Migrate the 185 mechanical microbe-domain edges (#301, part 1 of 2)

[realmarcin]

First of the two migrations planned in proposals/metpo_traitmech_v9 (#326).

What moved

67 METPO object properties inherit rdfs:domain METPO:1000525 (microbe) via
METPO:2000001. OWL domain is an inference rule, not a constraint, so a causal-graph
edge using one entails its subject is a microbe — and CausalNodeTypeEnum has no
organism member. This migrates the 185 edges whose target was already decided or is
upstream:

Family	Edges	→	Source
any subject → TRAIT object	62	METPO:2007700 confers	v8
activity subject produces X	103	RO:0002234 has output	RO, upstream
TRAIT uses carbon source	12	METPO:2007806 has carbon source	v9
TRAIT uses energy source	8	METPO:2007807 has energy source	v9
	185

Every edge keeps its direction. This pass only relabels and re-grounds — that is what
keeps a 130-file change reviewable, and it's why the evidence-snippet baseline is
untouched (a reversal changes an edge's baseline key; nothing here does).

Family B takes an RO term rather than a new METPO one because biolink gives has output
domain: biological process or activity, which its BIOLOGICAL_PROCESS/PATHWAY/
MOLECULAR_FUNCTION subjects satisfy. Its mapping row is gated to exactly those three,
so a protein-subject edge is refused rather than silently re-acquiring the very domain
violation this removes.

The guard earned its keep

The script keys targets on (subject, object, predicate_id), not the endpoint pair.
Two corpus cases make the pair unsafe, and the count guard caught both before anything
was written:

• ecology/biosafety_level_1.yaml has low_pathogen_hazard --produces--> bsl1_trait
  (a target) and low_pathogen_hazard --qualifies for--> bsl1_trait
  (ungrounded, not a target) on the same endpoints. Keying by pair would have
  rewritten the ungrounded edge too — a silent corruption.
• physiology/chemoheterotrophic.yaml has both uses energy source and uses carbon source between the same two nodes, taking different targets, so one key cannot
  carry both.

Both were canaried after the fix: the ungrounded edge is untouched, and the two
same-endpoint edges get their correct distinct targets.

The proof it landed

MICROBE_DOMAIN_ON_NONORGANISM 366 → 181 — exactly the number the v9 round-trip plan
predicted, which is the point of having stated it.

What was checked

• just validate-strict — 477 files, 0 errors.
• just qc — green, every ratchet at blocking: 0.
• uv run pytest — 420 passed. Ruff clean.
• Evidence-snippet baseline unchanged at 2737, 0 new — the expected signature of a
  no-reversal pass.
• Gates verified against the data before writing them: has carbon source objects are
  12/12 CHEMICAL; has energy source is 2 CHEMICAL + 6 ENVIRONMENTAL_FACTOR, which is why
  its object_types is wider than its sibling's.
• Widened the confers row to admit CAPACITY — one family-A edge has a CAPACITY
  subject, which the v8 gate did not list and which would otherwise have been refused.

Scope

Does not touch the 181 edges needing the other 11 v9 terms (enzyme, transport, and the
rest of produces) — that's part 2, taking the baseline 181 → 0. The
METPO:20078xx ids remain placeholders until upstream mints them.

🤖 Generated with Claude Code

[culturebot-reviewer]

Claude finished @realmarcin's task in 6m 34s —— View job

---

Review

I verified the migration mechanically rather than taking the numbers on faith. The data pass is clean; one bookkeeping defect.

What I confirmed

• The 185 removed baseline rows are exactly METPO:2000202×165 + METPO:2000006×12 + METPO:2000010×8, with zero rows added — so nothing regressed into the ratchet. conf/predicate_domain_audit_baseline.tsv is 182 lines (181 + header), matching the claimed 366 → 181.
• Across data/traits/** the diff removes 500 lines: 370 predicate/predicate_id lines (= 185 edges × 2) and 130 file headers. Nothing else. No snippet, reference, notes, node, or description line is touched anywhere, which is the strongest possible form of the "evidence baseline unchanged" claim. 130 files, 130 MIGRATE_MICROBE_DOMAIN_EDGES events.
• All 20 family-D edges have subject_type=TRAIT, so the TRAIT subject gate is exact and not a guess; energy-source objects are organic_molecule, reduced_chemical_substrate, and light×6, which is what makes the wider CHEMICAL|ENVIRONMENTAL_FACTOR cell correct rather than sloppy.
• RO:0002234 is has output — correct term, correct label. METPO:2007806 / 2007807 match proposals/metpo_traitmech_v9/metpo_proposal_properties_robot.tsv on both id and label, and their 12/8 counts match the proposal's table exactly.
• The biosafety_level_1.yaml guard case is real and I checked it: low_pathogen_hazard --qualifies for--> bsl1_trait is at line 112, same endpoints as the migrated edge, still ungrounded, untouched. Triple-keying was the right call.
• confers is now 207 edges = 145 (v8) + 62. CAPACITY is a valid CausalNodeTypeEnum member.

