EVOAI Published structures, recomputed in your browser, recorded where nobody can edit them

Evo 2 was trained on a nucleotide corpus and generates DNA one base at a time. The Nature paper reports training on 9 trillion DNA base pairs from all domains of life, a context window of 1 million tokens and single nucleotide resolution, with the weights, the code and the OpenGenome2 training set released publicly. Those two figures are the paper's own, and this site does not restate them as its own measurement.

Evo 1, published in Science in 2024, established the approach at the scale of molecules and small genomes. Evo 2 carried it to eukaryotes and to whole genome context. Neither model is run by this site. Nothing on this page is model output.

The work reports the generative design of complete bacteriophage genomes, written by Evo 1 and Evo 2, both fine tuned on Microviridae genomes, using PhiX174 as the template, 16 viable phages, and a cocktail of them that overcomes PhiX174 resistant Escherichia coli. It also reports that cryo electron microscopy confirmed one generated phage using an evolutionarily distant DNA packaging protein. The Stanford Report story on the work says nearly 300 phages were synthesised and tested, of which 16 were exceptional.

These are bacteriophages. They infect bacteria. They do not infect humans, animals or plants. The study reports them against laboratory strains of E. coli, and none of the synthesised genomes infected E. coli K-12.

Two structures came out of that study and both are in this repository as baked integer payloads: the written phage as 36CR, solved at 2.9 A, and its natural template as 36CQ, solved at 2.76 A. Both are CC0, both are public, and every number in the comparison below is computed from those coordinates. If a number appears anywhere on this site and the engine in front of you did not compute it, it comes from something on the references page.

The engine is EVOENGINE v1. On the path whose output is hashed there is no floating point arithmetic and no transcendental call, because the ECMAScript specification only requires those to be approximated. Lengths come out as integers in units of a millionth of an angstrom, ratios in parts per million. The result is a canonical text block and its keccak256 is the hash that goes on chain. Two runs of the same question are byte identical on every device.

This site never runs Evo. It never simulates physics. The words for what it does are recomputed, recorded, replicated and diverged.

Writing a run to the shared log will require 1000 $EVOAI staked, with a three day cooldown, which buys six runs per address per day. Two thousand staked gives nine and the ceiling is twenty four at seven thousand. The curve is deliberately flat, because this is a rate limit and not a yield ladder.

A stake rather than a balance check, because a balance at call time can be borrowed: one thousand tokens can gate a thousand wallets inside a single block. A stake with a cooldown cannot.

The engine is never gated. Every wallet state, including no wallet at all, can press Run and see every number and the whole 3D view. The chain is only ever asked to record a claim.

A single question can ever mint at most 300 credits, so the supply is bounded by the number of distinct valid questions rather than by the number of wallets. Credits buy a tier 2 run at 50, a tier 3 run at 150, a note on your own revealed run at 200, and a label on a specimen at 1000.

They are not money. There is no transfer, no approval, no mint by the owner, no redemption, no fee discount, no revenue and no claim on anything. They only leave an address by being spent inside the contract.