Galaxy Brain

figure5-draft-resolves

Figure 5 — The workflow draft resolves (draft spec)

Source-of-truth content for Figure 5: one real workflow step shown at three resolution tiers over a fixed, concrete topology, ending in promotion to class: GalaxyWorkflow and a green gxwf check. The “under-determination as a typed state” payoff figure.

Step shown: DESeq2 differential test — the central step of the UC3 differential-ATAC construction run, and the step where gxwf caught a fabrication (below).

Source artifacts (UC3 interview→galaxy run, _emulated-runs/interview-atac-uc3/):

Tiers trimmed for legibility; … marks elided _plan_* prose and tool_state keys.

Tier 1 — Deferred (reconstructed)

DESeq2 differential test:
  tool_id: TODO                 # wrapper not yet resolved
  tool_version: TODO
  in:
    TODO_counts: ATAC counts
    TODO_factor_table: ATAC sample metadata
  out: [{id: TODO_result}, {id: TODO_normalized}, {id: TODO_plots}]
  _plan_state:   "factor=condition (2 levels) from metadata; reduce whole counts list;
                  emit results+normalized+plots; downstream assumes c3=log2FC, c7=padj"
  _plan_context: "identity from IWC rnaseq-de exemplar (Medium); UC3 gives one counts
                  collection + sample_metadata, not the exemplar 2-collection idiom"
  _plan_in:      "TODO_counts→countsFile/select_data; TODO_factor_table→metadata TSV"
  _plan_out:     "TODO_result→clean/filter/volcano; normalized+plots promoted"

Topology concrete (edges below); wrapper, version, ports all TODO; full planning intent carried.

Tier 2 — Identity-pinned (real: galaxy-workflow-draft.gxwf.yml)

DESeq2 differential test:
  tool_id: toolshed.g2.bx.psu.edu/repos/iuc/deseq2/deseq2   # identity pinned on exemplar evidence
  tool_version: TODO                                        # version still deferred
  in:
    TODO_counts: ATAC counts
    TODO_factor_table: ATAC sample metadata
  out: [{id: TODO_result}, {id: TODO_normalized}, {id: TODO_plots}]
  _plan_state:   "…"   # intent retained for the per-step loop to discharge
  _plan_context: "…"
  _plan_in:      "…"
  _plan_out:     "…"

tool_id now concrete; tool_version + ports still TODO; _plan_* retained. Pinned only because an exemplar named the wrapper — never on plausibility.

Tier 3 — Resolved (real: galaxy-workflow.gxwf.yml loop endstate)

DESeq2 differential test:
  tool_id: toolshed.g2.bx.psu.edu/repos/iuc/deseq2/deseq2/2.11.40.8+galaxy2
  tool_version: 2.11.40.8+galaxy2
  tool_shed_repository: {changeset_revision: 6c363db8c702, name: deseq2, owner: iuc,
                         tool_shed: toolshed.g2.bx.psu.edu}
  in:
    select_data|countsFile:   {source: ATAC counts}            # real wrapper ports
    select_data|sample_sheet: {source: ATAC sample metadata}
  out: [{id: deseq_out}, {id: counts_out}, {id: plots}]        # real output names
  tool_state:
    select_data:
      how: sample_sheet_contrasts
      design_formula_mode:
        factor: ["2"]          # ◄ COLUMN INDEX — not "condition". gxwf rejected the name.
        reference_level: control
    output_options: {output_selector: [pdf, normCounts]}
    …                           # full tool_state bound
  # no _plan_* — planning intent discharged

Wrapper version + changeset pinned, real ports bound, tool_state complete, planning fields gone.

Fixed topology (constant across all three tiers)

The edge graph for this step never changes — only the wrapper tier resolves:

[ATAC counts] ─────────────┐
                           ▼
[ATAC sample metadata] ─► DESeq2 differential test ─► result ─► Clean DESeq2 table ─► …
                           ├─► normalized counts  (promoted output)
                           └─► diagnostic plots   (promoted output)

Promotion + validation (the endstate)

Tier 3 has no TODO/_plan_* anywhere, so the draft promotes whole: class: GalaxyWorkflowDraft → class: GalaxyWorkflow, no transformation. gxwf validate on the promoted workflow:

The schema-catch (annotation on Tier 3): the first hand-authored Tier-3 tool_state set design_formula_mode.factor: ["condition"] — the factor name, which a fluent model writes by default. gxwf validate rejected it: the field is a column index, "2". Corrected → validates. This is “under-determination as a typed state” closing the loop — the deferred step resolves, and the schema, not a prose caveat, catches the fabrication that resolution introduced.

Panel layout sketch (for the designer)

Three columns left→right (Deferred · Identity-pinned · Resolved), the fixed topology strip drawn once across the top spanning all three to show it does not change. Within each column, the DESeq2 step block; highlight what hardens between columns:

        ┌───────────── fixed topology (counts + metadata → DESeq2 → result/…) ─────────────┐
        │                                                                                  │
  ┌─────────────┐            ┌─────────────┐               ┌─────────────────────────┐
  │  DEFERRED   │  ──────►   │ IDENTITY-   │   ──────►      │       RESOLVED          │
  │ tool_id:TODO│            │  PINNED     │               │ deseq2/…/2.11.40.8+gal2 │
  │ ports: TODO │            │ tool_id ✓   │               │ real ports + tool_state │
  │ _plan_* ████│            │ ver: TODO   │               │ factor:["2"] ◄ schema   │
  │             │            │ _plan_* ██  │               │ no _plan_*              │
  └─────────────┘            └─────────────┘               └───────────┬─────────────┘
   evidence: none yet         exemplar pins id              promote ▼ class: GalaxyWorkflow
                                                            gxwf validate ✓ (5 ok / 2 cache-skip)

Visual encoding: _plan_* shrinks column-to-column (intent discharged); TODO tokens turn concrete; the green gxwf ✓ and the factor:["2"] catch anchor the right edge.

Honesty notes