about
zer0 research measures what agent harnesses, small models and generated interfaces actually do — cost, speed, memory, governance — and publishes the receipts, including the negative results.
what this is
zer0 evals is the public, reproducible evaluation layer of the zer0 agent stack. other repositories run the experiments and own the production behaviour; this one freezes the evidence and publishes what it does and does not show.
how a study becomes a post
- an experiment runs somewhere else and is committed there.
- an import spec pins that source by commit sha and allowlists the files that may be published.
- the frozen data is validated, analysed and turned into figures.
- the post says what was measured, what it does not prove, and links every receipt.
no claims without receipts. negative and incomplete results stay on the site: a miss against a pre-registered bar is still evidence. the research log lists every eval, experiment and idea, including the ones that never became a post.
evidence status
every post and study carries one of these. status is always a dot and a word, never colour alone.
- living
living- measured locally
evidence-local- mixed evidence
evidence-mixed- evidence pending
evidence-pending- unscored
unscored- synthesis
research-synthesis- architecture
architecture- draft
draft
research-log entries use their own lifecycle:
- published
published- frozen, not yet posted
frozen- measured, internal evidence
frozen-internal- draft study PR
draft-pr- in progress
in-progress- proposal
proposal- idea / paper
idea- stopped
abandoned
source
everything here is built from the z0evals repository: the posts, the frozen studies, their manifests and the scripts that render this site.