A hypothesis referee that says a confident no. Every scientist is caught between two floods — over a million papers a year and omics runs returning millions of measurements per sample. Wayfinder helps you navigate the gap: it surfaces the data's implicit hypotheses with literature-based discovery, then tests each one against real experimental data and returns a verdict with a receipt for every hop — supported, refuted, or untested (the knockdown failed). The library and the lab, on one bench.
▶ 3-minute demo video: https://youtu.be/MbgojaAFfz0
Built for Built with Claude: Life Sciences (Anthropic × Gladstone Institutes), July 2026 — Researcher track ("a researcher who also builds"). The whole pipeline was reproduced natively inside Claude Science, Anthropic's scientific workbench.
Three NIH reviewers killed the R01 behind this line of work (1R01LM015392-01) with one recurring critique:
"This application has the same major problem that has plagued all LBD work: it generates an enormous number of hypotheses, almost none of which ever get followed up."
They were right. Literature-based discovery overproduces hypotheses and tests almost none. Wayfinder is the answer to that critique — it puts hypothesis generation and testing on one bench, culls aggressively, and is willing to say a confident, receipt-backed NO. Falsification, not confirmation, is the moat.
On the genome-scale CD4+ T-cell Perturb-seq resource (Marson lab), Wayfinder posed 22,039 gene→program→ disease hypotheses; a deterministic referee culled them to 30 clean, receipt-backed survivors. The standout: NAB2 → Th1/Th2 polarization → atopic eczema — a connection the literature had never drawn, re-derived here with a receipt at every hop.
- The confident NO (the moat). A failed knockdown returns untested, never "no effect": IL2's knockdown-QC gate failed (0/2 guides reached significance) → untested, an artifact caught rather than a false negative. NAB2's gate passed (2/2 guides, adj_p 1e−16), so NAB2 can be judged.
- A confounder refuted against the authors' own data. NAB2 and STAT6 (a master Th2/atopy regulator) are convergent tail-to-tail neighbours (~43 kb between promoters), so a CRISPRi cis-artifact was the obvious worry. Reading the authors' own deposited genome-wide DE matrix, STAT6 is unmoved by NAB2 knockdown (log2FC +0.09, adj_p 0.79) while NAB2's own transcript drops (−3.08 log2FC, z≈−17) → the cis-artifact is refuted.
- Calibrated, not overclaimed. The eczema link is an Open Targets GWAS-genetic nomination (no colocalization); the perturbation-backed claim is NAB2 as a Th1/Th2 regulator. NAB2 stays a nomination — never "proven" or "discovered."
- QC gate — a null result counts only if the knockdown actually worked (receipt: guide knockdown-efficiency).
- Gene → effect — did perturbing the gene produce a real, reproducible transcriptional effect? (receipt: DE stats + cross-donor/guide reproducibility)
- Gene → program — do the downstream effects shift a real T-cell program? (receipt: Th1/Th2 polarization signature)
- Program → disease — does the downstream cluster enrich for an autoimmune disease? (receipt: cluster→disease odds ratio + intersecting genes)
Every causal edge traces to a data receipt (odds ratio / p-value / effect size). Claude interprets receipts; it does not assert biology. Deterministic data lookups are tools, not "agents" — visible agency is reserved for judgment.
The generation → referee → provenance pipeline ran natively in Claude Science (Anthropic's scientific workbench), which authored and ran the literature-based-discovery generator live and executed a multi-model actor–critic: an Opus 4.8 author with an independent Sonnet 5 reviewer that verified every number and enforced calibrated language — it flagged the words "validated" and "definitive" as overclaims, and they were cut. The platform checked its own work.
I build with more than one model on purpose: combine their strengths, and use each to check the other — while keeping the science itself on one platform. Nothing here is hidden.
The science ran in Claude (Anthropic), end to end. The generation → referee → provenance pipeline executed
natively inside Claude Science — verified from its own audit store (operon-cli.db) as Opus 4.8 (author),
Sonnet 5 (the independent reviewer), and Haiku 4.5 (inline). Every scientific receipt and number in this
repo was produced by deterministic tools + Claude; no other model computed a result in the pipeline. Claude Code
wrote the fresh tool layer during the event and orchestrated the build.
A second model family — OpenAI's Codex (via the codex CLI) — was used as a cross-model engineering aid,
deliberately kept outside the science pipeline: (1) code review that caught real bugs before any result
depended on them (attributed inline in the source); (2) adversarial red-teaming of the plans, the LBD spec,
and the manuscript (docs/reviews/codex-*); and (3) independent replication — 2 of 5 agents in the
cross-model replication were Codex, clean-room re-deriving the NAB2 finding to confirm it reproduces across model
families (docs/replication_report_2026-07-08.md).
Using a rival model family to attack and re-derive the work is not a shortcut around Claude — it is the integrity posture. The finding is stronger because it survived a second model's scrutiny, and the science it describes was done in Claude.
bash data/fetch_data.sh # public supplementary tables (~25 MB); no raw data committed
# then, from a Python env with the repo installed:
# open notebooks/ — the evidence-chain notebook recomputes every headline number live from the tables
# or call the referee directly: from arbiter.lbd import referee_tripleFull method + build plan: docs/plan.md. Independent 5-agent replication: docs/replication_report_2026-07-08.md.
Native Claude Science pipeline + provenance: docs/cs-full-pipeline_2026-07-09/.
Public supplementary tables from the genome-scale CD4+ T-cell Perturb-seq study — Zhu, Dann, … Marson
(Gladstone/UCSF), bioRxiv 2025, doi:10.64898/2025.12.23.696273 — used under the hackathon's public-dataset
provision; all analysis performed during the event. Provenance in data/README.md; analysis reference:
github.com/emdann/GWT_perturbseq_analysis_2025. This repo commits no raw data.
Built by Dayanjan Shanaka Wijesinghe, Ph.D. ("Shanaka") — tenured Associate Professor at Virginia Commonwealth University School of Pharmacy and former Deputy Director of VCU's Lipidomics/Metabolomics Core Facility. Having spent years handing researchers data matrices of thousands of features and fielding the same question — "what does it all mean?" — he now builds AI instruments for precision medicine and drug discovery. linkedin.com/in/dayanjanwijesinghe
Code: MIT — see LICENSE. All files authored during the event (started 2026-07-07); the git history is the
new-work compliance proof.