MonopoleAI builds research instruments for teams who treat machine cognition the way a botanist treats a forest—with patience, taxonomy, and an unembarrassed reverence for the actual.
Structured ingest from the messy field: documents, traces, telemetry. Each artifact pinned, dated, and indexed by provenance.
Models that read like a careful taxonomist—prefer to refuse a label than to invent one—and cite the leaf in the herbarium.
Every output traceable to its inputs, prompts, and weights. The notebook closes; the experiment can be repeated.
Five years of quiet, peer-reviewed work on alignment, retrieval, and the small grammars of structured uncertainty. We publish less than we ought to. We measure twice.
arXiv:2603.12041 · 18pp · 2026
A method for teaching models to recognize their own boundary—the serrated margin where competence ends and confabulation begins.
An open evaluation set of 3,142 carefully annotated reasoning specimens. Curated by working scientists across six disciplines.
cc-by-4.0 · mirror.monopoleai.com
A study of branch-and-rejoin attention patterns that mirror the hydraulic logic of a leaf vein network.
field measurements
17
peer-reviewed papers since 2021—each one earned the patience of three rejections before publication.
We ship a small number of carefully kept tools. Each one earns its place on the bench by being honest about what it can and cannot measure.
v3.4.1 · pyloupe
A retrieval lens for working scientists. Reads your archives, refuses to invent, returns to the page.
→ field manualv1.9.0 · quadrat-cli
A sampling frame for evaluation. Lay it on your dataset; what falls inside is measured, named, and kept.
→ field manualv0.7.2 · private beta
Long-term storage for reasoning specimens. Annotated, dated, signed by the practitioner who pressed the leaf.
→ request accessin service
3
tools maintained by hand. We reject the third party who asks for a fourth.
“A loupe should make the small thing larger and the loose thing more honest.”
— Field handbook, p. 11
Short essays, written slowly, about the practical matters of building research-grade machine learning. Updated when there is something to say.
The most useful answer a model can give is sometimes “I do not know, and here is the smallest piece of additional evidence that would change my mind.”
7 min read
Why we annotate evaluation cases by hand, on paper, before they touch a JSON file. The slow gesture catches errors the IDE will not.
11 min read
Notes on aesthetic discipline as engineering practice: how a clean visual grammar makes mistakes louder.
5 min read
Forty-two notes since 2022, each one with revisions noted in the margin. The archive is a living document.
→ browse the archivesubscriber count
2,841
readers receive a quiet quarterly digest. No tracking pixels. No urgency.
We are a small group of researchers, engineers, and scientific editors. We grow the way an old-growth stand grows—slowly, with care for the soil.
Where an output came from matters more than how confidently it presents itself. We surface citations before claims.
A model that says “not enough evidence” is more useful than one that fills the silence with a plausible sentence.
We ship when the experiment converges, not when the quarter does. Few releases, and each one defended.