Evaluation & safety
Publish the failures.
At Geocentric, safety is the evaluation that decides whether a checkpoint gets released, and the disclosure that goes out with it. A separate department writing principles has no part in it.
A model we cannot characterise is a model we do not release.
Every Geocentric model is published with its full evaluation. Arc's model card, for example, publishes all 24 responses from its diagnostic suite and all 12 writing probes, unedited. Roughly half of them are failures: invented facts, broken arithmetic, ignored instructions. They are on the page because removing them would misrepresent the model.
This is the smallest useful version of a safety practice, and the one we can honestly claim at our current size. We will not describe an alignment programme we have not built.
Kestrel, our flagship, is announced but not publicly available because of security and safety concerns. Read why.
Test the failure modes separately.
A single score hides the thing you need to know. PARALLAX measures the axes independently so a regression is visible where it happens.
Predeclared cases
Evaluation prompts are fixed before the run. One response per case, no retries and no best-of selection.
Separate axes
Text modeling, context use, completion ranking, instruction following, degeneration, and serving performance are reported separately.
Failures published
Every response is published, including the ones that are wrong or incoherent.
Stated methodology
Sampling settings, hardware, and scope accompany every number, so a result can be repeated or disputed.
What our models should not be used for.
Consequential decisions about people
Geocentric does not deploy its models to make or substantially assist decisions about employment, housing, credit, insurance, education, health care, or access to government services, and our Acceptable Use Policy asks customers to do the same. Asking a model a question about these subjects is fine. Letting it decide something about a person is out of bounds.
Unreviewed code
Generated code can be incorrect or insecure. Review and test it.
Sources of fact
Our models have no retrieval and no access to current information. Its fluency is no evidence that a statement is true.
Disclosure, in full.
Model behaviour, training data, and regulatory position are documented rather than summarised.