Manifesto
Six positions
We wrote these in a room in Zurich in 2021 and have revised them twice since. Both revisions are in the footnotes, because a manifesto that never changed was never load-bearing.
A capability you cannot inspect is a rumour.
Benchmarks measure whether a model was right this time. They say nothing about whether it will be right on the case you actually care about. Inspection is the only tool that generalises.
Legibility is a capability, not a tax.
A model forced to produce a derivation it can defend makes different mistakes than one optimised only for the answer — fewer of them, and more findable. We keep measuring this because we keep being accused of wishful thinking.
Publish the negative results or stop calling it research.
Our most-cited paper reports that derivation traces did not reduce user overreliance. It cost us a product launch. It was the most useful thing we published that year.
Scale is a budget line, not a philosophy.
We train large models because large models are currently better, not because size is the point. The day a 3B model reads as clearly and answers as well, we will happily delete the cluster.
Refusal without a reason is just failure with better manners.
When Noema declines, it names the predicate it invoked and links the clause. If we cannot write the clause down, we have not thought the policy through — and we do not ship it.
Somebody outside this building has to be able to check.
Hence open weights for Micro, a source-available probe stack, and a residency programme with an unconditional right to publish. Trust that cannot be audited is just branding.
History
Five years, abbreviated.
A fictional history for a fictional lab — written to make the design problem concrete rather than to describe anything real.
-
2021
Founded in a rented room
Four people, one leased cluster, and a shared conviction that interpretability was being treated as a garnish.
-
2022
Naming the heads
The first stable taxonomy — 240 named attention heads that survived retraining. It is still the backbone of everything we do.
-
2023
Noema-Micro, open
We published weights before we published an API, which several investors described as backwards.
-
2024
The negative result
Traces did not reduce overreliance. We shipped the paper and cancelled the launch.
-
2026
Noema-1
340B parameters, 1.2M context, and a derivation for every answer. The first model where legibility did not cost us the frontier.
People
Who signed it.
Invented people, invented roles. Any resemblance to actual researchers is coincidental.
-
Ilse Vantree
Director, Interpretability
-
Ravi Kalindi
Head of Training Systems
-
Mira Østergaard
Alignment Predicates
-
Tomás Abreu
Probe Tooling
-
Sena Nakamura
Long-context Research
-
Joss Bariteau
Evaluation
-
Adaeze Diallo
Residency Programme
-
Petra Wu
Technical Writing
Disagree with us.
The residency exists specifically for people who think we are wrong.