⚒️ The Foundry Pattern

The Case

Related Projects

This is a map, not a scoreboard. Every project below is good at what it does; we link to it and credit its strengths. A project can be excellent and still leave some of these axes open — that is the nature of a map, not a knock. The columns are the axes the Foundry Pattern cares about, so a system whose strength lies elsewhere will read sparse here: a sparse row means off-axis, not deficient. (POPPER is the clearest case — its real strength, an error-controlled empirical referee, isn’t a column at all; we credit it in prose below.) Where a “distinction” fails verification, we strike it. See also comparisons for the compile-time-vs-runtime framing and skills-package-not-source for the argument these rows support.

The values table

ProjectProduces skills?Progressive disclosureTraceability (humans + agents)CLI invocation: derived vs improvisedPortabilityHuman scrutiny of data behind skillsKB-backed?
bioSkillsYes (~547 files)None (flat skill files)In-prose cites; no provenanceDerived — version-pinned, introspects APIHigh (5 runtimes + conversion)Hand-authored; readable + editableNo
POPPERNo (in-process library)— (validation framework)Low (Python lib)No
BiomniTools into own registry (not portable skills)None (agent UI)Traceable outputs (credit); KB machine-facingLow (framework-bound)Source readable in-repo; no reader surfacePartial (machine-facing substrate)
knowledgebase-mcp (cluster)No (retrieval server)None (operator CLI)Source paths + line ranges (credit)Corpus portable; MCP-boundOperator-inspectable; not a reader surfaceYes (operator-facing)
awesome-genomic-skillsNo (index of others’ skills)None (flat catalog)Links out (discovery layer)n/a (a list)No
The Foundry PatternYes (cast targets)Yes (navigable reader’s KB)Yes — provenance, cited + linked, for humans and agentsDerived — from referenced manual pagesYes (cast targets)Yes — read / correct / trace / contribute to the KBYes

Cells use ”—” or “n/a” honestly where an axis doesn’t apply to a system’s shape (a library, a referee framework, and a catalog are not in the skill-packaging business, so their packaging cells are blank by kind, not by failing).

The closest neighbors

bioSkills — nearest on produces skills, and more formidable than a quick read suggests. Concede plainly: hundreds of skills across dozens of categories, genuinely multi-runtime (five targets plus conversion), and a signature CLI-version discipline — mandatory version-compatibility blocks, introspection commands, “adapt to the installed API rather than retry.” Their embedded statistical wisdom is excellent (“thresholds are conventions, not laws”) and their relationship fields (depends_on, Related Skills) are format-validated. The distinction is layer, not quality: in bioSkills the skill files are the source, hand-authored; in the Foundry Pattern they are derived artifacts cast from an inspectable abstract source-of-truth, each carrying provenance, with a human reader’s surface above and an external check beside. A different layer of the stack — not a better skills list. (We do not claim it is Claude-only, not-for-humans, or relationship-free; all three were checked and are false.)

POPPER — nearest on the empirical-referee axis, and the table’s axes don’t flatter it precisely because its strength is off this map. Concede the strongest thing first: POPPER is a genuine empirical, non-self-certifying referee with provable Type-I error control (e-values, Vovk–Wang), using permutation tests and negative controls — that posture is shared, not ours to claim. The distinction is the unit refereed: POPPER controls error over the falsification decision, assuming each experiment yields a valid p-value; the Foundry Pattern referees the prior question — is the method producing that p-value itself valid? — and delivers it as a foregrounded, inspectable KB that casts portable skills, not an in-process library call. The referee posture is common ground we credit; only the unit and the delivery differ. (We do not claim its checks are weak or non-empirical — that is refuted and unfair.)

Verification discipline

Every claim in these rows is checked against primary sources — repositories, papers, and live sites — and a distinction that fails verification is struck, not softened. This is not hypothetical: three hypothesized limitations of bioSkills (Claude-only, not-human-readable, no formal relationships) were refuted on inspection and dropped, and the empirical-referee axis is recorded as shared with POPPER rather than claimed. Positioning built on a strawman is worse than no positioning.

One row carries a caveat worth flagging in the open: knowledgebase-mcp is a name-cluster, not a single project — the name resolves to several agent-facing retrieval MCP servers (jeanibarz/knowledge-base-mcp-server, Geeksfino/kb-mcp-server, olafgeibig/knowledge-mcp), none science-specific. The cell values hold across the cluster; if a specific repo is meant, confirm the owner. Any row added later without a primary-source check will be marked unverified until one is done.

The throughline is the one these neighbors help locate: the two assets no static skill file carries — provenance and an enforced check — are what make traceable, check-backed knowledge non-commodity. That argument lives in the-two-assets; this page only shows where the neighbors sit around it.