Worked examples¶
Real open-science tools, made agent-ready with OSCAR. Each is chosen because it exercises a different mix of archetypes, so together they cover the whole doctrine.
NEMAR (built)¶
The Neuroelectromagnetic Data Archive and Tools Resource, plus nemar-cli and nemar.org. Archetypes: data archive, command-line tool, and website. The crux: the data is not on the dataset page (it lives at data.nemar.org), so the download path and the license must be machine-readable beyond the HTML. Includes a tested generator and a sample four-hour CI. See nemar/.
HED (built)¶
Hierarchical Event Descriptors, hedtags.org. Archetypes: standard, website (multi-page), library (Python and JavaScript), web app and API, and it already ships a Model Context Protocol (MCP) server. The complete example: a project with a strong action layer that lacked the discovery layer. Shows how a multi-page site adopts OSCAR with one llms.txt, server-rendered JSON-LD, and an AGENTS.md. See hed/.
EEGLAB (built)¶
The MATLAB toolbox for human electrophysiology, plus eeglab.org. Archetypes: library / toolbox and website. The honest hard case: MATLAB has no package registry or docstring standard, so OSCAR adds a checked-in, machine-readable function manifest. See eeglab/.
HEDit (built)¶
The Annotation Garden's natural-language-to-HED annotator. Archetypes: command-line tool and library. The small, clean case: a pip-installable tool that already prints -o json; OSCAR adds the AGENTS.md and llms.txt. HEDit generates, HED validates. See hedit/.
Candidates as the landscape grows¶
We do not try to cover everything. Likely additions include MNE-Python (library; it already ships a CLAUDE.md and an AI-contribution policy worth imitating), OpenNeuro and DANDI (data archives; DANDI's OpenAPI plus /info endpoint is a model, and both archives' client-rendered dataset pages are a cautionary tale), and Neurodata Without Borders (NWB) as a standard.
Each example lands as its own directory with the recommended after/ files and a write-up of what changed and why.