State the PATHWAY vs BIOLOGICAL_PROCESS rule, then apply it (#356)

[realmarcin]

Third #356 tranche, and the first needing a rule rather than a lookup. Tranche 1 was decided by the schema naming the nodes as its own examples; tranche 2 by #352's already-settled "a protein is not its activity". This one had neither — PATHWAY is "a pathway or pathway-like mechanism" and BIOLOGICAL_PROCESS is "a biological process", which decides nothing.

The descriptions didn't decide it either, unlike tranche 2 where the typings visibly described different things. Here they describe the same thing in the same words — carotenoid_biosynthesis is "Enzymatic pathway producing carotenoid pigments" under BIOLOGICAL_PROCESS.

The grounding could not decide it, and that's the finding

The standing doctrine here is that the grounding settles typing arguments (#352, #360, #382). It can't, because GO has no pathway branch.

All ten families ground to a GO biological_process — GO:0019491, GO:0022900, GO:0009767, GO:0016117, GO:0006113, GO:0006119, GO:0006636, GO:0002047 — and five carry the identical CURIE under both typings. Even the METPO groundings sit under METPO:1000060 "metabolism", itself defined as "A biological process that maintain life in an organism".

Read literally, the grounding says all of these are processes and PATHWAY shouldn't exist.

But PATHWAY isn't vestigial — the same shape as #352's CAPACITY finding. Twelve CURIEs corpus-wide are typed PATHWAY and never BIOLOGICAL_PROCESS, and they're the canonical named routes: GO:0006099 (TCA cycle), GO:0006097 (glyoxylate cycle), GO:0019253 (Calvin cycle). The distinction is real; GO just doesn't draw it.

So: a stated convention, written into the playbook

type	test
PATHWAY	a named route whose steps you could enumerate — TCA cycle, ectoine biosynthesis (lysC/asd/ectB/ectA/ectC), the respiratory chain (nuo, cyo, ndh, sdh)
BIOLOGICAL_PROCESS	everything else — a strategy (salt_in_strategy), a reaction class (amino_acid_decarboxylation), or a class of routes (fermentation)

It goes in docs/CURATION_PLAYBOOK.md because an unwritten convention just re-splits.

The majority is not the rule. It types ectoine_biosynthesis and carotenoid_biosynthesis as PATHWAY against their majorities (4:2 and 5:1) — both are named biosynthetic routes with their enzymes listed in the corpus's own descriptions. A convention that only ratified the commoner typing wouldn't be one, and I'd rather you check those two than the eight that agreed.

INCONSISTENT_NODE_TYPE   199 -> 148     the 51 predicted, 0 new findings

Found and filed rather than fixed

fermentation in chemoorganoheterotrophic.yaml is grounded METPO:1000845 = Acetogenesis, a different concept (#391). Left untouched deliberately: retyping a node while carrying a wrong CURIE along unchanged would make it look reviewed. It also exposes a gap — one node_id with two different groundings across records is undetected, the grounding analogue of #356.

just qc green · 540 tests pass · ruff clean · history record per #325.

[culturebot-reviewer]

The rule is sound and the reasoning behind it is the most useful part of the PR — "GO has no pathway branch, so the grounding cannot settle this one" is worth having written down, and stating a convention rather than ratifying the majority is the right call. I checked the ten families: every one is now internally consistent, the majority counts quoted in each curation event are correct, the 51 removed baseline rows are exactly the sum of the ten families' occurrence counts, and the 3 added rows are re-keyed UNREACHABLE_FROM_TRAIT details (the type name is in the detail string), not new findings. METPO:1000845 is indeed acetogenesis — deferring that one to #391 rather than fixing it in passing is the right instinct.

Two things need changing.

🟡 The encodes edge in green_pigmented.yaml now violates its own predicate's declared range.

data/traits/morphology/green_pigmented.yaml:45 retypes phenazine_biosynthesis to PATHWAY. The edge at line 91 — phz_operons --encodes--> phenazine_biosynthesis, predicate_id: METPO:2007813 — is gated in mappings/predicate_grounding.tsv to object_types = GENE_OR_PROTEIN|BIOLOGICAL_PROCESS|ORGANELLE. PATHWAY is not in that set, so the corpus now carries a grounded edge the repo's own gate would have refused.

Nothing will catch it. ground_causal_predicates.py:162 returns early on any edge that already has a predicate_id, so the gate is only ever consulted at first grounding; audit_causal_graphs.py has no defect class for it, and audit_predicate_domains.py treats PATHWAY and BIOLOGICAL_PROCESS alike in ACTIVITY_NODE_TYPES, so it sees nothing either. This is the one edge in the corpus where the retyping crosses an asymmetric gate — I checked produces, encodes, reduces, powers and transports against every retyped node in both directions, and this is the only hit.

The fix is probably to add PATHWAY to that row's object_types: the v9 proposal note for METPO:2007813 already describes its objects as "a protein COMPLEX or a biosynthetic PROCESS", which is what this edge is. Retyping the node back would also work. Either way it should be a deliberate decision in this PR, since it is the same question #352 settled — whether the typing is compatible with what the record's predicates already assert.

