The Diff: What Transfers
One implementation can make any local convention sound universal. The contrast between the Galaxy Workflow Foundry and the Statistical Genomics Foundry exposed a candidate pattern. The TDA Bioinformatics Foundry then applied the harder test: re-derive the domain model, consume common mechanisms through explicit contracts, and see which similarities survive without copying the application.
The result is a smaller claim than this page once made—and a stronger one.
Three implementations, different pressures
| Dimension | Galaxy Workflow | TDA Bioinformatics | Statistical Genomics |
|---|---|---|---|
| Focus | Construct and test Galaxy workflows from papers, Nextflow, and CWL | Harden frontier TDA into reproducible bioinformatics practice | Audit statistical-genomics claims and methods |
| Evidence | Workflow exemplars, tool interfaces, and schemas | Methods, papers, packages, locked environments, recipes, and replication records | Established-good methods and cautionary failure cases |
| Delivery | Cast skills composed into mature, ordered Pipelines | A typed corpus and its first authored Mold for scoring candidate structures | Committed standalone referee skills; composition remains optional |
| Check | gxwf validation while designing, then Planemo execution |
Corpus, build, and citation gates plus runnable environments and replication evidence; the scientific check is still evolving | Authored critique/referee skills today; empirical calibration is the stronger target |
The commonality is not a shared scientific workflow. It is the machinery that lets each Foundry state, inspect, package, and check a different one.
What survived all three
- A typed, human-readable Knowledge Base is the source of record. Strict schemas, controlled tags, and resolvable links make authored knowledge inspectable and mechanically usable without turning the rendered site or a generated skill into the source.
- Molds declare one action and the references it needs. All three carry the Mold contract; their live manifests agree on the core evaluation and scenario companions even though the actions and reference kinds differ. See the generated kind-catalog.
- Generated outputs have reproducible producers. Manifests, indexes, audit reports, and cast artifacts are checked for drift. Where casting exists, the artifact is compiled from the Mold and accompanied by provenance rather than maintained by hand.
- Shared mechanics take instance facts as inputs. Versioned packages can own schema assembly, manifest formats, registries, links, reading-shell behavior, reference contracts, and casting primitives. The domain still owns its kinds, vocabulary, routes, presentation, and policy.
The TDA Foundry is the clearest evidence for that last boundary. Its adoption work repeatedly asks whether a need is local knowledge, an existing shared mechanism, or a contract worth improving upstream. That discipline—not a common folder tree—is what now feeds improvements back into the other Foundries.
What remains local
The corpus, taxonomy, composition model, and meaning of correct do not transfer. Galaxy needs ordered Pipelines and two layers of workflow validation. Statistical Genomics can ship standalone referee skills and strengthen critique into calibration. TDA uses its own frontier → hardening → delivery arc, with reproducible environments and replication evidence, while its final scientific check remains unfinished.
Implementation maturity differs too. Galaxy exercises the broadest source-to-cast path. Statistical Genomics now casts its authored referee Molds and checks the committed artifacts for byte drift. TDA is the current architecture-hardening reference but deliberately has no caster or cast tree yet. The pattern names the delivery boundary; this page does not pretend every Foundry has reached it.
Bottom line
A Foundry keeps typed, human-readable knowledge as its source; uses Molds to select what an action needs; builds runtime artifacts through an explicit casting boundary; records provenance when it casts; and requires a check the work cannot award itself. The domain owns the vocabulary, composition, presentation, and meaning of correct.
For the normative boundary, read What a Foundry Needs. For live cross-instance details, use the generated kind-catalog and tag-catalog rather than freezing another inventory into prose.