The Statistical Genomics Foundry is a human-navigable knowledge base for doing statistical genomics honestly. Its domain spine is analyze → referee → revise: constructive Molds propose work, referee Molds test method validity, and a gate prevents doing from terminating in self-certification. This page maps that system and points to the focused records that own its implementation details.
System map
established methods + cautionary examples
│
▼
authored knowledge under content/
papers, tutorials, books, patterns
│
supports Molds
│
▼
analyze → referee → revise
│
validated and rendered today
cast into artifacts in future
The corpus is bipolar: established-good sources ground what to do, and cautionary-bad sources ground what must be detected or refused. corpus owns that evidence model. Molds are the primary navigation and action unit; Protocols are optional compositions that emerge only when real Molds prove they chain. molds and referee-loop own those domain decisions.
What is inherited and what changes
The Foundry inherits the parent pattern’s source-to-cast distinction, provenance requirement, typed references, progressive disclosure, shared substrate packages, strict content contracts, and human-first reading site. It adapts the corpus from workflows to methods and invalidity evidence, and adapts deterministic checking from workflow-schema validation to empirical refereeing.
Two parent choices are intentionally lighter here:
- Structured Mold IO schemas are deferred unless a genuinely structured artifact earns one; critiques and validation protocols are prose-shaped.
- Pipelines do not define the inventory. Standalone Molds are first-class, and a methodology Protocol appears only when composition provides real value.
The project-specific addition is the gate obligation: doing never self-certifies. Whether that becomes a Protocol convention, an inline [gate], or a validator rule remains a decision for the first real composed protocol, not a schema invented in advance.
Two design records the parent keeps have no counterpart here: the harness and pipeline narrative, and the schema-package convention. Both lighten under the stance above, and neither has an implementation to describe — the first waits on a real Protocol, the second on a Mold that earns a structured IO schema. They are deferred, not overlooked.
Major boundaries
- Domain design versus implementation. Referee roles, corpus posture, and Mold semantics live in guiding-principles, referee-loop, molds, and mold-spec. TypeScript and Astro mechanics live in code-architecture.
- Knowledge representation versus physical placement. content-model defines kinds, tags, references, and note identity. repository-layout defines where their files belong.
- Authored versus generated. Source notes and registries are authored. Book-level metadata copies and kind manifests are generated and drift-checked. build-and-validation owns the current flows.
- Present versus planned machinery. The content contract, validator, generators, and site exist. A caster, package workspace, committed cast tree, and fixture toolchain do not. Design records may state their contracts, but architecture must not describe deferred machinery as implemented.
Focused architecture records
- code-architecture — the current Astro/TypeScript implementation, dependency seams, and entry points.
- content-model — note kinds, source-note distinctions, metadata, tags, references, and companions.
- build-and-validation — corpus validation, drift generators, site builds, and the explicit casting gap.
- repository-layout — the current physical tree and lifecycle-based placement rules.
Architectural invariants
Independent of each other; each is a thing that must stay true, not a step.
- The knowledge base is authored for people and remains the source of truth for future casts.
- A Mold’s claims must be recoverable from explicit references rather than model memory.
- Doing does not certify itself; a referee or an external deterministic check owns validation.
- Note kinds and registries are declared vocabularies, not directory or prefix guesses.
- The site and the standalone validator consume the same assembled schemas and collection table.
- Deferred machinery stays named as deferred until code and a test make it real.
Update this map only when a top-level boundary, domain spine, or reading route changes. Put detailed implementation changes in the focused record that owns them.