Measure one
Useful work per unit of compute
Not benchmark points and not model size. How much real work gets done for the compute it costs, and whether that ratio improves as a system gains experience.
Status: PlannedResearch
Three questions, measured in the open.
Ymir Labs works on one problem: how to get far more useful intelligence out of every unit of compute. We have broken it into three questions, each of which can be checked.
01The questions
What we’re trying to find out.
How can a system reuse what it learned on earlier tasks so similar work costs less next time?
Status: In researchHow can an agent take a change of instructions in the middle of a long task and update only what is affected?
Status: In researchHow far can simulation results be trusted, and how do we measure that against real hardware?
Status: In research
These map directly onto what Candid is designed to do: get cheaper on familiar work, take corrections mid-task, and make predictions that are checked against real hardware. See Candid
02Keeping score
How we’ll know if it works.
Two numbers, reported together. The first is the goal. The second is how you decide whether to believe the first.
Measure two
Predictions that hold up
The share of our predictions that hold up when checked for real. For simulated experiments, that means comparing against measurements on a physical bench. The first such comparison is planned for November 2026, with timing dependent on funding.
Status: Planned03Publishing
What we publish, and how.
Results with their method
Every result is published with what is needed to reproduce it: the test set, the baseline it was compared against, and how uncertainty was measured.
Failures included
Negative and inconclusive results are recorded and published alongside positive ones. A result that sits within noise is reported as within noise.
Clear labels on everything else
Anything not yet measured is labelled as research, a plan or a concept. We don’t round intentions up to results.
Status: Measured A worked example
REN, a solo proof of concept by the founder, was evaluated on a held-out test set against its untouched base model, with 95% bootstrap confidence intervals. It counts a result as better only when the interval excludes zero. Six of eight measures cleared that bar. The one it was trained for did not, and the page says so.
04Method
On technical detail.
We publish questions, measurements and results. We don’t publish how Candid works internally while it is being designed. That is a deliberate choice, not an absence: the research specification exists.
Technical depth
We’re happy to walk researchers, programmes and investors through our technical approach in confidence. Get in touch.
The principles behind all of this are on one page. How we work