If you think your job is safe, think again.

01The trick that keeps working

The fastest gains in AI over the last two years came from one loop: let the model try, let a program check the answer, reward what passes. Researchers call it reinforcement learning with verifiable rewards, or RLVR. It's how models went from fumbling competition maths to solving it, and from autocompleting code to fixing real bugs against real test suites.

The loop has one precondition: a checker. Unit tests for code. An exact answer for a maths problem. Where a cheap, reliable checker exists, models climb fast. Where it doesn't, they plateau on vibes and human ratings.

So the useful question isn't "can AI do my job?" It's narrower and more uncomfortable: can a program tell whether my work is right?

02Checkers are everywhere

Most skilled work ends in something a program can check. A floor plan either meets egress code or it doesn't, and a shortest-path computation over the plan says which. A card dispute is either valid under the network's reason code or it isn't. A flight plan carries enough fuel under the reserve rules, or it doesn't. A Nutrition Facts panel is rounded the way 21 CFR 101.9 says, to the gram. A construction pay application ties out to the contract ledger, to the cent.

None of these need a judge model. The rules are already written down. Often the simulator or validator already exists, and it's open source.

03So I went looking

In November 2026 I ran a team of research agents over what already exists: 54 RL-environment ecosystems, 828 named tasks from the big gyms, the 2025–26 RLVR papers, and 213 open-source verifiers and simulators, flagging the ones nobody had ever used as an LLM reward. Then, field by field, I asked two questions: what work ends in an artifact a program can check, and has anyone built the environment?

The answer is the atlas below. It holds 380 candidate environments across 27 fields. Each has the task, the program that grades it, the knobs that make it harder, and the closest prior work I could find. Every entry also went through an adversarial novelty check, an agent whose only job was to find the paper that makes it not new.

Another 194 are "thin": a benchmark exists, or classic non-LLM RL, but nobody has built a reusable environment you could train a language model in. Only 17 had already been built, and they're listed at the bottom for completeness.

04Where the gaps are

The gaps aren't in exotic research corners. They're in the paperwork of the economy: industry operations, payments and insurance, regulatory filings, accounting, engineering code checks. Click a field to open it in the atlas.

05So is your job safe?

If your work ends in an artifact a program can verify, like a passing check, a balanced ledger or a valid filing, then the thing standing between you and a model that does it is a training environment. For 178 kinds of work, nobody has built one yet. Each is weeks of engineering, not a research breakthrough. The rules, the simulators and the validators are already sitting there.

If your work doesn't end in a checkable artifact, because it's judgment, taste, negotiation or trust, this method doesn't reach you yet. That's the honest version of "safe". It isn't about your title or your seniority. It's about whether your output can be graded by code.

The gaps cut both ways. Whoever writes the checker for a field gets to define what "correct" means there, and gets the first models trained against it. If you're the person in your field who knows where the rules live, that's a strong position to be in.