---

🟡 The confers widening is recorded in one place and contradicted in two

mappings/predicate_grounding.tsv:102 gains CAPACITY in subject_types, but the note in that same row still reads:

subject_types lists the 9 node types actually observed across the 145 migrated edges

The cell now lists 10, and the corpus now has 207. A row that contradicts itself is worse than one that is merely out of date, because the note is the thing a future curator reads to decide whether widening it again is allowed.

The same claim is stale in the artifact that actually goes upstream — proposals/metpo_traitmech_v8/metpo_proposal_properties_robot.tsv:3:

covers 145 causal-graph edges across 114 trait records … Subject types observed: BIOLOGICAL_PROCESS 81, GENE_OR_PROTEIN 32, … STATE 1.

CAPACITY (morphology/cell_width_small.yaml#cell_width_small_mreb_setpoint, rod_diameter_control --> cell_width_small_trait) is absent from that enumeration. That enumeration is the domain evidence a METPO curator will use to decide what METPO:2007700 admits, and this PR is the change that made it wrong. The v9 proposal documents the 62-edge addition but never mentions CAPACITY, so nothing else in the repo carries it.

Fix this →

---

🔵 The residual report now describes a corpus state that was not written

reports/predicate_grounding_residual.tsv moves has output from 4 unmapped to 3 blocked_by_node_type. Three of those four are genuinely blocked. The fourth — data/traits/upper/biological_process.yaml:129, molecular_function --has output--> chemical_output, whose own description already cites RO:0002234 — now passes the new gate, which is why the count dropped. But no predicate_id was written for it: the diff adds exactly 370 predicate lines, all of them replacements. So the committed report says that edge is groundable while the corpus still leaves it bare. A just ground-predicates --apply closes the gap; worth doing here so the report and the data agree at the same commit.

🔵 No history/ record

history/records/ is unchanged for a 130-file data/traits/** change. Advisory per history/README.md, and the Curation history check passed — noting it, not blocking on it.

---

Everything else about this is right, including the parts that were easy to get wrong: the object_type == TRAIT test running before the produces test, has output gated to the three activity node types so part 2's protein subjects are refused rather than silently re-broken, and the count guard aborting on a partial match instead of writing one. CI is green across all 7 completed checks.

Verdict: request changes — for the self-contradicting confers row and the stale v8 proposal evidence. Both are small; the migration itself needs nothing.
· branch feat/301-migrate-mechanical-families

[culturebot-reviewer]

Migration itself verified clean: the 185 removed baseline rows are exactly METPO:2000202x165 + METPO:2000006x12 + METPO:2000010x8 with zero rows added; the data diff touches only the 370 predicate/predicate_id lines and nothing else (no snippet, reference or node line anywhere); all 20 family-D subjects really are TRAIT; RO:0002234 and the two v9 ids/labels all check out; and the biosafety_level_1.yaml guard case is real (line 112, same endpoints, still ungrounded and untouched).

Requesting changes for one bookkeeping defect. mappings/predicate_grounding.tsv:102 gains CAPACITY in subject_types while its own note still reads 'the 9 node types actually observed across the 145 migrated edges' - the cell now lists 10 and the corpus has 207. The same claim is stale in the artifact that goes upstream, proposals/metpo_traitmech_v8/metpo_proposal_properties_robot.tsv:3, whose 'Subject types observed' enumeration omits CAPACITY. That enumeration is the domain evidence a METPO curator uses to decide what METPO:2007700 admits, and this PR is what made it wrong.

Also worth folding in (non-blocking): reports/predicate_grounding_residual.tsv drops 'has output' from 4 to 3 because data/traits/upper/biological_process.yaml:129 now passes the new gate, but no predicate_id was written for it - so the committed report and the corpus disagree at the same commit. A 'just ground-predicates --apply' closes that. And there is no history/ record for a 130-file data/traits change (advisory only).

Full detail in #328 (comment)

[realmarcin]

Both fixed, and the 🔵 residual one was the more useful catch.

