Come look inside

You will try to break our explanations. Specifically: quantise, distil and fine-tune Noema checkpoints and report which named circuits survive. We expect at least one paper a year telling us we were wrong about something. Experience with mechanistic interpretability preferred; stubbornness required.

Noema-Probe is the tool the entire lab lives in, and it currently has the ergonomics of a research prototype because it is one. You would own the library and CLI: activation capture, circuit diffs, the predicate log. Strong systems Python, an eye for API design, and patience for other people's debug sessions.

1.2M tokens is easy to claim and hard to serve. This role is kernels, memory layout, and eviction policy — plus the interface work that reports what fell out of context and why. CUDA or Triton experience expected.

Turn policy prose into predicates that can be evaluated during decoding, then argue with everyone about the edge cases. Half research, half drafting. If you have ever wanted to be a lawyer and a logician at the same time, this is the closest thing.

Own the legibility suite. Design the benchmarks we will later be embarrassed by, and retire them before anyone starts optimising against them. Statistical literacy and a very low tolerance for leaderboards.

Everything we build is worthless if nobody outside the building can follow it. You would write the probe documentation, the model cards, and the explanations that make the negative results legible. Bring opinions about sentence length.

A checkpoint of your own, a desk, and an unconditional right to publish whatever you find — including the parts that make us look bad. We deliberately favour applicants who disagree with the manifesto. Two cohorts a year, four people each.