99%
accurate when the graph can back the answer; it abstains otherwise.
99.03% of 35,951 blind derivations, on facts, from our knowledge graph.
A year of ARC3 learning in one self-improving requirements and code-writing harness.
You write the requirements as a model, and the harness breaks every decision in the code every way it knows, in one compile; each break must fail a test.
Every requirement, step, method, branch, break and test becomes a node in one graph, and that is the graph flying above, drawn with its names removed.
Claude writes its reasoning as a chain of steps, each citing what it stands on, and a citation that is not there fails the run.
A graph neural network proposes which requirement owns new code, as a question and never a verdict, and it has to beat plain rules.
Self-improvement is on by default: every pass records its own stats, checks them against its own rules and names what to improve next.
Each one says what it was measured on.
From our knowledge graph. Not measured on code.
99%
accurate when the graph can back the answer; it abstains otherwise.
99.03% of 35,951 blind derivations, on facts, from our knowledge graph.
616 of 616
top-tier facts right against outside truth (Wikidata), on a frozen population of 2,714.
One feature of a game, the one in the graph above.
96.5%
of 1,104 breaks of one feature caught by a test, leaving out those proven to change nothing (23), those with a written reason (20) and those with no verdict yet (52).
The 130 red points in the graph are those 95 and those 35.
294 of 317
statements held by a test, up from 31 of 82. Held means breaking it fails a test.
216×
less to read to learn who owns a method: 775 characters at the median against 167,665. This is the size of what is read; no agent was timed.
10
requirements nobody had written, found by walking the use cases, beside the designers' 17.
31 of 63
ARC3 lineage: its best agent reached 31 of 63 games where random reached 14.