Darkfield is, first, a lab. Our work is on building vision systems that decide their own pipeline, evaluate their own output, and improve themselves with no engineer in the loop. Below: the threads we're pulling on.
Each thread is led by a small team and reports against a single benchmark. Internal seminars on Wednesdays; external talks twice a year.
Resolving a noun phrase like "the third valve from the left" into a pixel cluster, with industrial-grade precision. THINK's grounding head is distilled from an internal annotation corpus we'd never use for training SEE directly.
A spiral mixer and a planetary mixer share most pixels. We train SEE on a hierarchy of industrial subtypes so a single prompt resolves to the right machine — not just the right category.
THINK samples SEE's output and scores it against a schema THINK itself authored. The hard part is making that audit trustworthy — and recovering when it isn't.
SEE specializes per-camera over its first 48 hours of operation, then continues to drift toward the scene as it ages. We treat the model as a population, not a single artifact.
An environment a model can operate in is a different design problem from one a human can operate in. Tools must be typed, retries must be cheap, and side effects must be reversible at the granularity the model reasons at. This is the longest-running and least public thread in the lab.
Open-vocabulary detection benchmarks test whether a model can find objects. They don't test whether it can tell a spiral mixer from a planetary mixer. DarkfieldOps-300 does.
300+ production-line objects across five industrial settings, each labelled at two levels: the parent class (mixer, filler, vehicle) and the subtype (spiral / planetary, rotary / linear / piston, tanker / flatbed / curtain-side). Footage is real CCTV — the angles, resolutions, and lighting conditions of actual deployments, not studio captures.
Standard mAP measures whether the right bounding box appeared. It treats spiral mixer and planetary mixer as indistinguishably correct if the box is in the right place. DarkfieldOps-300 introduces a type-aware score: macro-F1 over subtype labels conditioned on correct parent-class detection. A model that finds every mixer but cannot tell them apart scores near zero. A model that distinguishes subtypes reliably scores near one.
We have several papers written and ready, drawn from the threads above. We're holding them internal until the private beta closes — once partners are public, the work will be too. If you're a researcher and want to talk about the underlying methods, write to us.
We're onboarding a small number of partners in private beta.