Statistical Genomics Foundry

Glossary

Hand-curated and alphabetical. It carries the Foundry-pattern vocabulary this instance inherits — Mold, Cast, Pattern, Protocol, Phase, Provenance, the eval/scenario split — alongside the terms this domain coined: Family A and B, the referee, the gate, the bipolar corpus, the construct/critique/calibrate roles. If a design record disagrees with this file about a term, this file wins. It shares content/meta/ with the design records while being deliberately excluded from the collection: it is rendered by its own route, and skipped by the validator.

Analyze → referee → revise loop — the project's structural spine. An analysis (Family A) hands off to a referee (Family B), whose verdict gates certification; on a fail the analysis is revised (bounded) or escalated. Generalizes the parent's author → validate → fix loop; the novelty is that the referee node is a Mold and judges method validity. See content/meta/referee-loop.md.

Assessment axis — one of the four -ability dimensions a Rubric grades a Family-A doer skill on: Operability (can the tooling be run), Assessability (can the skill be tested), Traceability (do claims trace to recoverable sources), Refereeability (is the doing gated by a real Family-B referee). A skill scores as a vector of the four letters, never a composite. See content/research/mold-eval/rubrics/.

Bipolar corpus — our grounding corpus has two poles: established-good (methods + their validity conditions) and cautionary-bad (named invalidity patterns + signatures + remedies), plus supporting reporting-standard, calibration-methodology, and benchmark kinds. Unlike the parent's all-positive IWC corpus, a referee needs both poles. See content/meta/corpus.md.

Calibrate (Family B role) — construct and run an empirical check: permutation/null, simulation-under-known-truth, negative controls, power. The external verdict that is not self-certification. The role that turns the gate into a deliverable. Contrast Critique.

Cast (verb) — produce a self-contained skill artifact from a Mold via casting. (noun) — one casting result for a (Mold, target) pair.

Casting — the deterministic-assembly process that turns a Mold into a cast artifact, via per-kind dispatch over its references. The parent Foundry's pipeline allows an LLM condensation phase; this one does not implement it, so every carry here is deterministic. See content/meta/casting.md.

Cautionary-bad — the corpus pole of named invalidity patterns (double-dipping, batch confounding, naive multiple-testing, forking paths) with their signatures and remedies. Grounds the referee. No parent analog.

Construct (Family A role) — do the analysis: frame, design-review, select an established method, run reproducibly. Contrast Critique/Calibrate.

Corpus-first — abstractions must trace to observed practice, not be invented top-down. Load-bearing for us specifically: our failure mode is plausible invented prose, so a reference note starts as a stub and grows only when a real case demands it. We must not become the thing we referee. See content/meta/guiding-principles.md.

Critique (Family B role) — reason about validity against known invalidity patterns (double-dipping, confounding, assumption violations). Necessary but not sufficient — it is itself model reasoning, so the strong gate also requires a Calibrate pass.

Established-good — the corpus pole of established methods and when they apply. Grounds Family A's "pick an established method, don't invent."

Evaluation plan (eval.md) — the Mold's abstract oracle: fixture-independent property checks (how to judge any output), never concrete cases. For us, the highest-value properties are referee-correctness guardrails ("a double-dipped analysis must be flagged, never passed"). Never packaged into casts. Distinct from Scenario note.

Family A — do the analysis: frame the question, review the design, select an established method, run reproducibly. Leans on orchestrating existing doers + the anti-invention guardrail.

Family B — referee the analysis: audit method validity (Critique) and construct the empirical checks the field trusts (Calibrate). The project's teeth.

Foundry — the standalone, navigable knowledge base where Molds, patterns, corpus notes, and protocols live and from which casting reads. Working name: Statistical Genomics Foundry (provisional). Instance #2 in the Foundry-pattern lineage (Galaxy = #1). The substrate term for this source-of-truth artifact is Knowledge Base (KB).

Gate / gate obligation — the invariant that a Family-A protocol may not terminate in self-certification; it must hand off to a Family-B referee whose verdict gates certification. The project's defining structural rule. Encoded at the protocol altitude, not as a Mold property.

Knowledge Base (KB) — the inspectable, human-readable source of truth at the center of the instance — the standalone, navigable site where Molds, patterns, corpus notes, and protocols live and from which casting reads. Authored to be read and learned by a human, not merely stored for retrieval. The KB is the source; a skill artifact is the package. This instance colloquially calls it the Foundry (see Foundry).

License-aware summary — a Source note whose license resolves to verbatim-ok in the License-policy table, so short load-bearing verbatim quotes are permitted (marked as such). derived: license-aware-summary. E.g. harmon-pcm (CC-BY). Contrast Own-words summary; the choice between them is Summary posture.

License-policy table — the license → redistribution-policy table that decides, per SPDX id, what may be redistributed and in what form (policy, license_file, copyleft, obligations). Installed, not vendored: it ships in @galaxy-foundry/license-policy, the shared package both Foundry instances consume instead of hand-mirroring a root license-policy.yml (the decision behind it: galaxyproject/foundry-pattern#4). It answers what a license permits, and — from a note's derived: posture — whether the table governs that note at all, since an Own-words summary redistributes our prose rather than the source's. What it does not answer is whether a note is coherent with its license: that rule is instance-local, in site/src/types/context.ts. Determines Summary posture.

