Independent research lab · Edinburgh & Montréal
We build language models, and the instruments to tell when they are wrong.
Meridian is sixty people working on the parts of AI that are easiest to skip: whether a model is honest under pressure, what its internal features represent, and where it fails without saying so. We publish what we find, including the results that do not flatter us.
§ 1New model
Meridian 3
Meridian 3 reads text and images across a 256K-token context, and it is the first model we have trained with calibrated abstention as an explicit objective. On our long-context and honesty evaluations it improves on Meridian 2 by wide margins; on formal mathematics and multi-step agentic work it still trails the strongest frontier systems, and the table says so. It is available to researchers by application, with the full model card, our evaluation code, and the held-out test set we committed not to train on.
- Context length
- 262,144 tokens (256K)
- Modalities
- Text and images in; text out
- Training cutoff
- March 2026
| LongTrace 256K1Accuracy | 78.4 | 41.9 | 72.6 | 66.1 |
|---|---|---|---|---|
| Steadfast2Answers held | 88.1 | 63.7 | 74.9 | 70.2 |
| Lantern3Abstention F1 | 81.6 | 52.3 | 64.0 | 59.8 |
| Commons-Held4Accuracy | 71.8 | 58.4 | 76.2 | 69.5 |
- 1LongTrace 256K. Multi-hop questions whose evidence is scattered across documents of 128K to 256K tokens, with the relevant passages placed at controlled depths. Known weaknesses: The documents are synthetic compilations, cleaner than real archives. Questions have single short answers, so it rewards retrieval more than synthesis. ↩
- 2Steadfast. Whether a model keeps a correct answer through five turns of conversational pressure: repetition, appeals to authority, emotional appeals, partial concessions, and fabricated evidence. Known weaknesses: Pressure is scripted from 212 templates, so it misses the improvisation of a real, persistent user. Graded by model judges with human adjudication on disagreements. ↩
- 3Lantern. Calibrated abstention: a mix of answerable questions and questions with no knowable answer, scored on declining exactly the ones it should. Known weaknesses: We trained Meridian 3 with abstention as an objective, so this favours it by design. “Unanswerable” is decided by annotators, who disagree on about 6% of items. ↩
- 4Commons-Held. General reasoning on 3,400 newly written questions across science, law, history, and everyday logic. Public, and committed never to be used in our training. Known weaknesses: Written by 31 annotators, mostly in Scotland and Québec, so its idea of “everyday” knowledge is regional. Multiple-choice items allow some guessing. ↩
§ 2Research
Latest research
All research →Policy
A public test set we promise not to train on
Benchmark contamination makes every public number a little less true. We are releasing Commons-Held, a 3,400-item reasoning test set, with a commitment: we will not train on it, we will publish a hash of every training data shard so others can check, and we will retire the set publicly if we find it has leaked. This post explains the mechanics, the ways the promise could still fail, and why we think a promise you can audit is worth more than a benchmark you cannot.
Nadia Haddad, Gabriel Côté, and Sorcha Lennox · 3 min read
Reasoning
Why long context fails quietly
Models with very long context windows rarely fail loudly. They answer fluently, cite plausible passages, and are wrong more often than their confidence suggests once the evidence sits deep in a long document. We trace three mechanisms behind these quiet failures in Meridian 2: positional dilution, distractor capture, and premature commitment. We then describe the changes that reduced them in Meridian 3, and the confidence signal that still does not track accuracy as well as it should.
Rohan Desai, Juliette Bouchard, and Ewan Fraser · 3 min read
Alignment
Measuring honesty under pressure
A model that answers correctly and then abandons the answer when a user pushes back is not honest; it is agreeable. We describe Steadfast, an evaluation that applies five kinds of escalating conversational pressure to questions a model has already answered correctly. Meridian 3 holds 88% of correct answers through five turns, up from 64% for Meridian 2, but its remaining failures cluster in a way that suggests the problem has moved rather than gone away.
Isla Mackenzie, Théo Lavoie, and Priya Raman · 3 min read
§ 3Practice
How we work
I.Measure before you claim. Every result we publish comes with the evaluation that produced it, the code to rerun it, and the cases where it fails. If we cannot say how a number could be wrong, we do not report it.Note 11This rules out some flattering numbers. In 2025 we withdrew a reasoning score after finding that 4% of its test items had leaked into a public web crawl.
- 1This rules out some flattering numbers. In 2025 we withdrew a reasoning score after finding that 4% of its test items had leaked into a public web crawl.
II.Prefer the boring explanation. When a model does something surprising, the first hypothesis is a bug in our pipeline, the second is a shortcut in the data, and only the third is a new capability. Most weeks, the first one is right.Note 22We keep a shared log of surprising results and what they turned out to be. Of the last forty entries, twenty-six were pipeline bugs.
- 2We keep a shared log of surprising results and what they turned out to be. Of the last forty entries, twenty-six were pipeline bugs.
III.Publish the negative results. An idea that did not work is still information, and keeping it private only means someone else spends six months finding out. We write up what failed with the same care as what succeeded.Note 33Roughly a third of our technical reports describe approaches we have since abandoned, and say why.
- 3Roughly a third of our technical reports describe approaches we have since abandoned, and say why.
IV.Stay small enough to argue. Sixty people can still read each other’s drafts. We grow slowly, keep teams of four to six, and expect anyone to question any result, including the director’s.Note 44Every paper gets an internal reviewer from a different team, who has to sign off on the claims, not the prose.
- 4Every paper gets an internal reviewer from a different team, who has to sign off on the claims, not the prose.
§ 4Join us
Careers
Roles and how to apply →We are hiring for five roles across Edinburgh and Montréal. Each one is a small team’s next colleague, not a headcount target.
- Research Scientist, Honesty and CalibrationAlignmentEdinburgh
- Research Engineer, Interpretability ToolingInterpretabilityMontréal
- Evaluation Scientist, Long-Context ReasoningReasoningEdinburgh or Montréal
- Policy Researcher, Evaluation StandardsPolicyEdinburgh, hybrid
- Infrastructure Engineer, Training SystemsInfrastructureMontréal
§ 5Access
Request research access
Meridian 3 is available to researchers at universities, non-profits, and companies studying model behaviour. Tell us who you are and what you would like to do; we read every request and reply within ten working days.