How we work
Measure, then claim.
A research lab is only as good as its claims. A deep-technology company that claims what it cannot prove does not last. So before building anything, we decided how we would make claims, and how we would keep them honest.
01Principles
Ten working rules.
Evidence over spectacle.
Tests, sources and reproducible experiments come before announcements. If we can’t show it, we don’t say it.
Build for decades.
Research before hype. Prototypes before promises. We are building a lab meant to last, so each step has to be real before the next one starts.
Available hardware before custom hardware.
We learn real workloads on hardware anyone can get before taking on the cost and risk of anything bespoke. REN was trained entirely on free GPUs. See REN
Gates before scale.
Every expansion has to pass a gate first: technical, safety, legal and funding tests, each with a condition under which we stop. Growth is earned, stage by stage.
Publish only what exists.
Everything else is labelled as research, a plan or a concept. This site follows the rule: every claim that isn’t a plain fact carries a label.
Write the hypothesis first.
We write down what we expect before running an experiment, and compare against a cheap, simple baseline. Otherwise almost any result can be made to look like a win.
Keep the failures.
Negative and inconclusive results are recorded and published alongside the rest. A result within noise is still information, and it is reported as exactly that.
Data you can trace.
Training data should be licensed and traceable, and test sets checked for overlap with training data. REN’s data came from openly licensed sets and programmatic generators, with no outputs from OpenAI or Anthropic models.
People stay in charge.
Systems should be easy to correct partway through, and their work easy to undo. It is one of the outcomes Candid is designed for, and a standard for everything we build.
Earn reliance. Never trap it.
We want people to depend on what we build because it is reliable, documented and efficient, never because leaving is hard.
02Reading this site
Every claim carries a label.
The doctrine says to publish only what exists and label everything else as proposed. These are the labels you will see, and what each one promises.
- Status: Measured
- A result that exists, tested and published with its method.
- Status: In development
- Being designed and validated. Not launched.
- Status: In research
- Being studied and designed. Not built yet.
- Status: Planned
- Something we intend to do. Timing depends on funding.
- Status: Concept
- An idea we have written down and will test before building.
- Status: Exploratory
- Long-range reading and thinking. No claims attached.
The labels change as the work does. When something moves from planned to measured, the change will be dated in the log.
03A worked example
The row we could have hidden.
REN, the founder’s proof of concept, was trained to handle being interrupted mid-answer. On that exact measure, it improved from 33 to 35 out of 42, and the confidence interval includes zero. So the results table says “within noise: not a win yet”, in the same type and the same size as the six rows that cleared the bar.
That is what these rules look like in practice.