two survives being asked to testify. Whether the work is good, once it's confirmed to be in-domain, is still — and will likely always be — a human judgment, and possibly a "sharp knife" judgment at that. I'm not replacing the knife. I'm building the net that catches it before it cuts into the wrong territory entirely.
Here's the reframe that took me the longest to earn, and it's the one thing in this post I'm actually certain of: empathy is not a moral posture. It is a high-bandwidth sensor for the gap between the map and the territory. If you cannot feel friction, you cannot detect drift — you're numb to the exact signal that tells you a representation has quietly stopped matching reality. People who called me "too sensitive" were describing a real, high-bandwidth receiver running without a disciplined enough grid to filter the signal yet. That's not an indictment of the sensor. It's a diagnosis of an unfinished filter — and the filter, not the sensor, was the thing worth building.
This is also, deliberately, business ethics rather than virtue ethics, and the distinction matters more than it sounds. Virtue ethics asks whether an action is good. Business ethics — the kind that actually gets you funded, insured, and taken seriously in a room full of people whose job is risk — asks whether a claim is structurally sound: can this be verified, can it be priced, can it be defended when someone official asks where the number came from? "Understanding," in a challenger sales conversation, sounds like empathy, and empathy sounds like a soft skill you bring to a room that wants a hard number. Stripped of the sentiment, what "understanding" actually means here is structural grip — full-deck triangulation, not warmth. You need the whole deck, sensory input included, to play the hand at all. You just can't get paid for holding it if you can't also produce the receipt. That's the whole reason zero-latency is replacing zero-trust as the real security posture: zero-trust assumes the breach and adds friction everywhere; zero-latency just makes intent and execution the same physical event, so there's no gap left for the breach to hide in.
Here's the part of this pitch I used to soften, and I've stopped, because softening it was doing you a disservice: you cannot build this internally, and the honest move is to stop pretending otherwise. Not because your engineers aren't good. Because of three things your organization is actually built to do, and none of them is this.
One, the hiring trap. Your hiring pipeline is built to find sharp knives — deep, narrow, credentialed optimizers who are extraordinarily good at making an existing model more elegant, and structurally unequipped to ask whether the model should exist in this shape at all. Hand a sharp knife a hallucination problem and you will get, with great confidence, a better-dressed hallucination. Two, the proxy comfort zone. Your culture is fed by the metrics that make friction disappear — confidence scores, quarterly OKRs, a dashboard that's been green for a year. Grounding semantics in physics is, by definition, the process of making those metrics stop lying to you, and the first thing an honest internal audit finds is how much of that comfort was borrowed against reality nobody checked. Nobody survives, politically, being the person who says so out loud in the room that funded the dashboard. Three, the disaster deficit. Building a net wide enough to catch this means having been the disaster it's built to catch — watching an ungrounded story eat real capital, in public, on your own account, and rebuilding the filter afterward instead of just building a thicker wall. Most internal teams have excellent insulation and no line item for that specific experience, and there is no procurement process that buys scar tissue.
None of that is an insult, and — this is the part worth sitting with — pointing it out isn't cruelty either. Shattering an executive's proxy metrics is not mean. It is the only form of structural empathy that actually helps them, because the alternative is watching them spend two years and a real budget building a more confident version of the exact hallucination they already have. Every executive already suspects, quietly, at whatever hour the demo looked a little too good, that they might be the next Warren Harding. That fear is correct, and the only thing that resolves it is a receipt, not a reassurance. So here's the pitch, uncompressed to four words, because dressing it up further would be its own small act of dishonesty: shut up and take my money.
Run this audit on your own organization, and try not to wince: how many of your key metrics are actually proxies for the territory rather than the territory itself? A "confidence score" is a Harding hairline. A "quarterly OKR completion rate" can be a Groupon thesis with a nicer font. A dashboard full of green lights that nobody has ever traced back to an actual physical event is, structurally, a beauty pageant with a spreadsheet instead of a runway — and the judges, like the 1920 Republican operatives, are voting on presentation because nobody built them a way to check the substance underneath it in the time they have.
