Say less. Demonstrate more.

On capability claims

We describe models only by what they demonstrably do in our own products. We are interested in developmental approaches — how capability grows through use and release — and we make no claims about having achieved artificial general intelligence.

On evaluation

Our evaluation framework is public before our results are: see /research/evaluations. When results exist, we publish them as runs complete — including results that reflect poorly on our models.

On misuse

Products include reporting surfaces; uploads are checked; agent actions remain visible to the user who owns the machine. API keys are revocable, and abuse of the platform ends with the key.

On provenance

Generated media we publish as brand work is labeled (for example, "Generated with Caelim · motion reel 001"), and our public dataset ships with provenance and checksums so its origin is checkable.