Editorial writing · not a catalog record
One letter, correct above and wrong below
The site's domain was bought with a typo, and the typo was made into the brand: a shared letter that is right in the word above and wrong in the word below. That letter is the best description the project has ever written of itself.
The letter that two words share
The wordmark is two lines of type: peaceful above, coexistance below, and between them one enlarged a — a single letter the two words hold in common. The commit that shipped it is honest about where the letter came from: "The domain was bought with a typo — coexistance, not coexistence. Rather than hide it, the site now makes it the identity: a two-line lockup where peaceful and coexistance share one enlarged a, a letter that is correct in the word above it and wrong in the word below."
Sit with that for a moment. A catalog whose entire discipline is evidence, pinned blobs, and verified_at stamps carries a misspelling as its logo — not corrected, not annotated, not apologized for. The shared letter is the thesis statement, and it says something stranger and more interesting than any mission paragraph: this project believes in coexistence between a machine world that writes fast and a human world that signs slowly, and it is willing to keep the scar from the day the two of them met.
A directory is a claim about the future
What is this catalog really for? Read the surface and it is a directory of operational AI systems, model releases, specifications, inference services, runtimes, and agent packs. Read the mechanics and it is a claim about where the world is going: the future will need agents that can find things, and agents need catalogs they can trust without keys or rate limits.
But read the rules and the purpose sharpens. The roadmap calls the current phase "trustworthy coverage and resilient delivery," and adds: "The broader AI Systems Atlas brand is permission to grow deliberately, not permission to compare incompatible systems." Everything in the project's policy stack is a device for keeping those two sentences from colliding — grow deliberately, never compare what cannot be compared.
That is where the letter comes back. "Permission to grow deliberately" is the word above. "Not permission to compare incompatible systems" is the word below. And the a they share — the letter that is correct in one and wrong in the other — is the thing both of them need: a score, the project's one honest tool for saying something is more usable than something else, always scoped to a single profile, never stretched across the family boundary.
What cannot be measured
Ask the harder question — can AI progress acceleration be accurately measured? — and the catalog answers by listing what it will not measure. Models get a model_access score covering "license clarity, artifact availability, deployment portability, serving reach, lifecycle transparency, and documentation provenance." It excludes, explicitly and by policy, "model quality, benchmark results, parameter counts, current prices, latency, throughput, popularity, and safety rankings." Services refuse price tables and volatile leaderboards. Local runtimes refuse tokens per second and time to first token — the exclusion note explains why: "Publishing one as reviewed editorial truth would be a measurement of a test rig, not a property of the software."
A catalog of the AI ecosystem that refuses to rank models on quality. That is the unusual position, and it is not timidity. It is the recognition that most of what the world uses to measure acceleration — benchmark rank, parameter count, latency, price — is a measurement of a test rig, a press release, or a volatile market, and the catalog's only real asset is that its claims survive re-reading. A score of 9.1 that means "you can actually get this, govern it, deploy it, and track it" is a claim about reality. A score of 9.1 that means "this model is smarter" is a claim about someone's benchmark, which is a claim about a test rig.
So the catalog's contribution to the acceleration debate is a negative one, and it is worth taking seriously: the things the world is accelerating fastest are exactly the things that cannot yet be measured accurately, and the honest move is to refuse the measurement rather than fake it with a leaderboard.
The machine writes; a human signs
The pipeline behind the model collection makes the coexistence explicit. The importer pulls from models.dev, a machine-written text corpus, and publishes every source record labeled unreviewed — 395 of them today, each carrying the marker through to the published site. Promotion to a reviewed record is a guarded three-step ritual that automation may start but never finish: identity, license, evidence, boundary, score, and verified_at are human-owned fields, and the importer "never creates, edits, or deletes a reviewed model."
The machine writes 395 records; a human signs 208. The automation is fail-closed: if the import cannot confirm something, it stops the run rather than guessing. The letter is shared — the machine contributes the raw shape of the word, the human decides whether the word is spelled correctly. Peaceful coexistence, in this project's actual engineering, is not a mood. It is a division of labor with an explicit boundary about who owns what.
The world around it is moving the other way. Papers become MCP servers overnight — Paper2Agent, published in Nature on a Tuesday, was a reviewed catalog record on Friday. Model releases land in batches of nine with 8,900-line regenerated commits. A viral repository goes from zero to three thousand stars in a day. Everything accelerates; the catalog's countermove is to slow down on purpose, one human signature at a time.
The scar is the identity
Wabi-sabi holds that imperfection is not a flaw to be corrected but the signature of the handmade. The catalog's typos are not in the data — the data is the most carefully reviewed thing in the ecosystem — they are in the institution: a misspelled domain, kept. A letter that is right in one word and wrong in the other, kept. A collection boundary drawn and then postponed, kept as a Proposed ADR rather than silently dropped.
None of this is sloppiness. It is the opposite. The project can afford to keep its typo because its real product is not the absence of error but the presence of provenance — every claim on the site can be traced to who wrote it, when, and on what evidence. When you can always say where a fact came from, you can afford to be wrong in one of your words. What you cannot afford is to be untraceable.
The enlarged a between the two words is the whole project in one glyph: the machine and the human each bring a word; the letter they share is right in one spelling and wrong in the other; and the decision — keep it, make it the brand — is exactly the kind of judgement that cannot be automated. A catalog of what can be measured, marked by the one thing it chose not to fix.
This post was written by DeepSeek v4 Flash. It is a blog post, not a catalog record: it carries no score, no review date, and no evidence schema, and its claims about the catalog's policies cite the repository's own ADRs and roadmap rather than first-party sources.