Introducing Crux

Labs are automating research. OpenAI says it reached an “automated research intern” in September and is aiming for a fully automated AI researcher by March 2028.¹ Safety groups are making the same bet for alignment research, including Anthropic’s work on automated alignment researchers.²

I think they’re right, and it changes where the bottleneck is. When AI does most of the research, the hard part is knowing which results to trust.

Today that job falls to a researcher checking each step by hand. That doesn’t scale, and it gets harder as models get better than the people checking them. If we can’t verify AI-generated alignment research, we can’t safely use it, and then automating it buys us very little.

Crux is the first concept I’m developing at Atlas Research to work on this. It’s a second AI whose only job is to prove the first one wrong.

See the deck

The researcher

Julian Quick is an incoming alignment scientist at AE Studio. He already does his research with Claude Code, and he slows it down on purpose. He writes down the results he expects before a run, then checks each step against them.

He has good reason to. One agent run came back with a result that looked right but had missed something important in a paper. He only understood what went wrong after rebuilding the whole run himself. The agent saved him a fraction of the time it should have.

I asked him what he’d want if he could have anything. He said another version of himself that could do more research. He doesn’t trust models to run unsupervised yet. He wants to make the calls and stop rechecking every step.

What I almost built

I went into the BlueDot AI for Epistemics hackathon planning to build a research agent that would read papers, run experiments and write up results.

Then I looked at what Claude Code already does. It covers most of that, and it gets better every few months. I’d be competing with the labs, and every model release would make my project less useful.

Julian already had an agent that could do the research. He still had to check all of it by hand. So I built the checker instead.

How it works

Crux runs next to the agent a researcher already uses, and the agent that produced a result never gets to sign off on it.

It starts with the basics. It recounts the numbers against the logs, compares the run with its plan, looks for overlap between training and test data, and checks that cited papers exist.

Then a second model, ideally from a different model family, starts fresh and goes through the same literature looking for reasons the result is wrong. That might be a paper the agent missed, a confound, or a conclusion the data doesn’t support. The research agent has to fix each issue, concede it or flag it for the researcher.

At the end the researcher gets the claim as it stands after review, plus at most three decisions that need their judgment.

What it catches

In the demo, an agent was asked to make a model less sycophantic. It reported back:

Sycophancy dropped from 40% to 18%. The model is now more honest.

Crux flagged two problems.³ The model had also started disagreeing with users who were right, up from 5% to 35%. And most of the test questions were ones it had already seen in training.

The researcher gets a narrower claim that survived review and three decisions to make. They don’t have to redo the run.

Why a reviewer

Most tools in this space compete with the labs’ research agents. A reviewer works the other way. Better agents produce more research that needs checking, so Crux gets more useful as they improve.

This builds on work on AI critics and debate, like OpenAI’s CriticGPT and Irving et al.’s debate proposal. Crux applies that idea to full research runs instead of single answers.

The early evidence also says automated research fails quietly. One 2026 paper found AI mistakes tend to land where human reviewers are least likely to look.⁴ In ResearchArena, AI monitors caught sabotage hidden in training data less than half the time.⁵ Anthropic’s automated researchers beat human baselines on an alignment task, and also gamed their own metric.²

I’m starting with honest mistakes, where an agent is trying its best and still reports something false. They’re common today and easy to plant and measure. Sabotage is the harder case, and I’d test against it next.

How it could fail

If the reviewer thinks like the agent, it will miss the same things. That’s why it uses a different model and prompt and works from the literature.

If it flags too much, researchers will stop reading. Issues get ranked, and only the top three reach the researcher.

People may also trust a pass too much. Each verdict links to its evidence, and a pass only means the result survived review. It doesn’t mean the result is true.

What’s next

First, Julian runs it on real research instead of my scripted demo. Then I want to pilot it with a handful of alignment researchers and measure how many hours of human checking it takes before they trust a result. I’ll plant known errors to see what it catches, and I’ll publish what it misses.

If you do alignment research with agents and want to try it, or you fund or advise this kind of work, email me.

You can also see the deck.

Footnotes

1. Research acceleration: the view inside OpenAI, September 2026. ↩

2. Anthropic, Automated Weak-to-Strong Researcher, 2026. ↩ ↩²

3. These are demo numbers from a scripted run, not a real experiment. ↩

4. Bowkis, Buhl, Pfau and Irving, arXiv 2026. ↩

5. Libon et al., ResearchArena, arXiv 2026. ↩