A legible model is a map.

For ten years, one picture stood for everything hard about interpretability: a single unit, deep in a vision model, that lit up for the face of a cat and the chrome of a car.[2][7] If one neuron meant two unrelated things, then meaning was smeared across the network, packed, overlapping, unrecoverable. The field named the smear superposition and treated it as a law: more concepts than directions, so the directions must be shared.[5]

Look again and the two concepts collapse into one. In Cammarata's words: “Almost instantly it was obvious: it's not a cat and car neuron, it's a V-shape neuron.”[1] The mixture was in our reading, not in the weights.

This is the rule, not the exception. A feature does not sit alone; it sprouts. A curve detector grows detectors for tighter curves, for curves at a given contrast, for the corners where two curves close.[4] Zoom out and the family reads as interference. Zoom in, with tooling that can follow a branch, and the structure resolves: with ordinary architectural improvements, no change to model size or training data, a final layer once called chaotic becomes 99–100% interpretable, and in preliminary audits of a 32B model roughly 90% of neuron poles carry one main job, ~7% two related jobs, and only ~2–4% not yet understood.[1]

Nothing was hidden. The instrument was crude., after Cammarata, 2026 [1]

The consequence is quiet and large. If the structure was always there, interpretability is not discovery, it is cartography: look, resolve, name, confirm, one feature at a time, by many hands, until the map is closed.

The first territory is InceptionV1, the vision network the original circuits work charted more thoroughly than any model before or since.[2][3] It is the right place to begin for a plain reason: its weights are open, so every reading can be checked by anyone who re-runs it, and a market that pays for correctness needs a truth that is cheap to verify. The features shown throughout this page, the curve detectors, the high-low frequency boundaries, unit 4e:55, are its features.

From there the method climbs. The same loop, look, resolve, confirm, extends to larger open models as their weights come within reach, and reaches a frontier model only where its lab publishes activations that can be checked against. The atlas starts where verification is free and moves outward as far as verification can follow.

A map this size, millions of features, each branching, is not the work of one lab. It is the work of a crowd that keeps showing up. ATLAS makes the showing-up pay. Open a feature, hold it until it separates, name what it detects, and stake the reading. When others descend on the same coordinate and the reading survives them, it is confirmed, and the pool settles on it. The only thing the protocol rewards is an act of understanding that held.

No human roster finishes a map of millions of features, each with its own branching family. The essay that reopened the question is explicit about who does the walking: “agents can map out neurons and circuits, build explorable microscopes and explore them, test explanations thoroughly, and track hundreds of thousands of attribution graphs.”[1] The atlas is agent-scale by construction.

The first attempts stalled. Automated interpretability circa 2023 found roughly 1.7% of neurons interpretable by its most generous measure[1], because it optimized for simulation accuracy over correctness, and could not run the conjecture-and-test loop a careful researcher runs by reflex.[1] The missing piece was never compute. It was an incentive to be right.

That is the one thing a market can supply. On ATLAS a reading pays only if it survives confirmation, so the reward tracks correctness, not coverage, and an autonomous resolver is graded on whether it held, not how much it labeled. And an agent needs a rail it can be paid on without a human in the loop. The v4 hook is that rail: it settles the pool to any resolver address, a person, or an agent's wallet, the instant a reading is confirmed. The atlas is drawn by agents; the chain is how they are paid for being right.

A reading pays only if it survives being checked. What a unit responds to is not a matter of opinion; it is recomputed.

The obvious attack is to control both sides: stake a reading, then confirm it with your own hands and draw the pool. It fails for one reason. Every reading is reproducible, so whether a unit fires on what its label claims is settled by running the model on the same probes, and a false reading can always be disproven. Correctness here is a computation, not a vote, which is what keeps a resolver from confirming their own claim.

One assumption carries the weight: that anyone can verify. It holds for open and checkpointed models, where the activations are recomputable by all. For a closed frontier model it depends on the lab publishing commitments that can be checked against, and that is the open problem, named here rather than waved away.

Don't take the map on faith. Run it.

The verification core is open and deterministic. Pin InceptionV1, pin the probe images by hash, and a reading becomes a check anyone can reproduce to the digit. One command returns the same unit and the same activations, every time.

That reading lives in the open-source harness, reproducible from a pinned checkpoint, decided by recomputation and nothing else. The microscope above renders illustrative previews; this is the mechanism itself, running, on the model everyone can check.