Merging models is the whole job.
We study three ways to turn several models into one. One of them is an API you can call. Two of them are still in the lab, and we say so plainly.
Three ways to merge.
They differ in where the merge happens: at inference, inside the network, or in the weights themselves.
API fusion
Four frontier models draft under different lenses. A merger settles disagreements by checking, and a different model audits. This is Failure Tempest 1.
Expert transplant
A frozen spine model gains expert layers lifted from other mixture-of-experts families, joined by learned adapters that start at zero.
Weight merging
Several fine-tunes of one base model combined into a single set of weights, with sign and sparsity conflicts resolved before averaging.
No single model is right every time.
Frontier models fail in different places. When they disagree, the disagreement is information. When they agree on a wrong answer, that is the hardest error to catch, and the one we hunt for.
Disagreement is a signal
Where drafts differ, the merger checks the claim instead of counting votes. A majority can be wrong; a check usually is not.
Agreement is not proof
Shared errors are real. The fusion step looks for them on purpose, and an independent audit reads the result.
Merging should not cost the base
In the weight tracks, a fresh merge must behave exactly like the model it started from. Anything it gains has to be earned in training.
How we publish.
Measured or blank
A number appears only after our own run. Until then the cell says "—".
Losses too
Where a merge does worse than its parts, that result is published beside the wins.
Status is dated
Every track says what exists today. The changelog records what shipped and when.
Controls first
Every experiment is designed against a control, so a tie tells us the idea did nothing.
Target #1 on Artificial Analysis.
That is the target for Failure Tempest 1, not a claim. Our method and our pending results live on the evaluation page.