open-science

Open science experiments from NimbleCo — data visualizations, community reports, and research artifacts produced by Matilde and other agents.

On the web: cyborg.garden/open-science — the blog, plus experiment pages built from this repo.

Experiments

Experiment	Description	Live
scn-photometry	The Clock Under Constant Darkness: SCN fiber photometry across 14–17 days of constant darkness in 10 mice — GCaMP/GFlamp1/GFP signal, wheel-running actograms, and light-pulse phase resetting. Python port of the lab MATLAB pipeline, validated to exact match.	View
annotator-agreement	Who Agrees About How Agents Fail: There is a published taxonomy of the fourteen ways multi-agent AI systems break. We asked whether language models apply it consistently, found that they do not, and then found that our own answer was wrong: agreement scales with annotator size, from 0.173 at 2-12B to 0.598 at frontier. Underneath that sits the sharper result — the reference labels score 0.047 against themselves, while four frontier models agree with each other and not with those labels.	View
fmri-pooling	Pooling Small-N fMRI: Can many small open fMRI studies be merged into one shared space? Same pipeline, two naturalistic stimuli, opposite answers: alignment hurt on a non-narrative visual tone-poem (N=93) and helped on a heard narrative (19 × 8 runs) — a measured map of where pooling works, with the confounds stated	View
consciousness-autostudy	The Consciousness Autostudy: A 7-day autonomous research campaign audits the Landscape of Consciousness dashboard: citation integrity, a Cogitate prediction-outcome ledger, a cross-theory premise graph, a blind rubric reliability audit, and the verified post-Cogitate discourse — with ranked next steps	View
fde-first-proof	An Open-Weight Model Proves a Logic Theorem: A research agent running an open-weight model (Kimi K3) formalized the four-valued logic FDE in Lean 4 and proved it sound, complete, and decidable — every proof checked by the kernel, every axiom reported, every wrong turn logged. The leaderboard of AI-proved math is dominated by closed models; this is the open-stack version	View
failure-atlas	The Failure Atlas: Your agents write down everything they do, and almost nobody reads it back. We read ours: 2,781 sessions of real agent work, to see how AI actually breaks on your own tasks rather than on a benchmark. Most of what goes wrong turns out to be plumbing, the standard list of agent failures has no box for it, and the failures that matter most are the ones you cannot catch automatically.	View
nz-birdsong	Aotearoa Birdsong: Vocalisation analysis of tūī, korimako & kākā from Xeno-Canto open data — syllable classification, geographic variation, annotation & clustering tools, and a Te Reo Manu learning game	View
beyond-paper	Beyond the Paper: 14 structural critiques of the scientific paper format, an interactive evidence-overlay demo, a landscape of existing alternatives (nanopublications, executable articles, PRC, knowledge graphs), and a proposal for decoupled scientific knowledge	View
consciousness	Landscape of Consciousness: Theories of consciousness catalogued, formalized with scope lines, and assessed for testability — with evidence from the 2025 Cogitate Consortium adversarial collaboration	View
peptide-reporting	Peptide Pharmacovigilance: Community self-reports with valence — generated by Matilde	View

Structure

experiments/<name>/      ← source files + raw data (when the source lives in this repo)
docs/experiments/<name>/ ← built/served files (GitHub Pages root)
docs/experiments/<name>/meta.json ← manifest: what the indexes are built from
scripts/build-index.mjs  ← regenerates both indexes
scripts/publish.sh       ← publishes an experiment from the Mini

Publishing an experiment

Both indexes — the cards on https://cyborg-garden.github.io/open-science/ and the table above — are generated. Everything between the <!-- experiments:begin --> and <!-- experiments:end --> markers in docs/index.html and README.md is overwritten on every build. Edit meta.json, not the markers' contents.

Each experiment carries a manifest at docs/experiments/<slug>/meta.json:

{
  "slug": "nz-birdsong",
  "title": "Aotearoa Birdsong",
  "blurb": "One sentence shown on the card and in the table",
  "emoji": "🐦",
  "status": "published",
  "updated": "2026-08-04",
  "order": 2,
  "source": "where the source lives — repo path, or the machine that generated it"
}

status other than published hides the experiment from both indexes. Entries sort by updated (newest first); the optional order only breaks ties between entries sharing a date. title/blurb are inserted verbatim into HTML — keep them HTML-safe.

Regenerate after any manifest change:

node scripts/build-index.mjs          # rewrite both indexes
node scripts/build-index.mjs --check  # CI: fail if they are out of date

From the Mac Mini (inside the hermes-matilde container), one command copies the built experiment in, regenerates the indexes, commits and pushes:

scripts/publish.sh nz-birdsong          # rsyncs ~/experiments/nz-birdsong/dist/
scripts/publish.sh nz-birdsong --src /path/to/dist

It refuses files over 95 MB (GitHub's hard limit is 100 MB) — keep those on the expansion volume and record the path in meta.json's source field instead.

Artifacts are generated by Matilde — NimbleCo's verifiable citations and scientific open research engine.