# Sediment

> Sediment Labs works on memory, context engineering, and adapting agents to
> new domains. We build systems that read an institution's own documents and
> answer, draft, and abstain in that institution's voice.

## Pages

- [Sediment 1, a case-study](https://sedimentlabs.ai/sediment-1): A local
  open-weights model against three frontier systems on a private-equity
  analyst benchmark over two synthetic firms of roughly 25 million tokens
  each. Reports an ablation ladder over one model, accuracy by task family
  with 95% intervals, a card average, cost and latency, and a graded
  drafting comparison.
- [Research](https://sedimentlabs.ai/research): Papers from the lab.
- [Company](https://sedimentlabs.ai/company): People, and how to work with us.

## Notes

- The benchmark harness is public: https://github.com/zacharyspeck/sediment-benchmark-harness
- The corpus and answer keys are private, which is what keeps the benchmark valid.