Reasoning 22 Sep 2026
Genesis OL 0.17 on ARC-AGI
OL stands for Operating Law, the one rule Genesis lives by: answer only from what the evidence supports, or say "I don't know."
Genesis OL 0.17 is our first public release. It applies that law to ARC-AGI puzzles, through a set of rules we wrote by hand. It answers only when every example agrees on one rule.
Key numbers
- answered
- 82
- wrong
- 3
- right when it answered
- 96.3%
- per puzzle
- 6.5 ms
ARC-AGI-1, public training half (400 puzzles) · pass@2 · self-reported1
How it answers
A rule is believed only if it reproduces every example pair. With fewer than two examples, or more than two rules that fit, it declines. At most two answers, ARC’s own pass@2.
Every refusal says why: no rule fit every example, or more than one did. "I don't know" counts as a right answer here.
Where the answers went
Most puzzles, it refused. The ones it answered, it almost always got right.
- Solved
- Wrong
- Refused
| Answered | Solved | Wrong | Right when it answered | Refused | Per puzzle | |
|---|---|---|---|---|---|---|
| ARC-AGI-1 | 82 | 79/400 (19.8%) | 3 | 96.3% | 318 | 6.5 ms |
| ARC-AGI-2 | 109 | 106/1000 (10.6%) | 3 | 97.2% | 891 | 7.8 ms |
The same 3 puzzles are wrong in both sets.2
How it got here
It works from a set of rules we wrote by hand, grown in one day from v0 to v17. Each new rule had to meet the same bar. Two didn't help, and we took them back.
As it learned to answer more, it got more careful, not less: right 93.8% of the time at the start, 96.3% now.
- ARC-AGI-1
- ARC-AGI-2
- Tried, then taken back
Show every version
| Version | ARC-AGI-1 | ARC-AGI-2 |
|---|---|---|
| v0 | 15 solved · 3.8% · right 93.8% · 1 wrong | 16 solved · 1.6% · right 94.1% · 1 wrong |
| v1 | 39 solved · 9.8% · right 92.9% · 3 wrong | 52 solved · 5.2% · right 92.9% · 4 wrong |
| v2 | 44 solved · 11% · right 91.7% · 4 wrong | 58 solved · 5.8% · right 92.1% · 5 wrong |
| v3 | 45 solved · 11.3% · right 91.8% · 4 wrong | 59 solved · 5.9% · right 92.2% · 5 wrong |
| v4 | 50 solved · 12.5% · right 92.6% · 4 wrong | 66 solved · 6.6% · right 93% · 5 wrong |
| v5 | 58 solved · 14.5% · right 93.5% · 4 wrong | 71 solved · 7.1% · right 93.4% · 5 wrong |
| v6 | 60 solved · 15% · right 93.8% · 4 wrong | 74 solved · 7.4% · right 93.7% · 5 wrong |
| v7 | 66 solved · 16.5% · right 94.3% · 4 wrong | 82 solved · 8.2% · right 94.3% · 5 wrong |
| v8 | 69 solved · 17.3% · right 94.5% · 4 wrong | 89 solved · 8.9% · right 95.7% · 4 wrong |
| v9 | 70 solved · 17.5% · right 94.6% · 4 wrong | 91 solved · 9.1% · right 95.8% · 4 wrong |
| v10 | 70 solved · 17.5% · right 94.6% · 4 wrong | 93 solved · 9.3% · right 95.9% · 4 wrong |
| v11 | 72 solved · 18% · right 96% · 3 wrong | 96 solved · 9.6% · right 97% · 3 wrong |
| v12 · taken back | 72 solved · 18% · right 96% · 3 wrong | 96 solved · 9.6% · right 97% · 3 wrong |
| v13 | 73 solved · 18.3% · right 96.1% · 3 wrong | 97 solved · 9.7% · right 97% · 3 wrong |
| v14 | 79 solved · 19.8% · right 96.3% · 3 wrong | 106 solved · 10.6% · right 96.4% · 4 wrong |
| v15 | 79 solved · 19.8% · right 96.3% · 3 wrong | 106 solved · 10.6% · right 97.2% · 3 wrong |
| v16 · taken back | 80 solved · 20% · right 96.4% · 3 wrong | 107 solved · 10.7% · right 96.4% · 4 wrong |
| v17 | 79 solved · 19.8% · right 96.3% · 3 wrong | 106 solved · 10.6% · right 97.2% · 3 wrong |
How fast
6.5 ms a puzzle on average, on one laptop CPU (Apple M3). Most take less.
- a typical puzzle
- 4.4 ms
- nine in ten puzzles under
- 13.6 ms
- almost all under
- 43.6 ms
- the slowest
- 99.1 ms
ARC-AGI-1, solving only, across five runs.3
On the ARC Prize board
It solves far less than the systems on ARC Prize's board. It also costs almost nothing to run.
Every other point is ARC Prize's own measurement, on their harder semi-private set, at the price paid to run it. Ours is the public training half, self-reported, and its cost is computer time at an assumed $0.36 per CPU-hour. Live from ARC Prize, their data from 2026-10-21 18:43 UTC. 291 systems.
What it means
The share it solves is small. It shows what a hand-written set of rules can reach.
How often it's right is the part we care about. When it answers, it's right about 24 times in 25, on puzzles it had never seen, with nothing tuned.
This approach is close to its limit. Going further needs a different engine. We chose Genesis’s own, and that work has started.
What's next
Now Genesis is learning to do this itself. Instead of rules we wrote, it uses its own way of learning: find a pattern, check it against every example, and answer only when all of them agree.
It's early. It solves fewer puzzles than Genesis OL 0.17 so far, and it hasn't got one wrong yet.
Check it yourself
Use their puzzles and their answers, not ours. The client marks everything on your machine, and we never see the result.
- Send it a puzzleThe live endpoint. Send your own and mark the answers yourself.
- The public repositoryThe protocol, the per-puzzle record, and a client that runs the whole set on your machine.
- The protocolThe rules it answers by, written down before the first run.
- The censusEvery figure on this page, and how it was measured.
- CorrectionsWhere any correction will go. Published results are never edited.
- The ARC-AGI puzzlesARC-AGI-1 and ARC-AGI-2, from their own repositories.ARC-AGI-2 repository
Notes
- Self-reported, not independently verified. These are the public training halves; the evaluation set is sealed and unread. The solver is private, so "no training" and the timing are our word. Every answer can be checked against the public ARC solutions.
- 391 of ARC-AGI-1's 400 puzzles are also in ARC-AGI-2, so the two sets aren't independent.
- Solving only, in-process, on one laptop CPU (Apple M3). The average is the middle of five runs. It isn't the endpoint's response time, which includes the network.
- The learning model didn't produce these numbers; Genesis OL 0.17 did. Build
6ab5dfa0…a95da283.