Evidence

Status tracker for what the paper can stand on. The manuscript’s Evidence section states the requirements; this file tracks what is built versus what is still owed.

Built (citable as system description)

• Typed content base: 54 Patterns, 41 Molds, 6 Pipelines, 13 Schemas, 6 CLI reference sets, 63 research notes (regenerates from repo; pin commit SHA in SI).
• Validator with strict frontmatter + controlled tags + cross-file checks (reference dispatch, pipeline-phase resolution, Molds = union-of-phases invariant, artifact graph).
• Casting pipeline producing 31 Claude-target casts, each with _provenance.json (schema v2).
• Five @galaxy-foundry/* packages; Astro site with raw-Markdown endpoints.

Still owed (the real gap — in progress)

The central efficacy claim is not yet demonstrated. Tracked in case-study.md:

• One clean construction case study (the path that runs): real analysis intent → freeform-summary → design Molds → exemplar comparison → gxformat2 draft → per-step resolution → gxwf validation → planemo test → provenance trace. Conversion (Nextflow/CWL/paper → Galaxy) follows once a path is exercised end-to-end; only summarize-nextflow ships runtime code today.
• A failure-comparison vignette (monolithic skill / unguided agent vs. the decomposed loop) on the construction task — fabrication-catch is the load-bearing metric.
• A provenance walkthrough (one SKILL.md paragraph traced to Mold + source ref).

Not evidence: the _emulated-runs/ dev test-drives in the project repo. They are internal harness shake-outs that surface gaps; they are not publishable end-to-end conversions and must not be presented as results.

Risks

• Without a completed case study this reads as architecture only — own that explicitly (the manuscript does).
• Keep comparisons primary-source backed and dated; no vendor-landscape overclaiming.
• Present as an early model, not a mature automated conversion system.