Statistical Genomics Foundry
The empirical-referee Foundry whose contrast with the original Galaxy implementation first made the shared pattern visible.
Read the instance’s own repo: https://github.com/jmchilton/statistical-genomics-foundry
Domain
Keep an LLM agent doing statistical genomics both productive and statistically honest. Left alone, such an agent is the only judge in its own loop: it produces plausible-but-invalid work — at worst inventing a method with a convincing name and no validity — and self-certifies it. The work splits into two families: Family A does the analysis (frame the question, review the design, pick an established method, run it reproducibly); Family B referees it (audit validity; construct the empirical checks the field trusts).
Corpus
Grounded in a deliberately bipolar corpus: established-good methods and cautionary-bad exemplars. The bad exemplars matter as much as the good ones — they are what the referee learns to recognize (double-dipping, confounding, naive correction, invented methods).
The check — from critique to empirical referee
No CLI can decide “is this statistical method valid,” so this Foundry makes the check authored knowledge rather than pretending a parser can supply it. Its current cast referee skills primarily critique analyses against known invalidity patterns. The stronger form also calibrates by constructing and running an empirical check—permutation under the null, simulation under known truth, or another field-appropriate calibration.
That empirical form remains the target rather than a fully implemented gate across the current Mold set. The intended analyze → referee → revise topology still matters: doing should not terminate in self-certification, but the profile distinguishes the critique skills that ship today from the calibration strength the architecture is working toward.
Targets
The Foundry now casts its authored Molds into committed portable skill artifacts using the shared @galaxy-foundry/cast machinery. Repository commands build individual casts or sweep the full tree, and the validation path checks byte drift so generated skills cannot quietly diverge from their Mold sources.
Status
Typed source and casting implemented; empirical calibration still maturing. The repository has a substantial typed corpus, deterministic corpus and contract validation, a human reading surface, authored Molds with scenarios and evals, and committed cast skills. The current skills make the critique side concrete; the stronger empirical referee remains the next correctness boundary to earn rather than a capability this profile assumes is complete.
Compare with the TDA Bioinformatics Foundry, whose stricter reference architecture is now being backported here, and with the Galaxy Workflow Foundry in the-diff. For the substrate tested across all three, see anatomy-of-an-instance.