ODIN LABS R&D B.V. · THE LAB ON A PAGE ·
How we evaluate models, tools and vendors.
Odin Labs R&D is the research lab of Odin Labs. It studies selected models, tools and vendors on bounded tasks from its own AI factory, where agents take part in software work under human approval. Published experiments do not establish a daily delivery rate or customer outcomes at scale. The lab works in the open or, by agreement, behind closed doors. Open work is published with the evidence: the question written down before the run, the method, the result and its limits, including when it went against us. Closed-door work stays private unless the partner decides otherwise.
How the lab works
- In the open
- A public experiment on the lab site: pre-registered, run, and published whatever it shows. Selected tools or vendors that fit the agreed question and terms, open source or commercial.
- Behind closed doors
- The same method under NDA, on a partner's code, data or model. Nothing is published unless the partner decides it should be.
Measured, with the record
Each number below is read from the published experiment record it links to, when the site is built. This page and selected evidence and open questions show examples, not a complete inventory of the lab's results or unanswered questions.
Price is not quality. On a sealed corpus of private changes, the cheapest reviewer cost $0.000631 per review and caught 29 of 60 seeded defects; the dearest cost $0.083274 and caught 31 of 60. The catch intervals of all reviewers overlap.
Measured . Source:
data/review-frontier/results.json· the field note.Brokers served what was asked. 720 identical requests through the brokers we use, at the same moment: 0 silent model substitutions for
claude-sonnet-5-5and 0 forgpt-4o-mini.Measured . Source:
data/broker-same-model/results.json· the field note.A negative result stays published. Can a reviewer model talk itself out of its own false alarms? The pre-registered answer was no: with a verify pass, its false alarms on clean changes stayed at 20 and 22.
Measured to . Source:
data/self-verify/results.json· the field note.
These results come from authored test cases and small private corpora. They do not establish a customer result at scale. Each record states its own limits.
Agents you can see and question
The lab's agent space is a research prototype with agents shown as characters a visitor can question. Published recordings include source cards: the question, the answer, why, and the cited source. A card records attribution; it does not establish correctness or guarantee that every answer has supporting evidence.
The partner programme
- Compute & token sponsors
- Providers expose tokens or credits to the factory and get real workloads and honest, pre-registered measurement. Sponsors never buy results, and no provider is benchmarked without its terms allowing it.
- Company-AGI partners
- Long-term companies that want their own governed factory, following the Odin vision. Exploratory, active and strategic levels; only measured figures are reported, the rest is marked not yet measured.
- Innovation sprints & research licences
- Fixed-scope sprints with evidence after every phase. Reports, datasets, methods and tuned gates are licensed; the intellectual property stays with the Holding. Selected tools or vendors within the agreed scope and terms.
No prices and no partner names here; every track starts with a written proposal, and no partnership is signed yet. Odin Labs reports an MIT application for Odin Next; this is author-reported, with no application or decision record linked here, and does not establish eligibility, approval or an award. Details: the partner programme.
Wording clarification · . The earlier headline and “Any tool or vendor” offer implied universal model and vendor coverage, and the earlier factory description implied a measured daily delivery rate. The published records establish neither. These are not “answers with proof”: the earlier label referred to source cards, which do not establish correctness. “Every answer” was too broad for the published recordings. The earlier subsidy qualification wording did not establish eligibility. The earlier “full list” link to Evidence implied a complete inventory that page does not provide. See our publication standard.
Odin Labs R&D B.V. · a wholly owned subsidiary of Odin Labs Holding B.V. · Start a conversation · odin-labs-ai.github.io/odin-rnd