LLM-driven statistical-genomics analysis fails in a specific, dangerous way: it produces statistically plausible-but-invalid work — at worst, an invented method with a fluent derivation and a memorable name and no validity — and certifies it with authoritative rationale. Monolithic “do the analysis” agents decay because they answer this with more reasoning, when the whole problem is that reasoning is what failed. These principles are the design pressure behind the alternative.
This project is not just a glossary, a docs site, or a pile of skills. It is an attempt to make statistical-genomics knowledge durable, inspectable, executable, and externally checkable while the surrounding tooling keeps changing.
Source Authority Beats Local Copies
Knowledge stays healthy near the project that owns it. Established methods belong in their papers and their reference implementations (Bioconductor, CRAN, tool repos). Tool behavior belongs in the tool. Reporting standards belong with their issuing bodies (EQUATOR, ClinGen/PRS-RS, MIQE).
This project should not become a stale mirror. It points upstream, quotes only what it must, and pins citations by DOI/commit when freshness matters. Concretely:
- Pattern and reference pages cite methods literature and cautionary examples by URL/DOI, not by importing them.
- Tool behavior comes from invoking the tool, not from prose we maintain about it.
- Vendored standards/checklists are copied into a cast artifact only when the artifact needs them; the issuing body remains the owner.
We add value by connecting, explaining, operationalizing, and refereeing upstream knowledge — not by competing to be its canonical home.
Reproducibility At Every Layer
Statistical analysis is itself a scientific act. A result is only useful if a maintainer can recover how it was derived, which assumptions it inherited, and which validity checks were applied. Reproducibility here is broader than rerunning the final analysis; it includes the provenance of every derived artifact: which Mold was cast, which model/prompt produced it, which references and methods grounded it, and which empirical checks (null, simulation, calibration, negative control) the referee ran. This is why casts record provenance and why refereeing is part of the loop, not a final cleanup.
The goal is accountable change, not perfect immutability.
External Instruments Do The Checking — the model never grades its own validity
LLMs are excellent at interpretation, synthesis, and translation; they are poor replacements for the things that establish statistical validity. What does that work here is the field’s empirical instruments: permutation and null calibration, simulation under known truth, negative controls, assumption diagnostics, error-rate control.
Do not let the model be the only judge of its own work. Fluent self-rationalization as the sole gate is precisely how a plausible-but-invalid analysis reaches a publication. The Foundry pattern states this principle in terms of schematic checks — does this parse? — because the parent Foundry has a CLI that answers it. The check here is empirical: is this calibrated, does it control error under the null, does it recover known truth?
Two consequences specific to us:
- Spend model context on judgment, spend the empirical check on validity. If a permutation test can reject a confounded result, run the permutation test; don’t ask the model to opine on whether the result is confounded.
- The empirical check is partly a deliverable, not just infrastructure. No parser ships for “is this method valid.” So constructing the right null / simulation / negative control is itself a first-class skill (Family B, the calibrate role). The gate moves from infrastructure into the work product.
Doing Never Self-Certifies
The structural bet, elevated to a principle. An analysis (Family A) must hand off to a referee (Family B) before it can be certified. Doing and refereeing are separated so that the agent cannot do-and-bless in one breath. The parent Foundry’s equivalent loop is gated by a deterministic CLI — infrastructure. This referee node is itself a Mold, and it judges method validity: the layer that even rigorous hypothesis-validators such as POPPER take as trusted input. referee-loop turns this into architecture.
This principle is why the project exists. Everything else serves it.
Progressive Disclosure Over Context Flooding
Agents — and humans — should see the right knowledge at the right time. Don’t flatten every method, diagnostic, checklist, and cautionary example into one prompt or one skill body just because it exists. Protocols disclose the journey; Molds disclose the action; typed references disclose the dependency surface; load policy distinguishes up-front from on-demand; casting decides copied vs sidecar vs inlined.
The goal is navigable depth: a human browses from protocol to Mold to reference; an agent moves from action to supporting evidence without dragging the whole library into every step. This is also Pillar 3 of positioning — knowledge foregrounded for a human reader, not merely stored for an agent to retrieve.
Portable Artifacts Over Platform Fashion
The agentic-coding landscape keeps changing. Core knowledge should not bind to one runtime, editor, model vendor, or framework. A Mold is a typed reference manifest plus a procedural skeleton; casting turns it into a target-specific skill, but the source of truth stays abstract and inspectable. Molds are durable source; cast skills are generated targets; a new runtime is a new cast target, not a rewrite.
Actionable Knowledge, Not Passive Notes
A passive knowledge base explains but cannot make an agent act. A standalone skill acts but compresses away the evidence and rationale that make it maintainable. We keep both: the site preserves the rich graph (methods, diagnostics, checklists, citations, rationale); Molds identify which knowledge a concrete action needs; casting condenses it into executable artifacts; protocols show how actions compose. The wager: a knowledge base becomes more useful when its structure makes it executable, and a skill becomes more trustworthy when its source stays inspectable.
Corpus-First, Not Invention-First
Learn from established statistical practice before inventing abstractions. Our grounding corpus is methods literature + cautionary negative examples (Bioconductor/OSCA, GWAS QC protocols, calibration/simulation standards, and named-invalidity literature like double-dipping and batch effects). Corpus-first means abstractions are justified by observed practice: pattern pages cite real methods; Mold behavior aligns with recurring analysis and refereeing tasks; new taxonomy appears after content demands it.
This principle is load-bearing here in a way it is not for the parent Foundry. The failure mode this project exists to stop is plausible-sounding invented prose that propagates the author’s priors and reads as authoritative — and a pre-written, comprehensive-looking reference note is indistinguishable, to a downstream agent, from an earned one. So the discipline is strict: a reference note starts as a stub — frontmatter, title, primary-source link — and grows paragraph by paragraph only when a real case demands it, a cast run or a refereed analysis or a logged failure. Write nothing until contact with the corpus forces it. We must not become the thing we referee.
How The Principles Connect
Keeping methods at their source makes citation-not-mirror possible, but only works if derived artifacts record provenance. Provenance is meaningful only if external instruments — empirical ones — perform the checks the model should not grade itself on. Those instruments are reusable when artifacts are portable, which needs an inspectable source of truth, which pushes toward a knowledge base, which becomes actionable through Molds, casts, and protocols. Doing Never Self-Certifies is the spine through all of it: the empirical referee is where provenance, external checking, and corpus-first meet. And Progressive Disclosure is the connective tissue that keeps the source record rich without forcing every artifact to carry the whole library.
The resulting division of labor:
- Upstream methods literature owns the facts.
- This project owns synthesis, structure, casting source, and the referee.
- Empirical instruments (null/simulation/calibration) own the validity verdict.
- Cast artifacts own target-specific execution.
- Harnesses own orchestration.
See Also
- positioning — how this shape is distinct from, and complementary to, the prior art, with verified evidence.
- referee-loop — the analyze → referee → revise spine that “Doing Never Self-Certifies” turns into architecture.
- architecture — the system map, and what this instance inherits from the Foundry pattern versus what it changes.