This is the entire joke behind "are you out of your pixel?" — the house question underneath everything I build. Your pixel is the one coordinate on the map where your actual competence and your actual authority overlap; being out of it doesn't feel like failure, it feels like a really good pitch deck, right up until the territory disagrees. The good news, and it is genuinely good news: it's okay. Yesterday, everyone was. The honest fence between what we can and can't tell you about your own pixel — the fun kind of "your data, your ledger, go look" — lives at /pixel. And if you'd rather stop laughing and start acting on it: the license that puts a signed, recomputable coordinate under your own agents' output, instead of a Warren Harding hairline, is one page over at /iamfim — $20 per agent, and the number itself never moves. What moves is the year it buys: lock in today and freeze a full year at that price; wait, and the same $20 buys less time every month you sit on it. A flat sticker price that quietly buys less over time is, once you notice it, its own small Warren Harding — looks stable, isn't.
None of what follows is offered as proof of anything above. It's the raw material this argument was built from — check it yourself, and if it points somewhere different than we did, that's the fact that changes this post. Malcolm Gladwell, Blink (2005) — the chapter "The Warren Harding Error: Why We Fall for Tall, Dark, and Handsome Men," the source of the name borrowed throughout. Alfred Korzybski, Science and Sanity (1933) — origin of "the map is not the territory." The McNamara/Vietnam systems-analysis failure, documented at length in Errol Morris's The Fog of War (2003) and in the broader historiography of the war. The 2008 financial crisis's model-vs-market gap, covered in Michael Lewis's The Big Short and the Financial Crisis Inquiry Commission's own 2011 report. Rice's theorem (1953) and the Turing/Church results (1936) it generalizes, for the formal boundary on what any checker — human or machine — can decide about a program's behavior in advance. W. Ross Ashby's Law of Requisite Variety (1956): a controller needs at least as many degrees of freedom as the system it controls, or it cannot regulate it — the formal reason a sharp knife, alone, cannot supervise a system wider than itself. Normalized compression distance — the model-free way of measuring how much two things share by how well they compress together — is a published, peer-reviewed comparator, not something we invented (Cilibrasi & Vitányi, Clustering by Compression); it's the mechanism underneath the recomputable coordinate the demo above returns.
The book carries the same claim in its own voice, worth reading in full rather than trusting our paraphrase: "The critical distinction: the grid doesn't represent meaning — it IS meaning… Position = Meaning. The map IS the territory." — Tesseract Physics, Chapter 0: The Razor's Edge. The same chapter is where the carrying cost of skipping this gets a number: $8.5 trillion in annual software waste is, in the book's own words, "the carrying cost of that confusion" — Chapter 0, again.
And the short-form piece that actually seeded this post's framing, embedded rather than just linked, because it earns the two minutes:
The mechanism underneath the lattice itself — the recomputable coordinate, not the history lesson — is Patent US 19/637,714, 36 claims, filed April 2, 2026.
Three things, in the order they actually cost you something. One: run npx thetacog-mcp attest-demo on your own machine and see the coordinate come back — that's the whole proof, and it's free. Two: go look at your own dashboards, your own vendor's confidence scores, your last three "this looks great, ship it" decisions, and ask, out loud if you have to, whether you're checking the map or the territory — the /pixel ledger is there if you want the framework to run the audit against instead of your gut. Three: when you're ready to put a signed coordinate under your own agents instead of a Warren Harding hairline, the license is at /iamfim — $20 per agent, and that number doesn't move. The year it buys does: lock in now and it's a full year at that price; wait, and $20 buys you less and less of one. Read that twice — a flat number that quietly covers less over time is a rising price wearing a Harding hairline of its own.
Or, compressed to the only sentence that was ever actually necessary: shut up and take my money. Everything above it was just the diligence you were owed before you were allowed to say that with a straight face.
So — are you out of your pixel? It's okay. Yesterday, everyone was. And "Fire Together, Ground Together" was never a metaphor for the book jacket — it's the literal thermodynamic requirement underneath everything above: if it doesn't ground, it isn't real, and no amount of confidence in the delivery changes that.
Cited, not claimed: Gladwell, Blink (2005) · Korzybski, Science and Sanity (1933) · Rice (1953) · Turing & Church (1936) · Ashby, An Introduction to Cybernetics (1956) · Cilibrasi & Vitányi, Clustering by Compression (2005) · Celastrina Calea, "Knowing Is Not Intelligence Ep. 2." The floor is filed: Patent US 19/637,714. Take this as an argument, not a verdict — check the ingredients yourself.