Published on: June 26, 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.
We believe you can't trust your AI for one reason: you've been grading it with a score — is it good, is it accurate, is it aligned — and a score is the one thing about an agent you can never pin down. To grade "is it good," you need a judge as capable as the work, and its verdict is a sample no stranger can recompute. So you run more evals, and the anxiety never resolves, because you're measuring the wrong kind of thing. Competence was never a score. It's a shape — a region on a map showing where the agent's work landed versus where you hired it to land. The question that has an answer isn't "is it good?" It's are you out of your pixel? — did the agent stay in the lane it was hired for, or did it wander into one you never authorized?
A score is a number you argue about. A shape is a region you point at. The agent that does flawless work in the wrong domain — the surgeon doing plumbing — gets a perfect score and is a catastrophe. Only the shape shows it.
Here is the map, and once you see it you can't unsee it. It is a fixed 144-node lattice — both edges are the same set of anchors (what you do, by what you act on), each labeled by a coordinate like A1,A2. Every place an agent's work says something (its declared intent) and does something (its actual execution) lights a cell. Colour tells you whether saying and doing landed in the same lane. Green is in-lane: you worked where you declared — this is the competence. Amber is bleed: the work touched a neighbour you didn't quite mention, which is normal because nobody declares everything. Red is drift: the work fired in an orthogonal lane it never declared — the rupture. Your competence is not a grade on this map; it's the region the green covers, and the catastrophe is wherever the red appears.
What good looks like is a shape: green concentrated where you declared, red near zero. Not a high number — a clean region. You read a glance of the map, not a leaderboard.
Why does a score miss the catastrophe a shape catches? Because a score quietly assumes risk moves in straight lines — a little off here plus a little off there sums to a small, survivable total. AI doesn't fail that way. It fails on the diagonal — the king move. Picture an agent a notch outside its domain and a notch past its role at the same time: each step alone is minor, the kind your controls wave through, but together they are one jump across the fence into a lane no single check was watching. Measured by the metric this map uses — Chebyshev distance, the king's move, where a diagonal shift weighs exactly as much as a lateral one — that move is a single large leap, not two small steps. A model that sums or averages (the straight-line math under most evals) literally cannot see it. The shape can, because it keeps the honest square boundary where the diagonal counts.
Price drift with king moves or you are blind to the exact corner where the catastrophe sits: the simultaneous small-domain, small-role shift that a summing score reads as nothing and a shape reads as a rupture.
This is not a diagram on a slide — it runs on every commit today. The drift receipt now shows you the shape: each green, amber, and red cluster is ringed in its own colour, numbered, and named by its coordinate and what that coordinate means — a red region reads not as an opaque "C3,B2" but as the lane it wandered into. The walk that places the work is the real recursive walk on a chip, about 14 milliseconds, no model in the loop. And the proof it emits is an ed25519-signed receipt of coordinates and one-way hashes — never your work product. It is an oracle, not a host: anyone recomputes the shape offline and gets the same regions, and changing one field breaks the signature. You don't take our word that the agent stayed in its lane; you re-walk it. On blind cross-domain tests the shape separates on-domain from off-domain at 0.90, rejecting off-domain work 10 times out of 10.
The shape is measurable, named, private by construction, and recomputable on your own machine. That is the difference between a screenshot you have to believe and a receipt you can check.
A shape you can see is already more than a score, but it's getting more useful fast — and we'd rather show you the road than oversell the destination. Two things are landing. First, the shape now narrates itself: a small local model reads the drifted regions plus your actual commit and tells the story in plain terms — "you asked for this lane, but the work pulled into that one; re-read this clause to fix it" — so the receipt stops being a picture and becomes a place to look. Second, the shape moves: rendered over a run of work, the regions become a film of competence over time, and the drift becomes a signal you can watch accumulate. Run the geometric walk and the narrative side by side and they cross-check each other — agreement is confidence, disagreement is exactly where to inspect. That is a system that can catch itself drifting, which is the thing every safety story gestures at and few can run.
Three honest questions, because a claim this large should answer them before you act. Why is there no alternative? Because the alternative is the score, and the score is grading a semantic property — which Rice's theorem (1953) says no program can decide in general. You cannot escape an undecidable question by running more evals; you escape it by changing the question to a decidable one: where, not whether. Why is this not too good to be true? Because it is bounded and says so — it reads domain placement, not quality; it's measured (0.90 separation, 10-of-10 off-domain) and names its open inch (a reworded sentence is read coarsely, on purpose, because surface paraphrase is the model's job, not the fence's). Why has nobody else done it? Because the field has been looking inside the box to certify behaviour — the undecidable problem — instead of placing a rigid geometry around the output, where the question becomes decidable and the answer becomes a shape. The full thesis lives in The Decidable Slice of Alignment, and what it means for the people holding the bag is in The Exclusion Is the Liability.
You can stop grading a number you'll never pin down. Find your pixel — the coordinate where your work is supposed to land — and watch the shape to see whether it stayed there.
See the shape, recompute the receipt: thetadriven.com/pixel. Competence was never a score you argue about. It's a shape you can point at — and now, one you can check.
![tolerance panel for commit 1910ba5 — chore(blog): attach commit tolerance panel as OG image [panel-attached]](/blog/post-panels/2026-06-26-competence-is-a-shape-not-a-score/1910ba5de48b98a26ec2e2a50f2a5ae4c56dbe46.png)