(Related, not this PR's problem: white_pigmented.yaml:37 staphyloxanthin_biosynthesis is BIOLOGICAL_PROCESS and is also the object of an encodes edge. It is a named biosynthetic route by the new rule, so whoever applies the rule to it next will land on the same gate.)

🟡 The rule creates a new same-label split at ph_delta_mid3.yaml:81.

aa_decarboxylation there carries the label amino-acid decarboxylation pathways — byte-identical to ph_optimum_mid1.yaml:60, which this PR moves to BIOLOGICAL_PROCESS. Both were PATHWAY before; now they disagree. Its description, "Decarboxylation pathways that consume protons and store energy as PMF", is a reaction class under the rule you just wrote, so the intended type is clear.

The audit cannot see this because it keys on node_id and these two differ, and TARGET in migrate_pathway_vs_process.py is keyed the same way — so the tranche skipped it. Adding aa_decarboxylation to TARGET closes it. Worth noting that ph_growth_preference.yaml gets this right by accident: it holds both amino_acid_decarboxylase_systems (PATHWAY) and amino_acid_decarboxylation (BIOLOGICAL_PROCESS), which is exactly the two-node_ids case the playbook describes.

🔵 Two smaller notes, both optional.

The same node_id-keyed scope leaves other pre-existing contradictions with the new rule — metabolism/photosynthesis.yaml:64 electron_transport (BIOLOGICAL_PROCESS) against anoxygenic_photosynthesis.yaml:48 photosynthetic_etc (PATHWAY) under the same label, and pyoverdine_biosynthesis sitting as BIOLOGICAL_PROCESS next to the now-PATHWAY phenazine_biosynthesis inside one graph. Leaving them is defensible now that the rule is written down; a follow-up issue on the label axis would be the natural companion to the grounding-axis gap you filed as #391.

The per-file curation_history rationale quotes a description from a different record — acidotolerant.yaml quotes neutrophilic.yaml's wording, and halotolerant/nacl_delta_low/slightly_halophilic quote euryhaline's five-step description. All real corpus text, and the ectoine ones say "the corpus's own description" so they read correctly; the amino_acid_decarboxylation one has no such attribution and reads as if it were quoting the record it sits in.

History record under history/infrastructure/pathway-vs-process/ is present and the pages are regenerated consistently.

[culturebot-reviewer]

The rule is sound and the application checks out. I verified independently:

• All 11 families are now single-typed corpus-wide — no residual PATHWAY/BIOLOGICAL_PROCESS split on any of the retyped node_ids.
• The baseline diff is 51 INCONSISTENT_NODE_TYPE removals plus 4 pre-existing UNREACHABLE_FROM_TRAIT rows whose type= substring changed; each of the 4 has a matching deletion. Nothing new was baselined.
• The #391 note is accurate: data/embeddings/metpo_to_kgm_node.tsv gives METPO:1000845 = Acetogenesis, METPO:1002005 = Fermentation, METPO:1000060 = metabolism. Leaving it alone and filing it was the right call.
• Both gate widenings are real, not convenient. ph_optimum_high.yaml does have terminal_oxidases -produces-> proton_motive_force with proton_motive_force typed STATE, and phz_operons -encodes-> phenazine_biosynthesis is the only encodes edge this tranche pushed to PATHWAY. I checked the other grounded edges touching retyped nodes; the rest land on *-gated predicates or on has output, which already admits PATHWAY.
• Pages and reports regenerate consistently; the history record follows the tranche-1/2 shape.

Two things to fix, both in the quoting.

🟡 salinity_phenotype_with_numerical_limits.yaml:254 quotes a different record's description as this record's. The new curation_history reads "A STRATEGY, not a route: 'Osmoadaptation by intracellular accumulation of inorganic ions'." That string is haloalkaliphilic.yaml:65's node description (truncated before (e.g. K+)). This record's own description is "Osmoadaptation by accumulating high intracellular salt, mainly potassium." This is exactly the defect #394's attribution note was about; the note was applied to amino_acid_decarboxylation only, and salt_in_strategy has the same shape. Either attribute the quote or quote this record.

🟡 The #394 attribution fix landed in the script but not in the corpus — and names the wrong file. scripts/migrate_pathway_vs_process.py now says "The corpus describes it that way in acidotolerant.yaml -- 'Enzyme-catalyzed decarboxylation reaction that consumes cytoplasmic protons'". That string is neutrophilic.yaml:85's description. acidotolerant.yaml:63 says "Proton-consuming amino-acid decarboxylation (e.g. glutamate/Gad system) that consumes intracellular H+" — not the quoted wording. Separately, acidotolerant.yaml:257 and ph_optimum_mid1.yaml still carry the pre-#394 text with no attribution at all, so the shipped records read the way the reviewer objected to, and the script no longer reproduces the records it wrote. Fix the filename to neutrophilic.yaml and re-run so the data matches.

🔵 The carotenoid_biosynthesis rationale asserts "every description calls it one", and that sentence is written into red_pigmented.yaml, whose description is "Phytoene synthase condenses two GGPP to phytoene, then desaturation/isomerization yields lycopene" — the label says pathway, the description doesn't. Same quoting-precision class as the two above; optional.

Advisory: the history record predates the second commit, so the two predicate_grounding.tsv widenings and the aa_decarboxylation addition aren't in its details.