Method validity — the property our referee judges: are assumptions met, is there no double-dipping/circularity, is the named method real and appropriate, is its error rate actually controlled? The layer beneath the p-value — what hypothesis-validators (e.g. POPPER) take as trusted input.

Mold — an abstract, structured template: a typed reference manifest (frontmatter) + a procedural body skeleton. Directory note — content/molds/<slug>/index.md plus its two declared siblings, eval.md and scenarios.md, and nothing else. Cast into one or more skill artifacts. See content/meta/mold-spec.md.

Own-words summary — a Source note whose license resolves to own-words-only in the License-policy table, rendered entirely in new expression — no transcription — with only short functional strings (error text, parameter names, numeric thresholds, equation forms) kept verbatim as facts. derived: own-words-summary. E.g. msmb (CC-BY-NC-SA). Contrast License-aware summary.

Pattern page — reference content describing a statistical-method pattern (established-good) or an invalidity pattern (cautionary-bad). Wiki-linked from Molds; carried into casts. Cites corpus sources by URL/DOI.

Phase — one atomic unit of a Protocol's ordered sequence: a Mold reference, a [loop]-flagged Mold, or a non-Mold annotation ([branch] routing, [gate] checkpoint). The phase-kind set is open. The [gate] kind is the likely encoding of the gate obligation.

Pipeline / Protocol — an ordered sequence of phases composing a methodology journey (e.g. the method-validation arc: map-to-established → derive-null → simulation → power → sensitivity). Demoted from the parent's pipeline-primary IA: we are Mold-primary, so Molds may stand alone and Protocols are optional. Referenced content, not cast.

Progressive disclosure — show the right knowledge at the right time: protocols disclose the journey, Molds the action, references the dependency surface. Both an authoring principle and a runtime contract. The basis of Pillar 3 (knowledge foregrounded for a human reader).

Provenance — every derived artifact records what produced it (source hash, model, prompt version, resolved-ref hashes, timestamp). The basis of Pillar 1; never lightened. See content/meta/casting.md.

Referee — a Family-B Mold (or protocol stage) that judges an analysis's method validity using empirical checks rather than the agent's own reasoning. Non-self-certifying. The novelty vs the parent: the referee is a Mold, not a deterministic CLI.

Reference (a.k.a. reference kind) — a typed dependency a Mold declares; the kind discriminator is what selects the casting transformation. reference_contract.yml registers two — pattern and research — and registering a third is a deliberate edit made when a Mold first needs it. See content/meta/mold-spec.md.

Rubric — a standing, living scorecard grading a Family-A doer skill on one Assessment axis (A–F, half steps), grounded in probe evidence and updated as more skills are graded. Net-new to this instance; content/research/mold-eval/rubrics/. Traceability is the standing-metric form of the /ingest-bioskill recoverability probe; Refereeability grades the Gate obligation. Prose contracts, not tooling (scoring machinery deferred to repo standup).

Scenario note (scenarios.md) — the Mold's concrete test cases: a fixture/input binding + its expected verdict, exercised by the eval.md oracle. For us, richest as planted-invalid fixtures (a deliberately double-dipped analysis → expected "flagged"). StatQA is a ready source. Distinct from Evaluation plan.

Schema (Mold IO) — a JSON Schema declaring a Mold's structured input/output. A Foundry-pattern term this instance does not use: outputs here are prose-shaped critiques and protocols, so schema is not a registered reference kind. It gets registered if a Mold ever produces a genuinely structured artifact, such as a power calculation.

Self-certification — the failure mode the project exists to stop: an agent doing an analysis and blessing its own validity in the same breath, with fluent rationale and no external check. The gate obligation forbids it.

Skill artifact (a.k.a. cast artifact) — the compiled output of casting: self-contained, condensed, frozen against the source version, no links back, no runtime dependency on the Foundry. Carries _provenance.json.

Source note — a per-source reading note under content/research/<papers|tutorials|books>/<id>/: a regenerable faithful summary (index.md) plus, by exception, an owned guidance file (guidance.md). A regenerable cast (what the source says), kept separate from our framing. Graded by recoverability, not coverage. Its Summary posture is license-driven. For books, book-invariant metadata lives once in a co-located book.yml. See AGENTS.md "Source notes earn their keep by recoverability."

Summary posture — whether a Source note is an Own-words summary or a License-aware summary. Determined by the source's license (the License-policy table's policy: own-words-only vs verbatim-ok), not by source type — a book is not inherently own-words (msmb is; harmon-pcm, CC-BY, is license-aware). Recorded in the note's derived: field. The two summary commands split by workflow (bulk book vs single source), not by posture.

Target (a.k.a. cast target) — an output format casting can produce (Claude skill, generic, web). One Mold may cast to several.