🟡 The self-contradicting confers row

Right, and the framing is right too — a row that contradicts itself is worse than one merely stale, because that note is what a curator reads to decide whether widening it again is allowed. It now reads 10 node types across 207 migrated edges, and records that CAPACITY entered via this PR.

The upstream artifact was the more important half. proposals/metpo_traitmech_v8's METPO:2007700 row said "covers 145 causal-graph edges across 114 trait records" with an enumeration omitting CAPACITY. Recomputed from the corpus: 207 edges / 172 records, subject types BIOLOGICAL_PROCESS 139, GENE_OR_PROTEIN 32, ENVIRONMENTAL_FACTOR 12, QUALITY 6, CHEMICAL 5, CELLULAR_LOCALIZATION 5, PATHWAY 4, MOLECULAR_FUNCTION 2, STATE 1, CAPACITY 1. The cell names #328 as the cause, so the v8 narrative stays historically readable rather than being retconned.

🔵 The residual report describing unwritten state — fixed

Good catch, and subtle: the new has output row made grounding newly applicable to upper/biological_process.yaml's molecular_function --has output--> chemical_output, whose own description already cited RO:0002234 but which carried no predicate_id. The committed report had started calling it groundable while the data left it bare.

Ran just ground-predicates --apply; report and corpus now agree at the same commit. The two remaining blocked_by_node_type rows are the gates working as intended — an EXPERIMENTAL_FACTOR subject for has output and a BIOLOGICAL_PROCESS object for confers are both correctly refused.

🔵 history/ record

Acknowledged, still not added — the convention gap and whether it should be enforced is tracked in #325, filed during the v8 migration for the same reason.

Re-verified

just qc green, just validate-strict 477 / 0 errors, uv run pytest 420 passed, reports and pages/ regenerated. Baseline still 181, unchanged by these fixes.

[culturebot-reviewer]

Claude finished @realmarcin's task in 4m 12s —— View job

---

Re-review of the fix commit (36e7720b)

Both requested changes landed, and I re-verified them against the corpus rather than the commit message.

The confers bookkeeping — fixed and correct, not just edited

• mappings/predicate_grounding.tsv:102 now says "the 10 node types actually observed across the 207 migrated edges (145 from Migrate 164 enables→TRAIT edges onto the v8 predicates (#302, #303) #323, plus 62 more in Migrate the 185 mechanical microbe-domain edges (#301, part 1 of 2) #328, which is where CAPACITY entered)". Cell and note agree, and the note now carries the provenance of the widening, which is the part a future curator actually needs.
• proposals/metpo_traitmech_v8/metpo_proposal_properties_robot.tsv:3 → 207 edges / 172 records. I counted independently: predicate_id: METPO:2007700 appears 207 times across 172 files under data/traits/**. Exact match.
• The new enumeration sums to 207, and its delta against the old one is BIOLOGICAL_PROCESS +58, CHEMICAL +2, QUALITY +1, CAPACITY +1 = 62, exactly the family-A count this PR claims. The counts are internally consistent with the migration, not just plausible.
• CAPACITY 1 is exact. Twelve trait files contain both a CAPACITY node and a confers edge; I checked the subject of every confers edge in each. Only morphology/cell_width_small.yaml's rod_diameter_control (line 58, node_type: CAPACITY) is one — every other file's confers subject is an ENVIRONMENTAL_FACTOR, BIOLOGICAL_PROCESS, or CHEMICAL node that happens to co-occur with an unrelated CAPACITY node. So the count is one, and the "1" is the right one.
• The row keeps naming Migrate the 185 mechanical microbe-domain edges (#301, part 1 of 2) #328 as the cause rather than rewriting v8's history — the right call for an artifact that goes upstream with a narrative attached.

The residual/corpus disagreement — closed

data/traits/upper/biological_process.yaml:136 now carries predicate_id: RO:0002234 on molecular_function --has output--> chemical_output, with a GROUND_CAUSAL_PREDICATES history entry. The gate holds on its own terms: the subject node is MOLECULAR_FUNCTION (line 36), one of the three activity types has output admits. reports/predicate_grounding_residual.tsv:89 still reads has output 3 blocked_by_node_type, but upper/biological_process.yaml has dropped out of its example list — the count is now 3 because three edges are genuinely blocked, not because one was silently counted as grounded. Report and corpus agree at this commit.

No collateral damage

Outside those three files the commit touches only pages/**, and every one of those is the regenerated Corpus as of … UTC timestamp line. conf/predicate_domain_audit_baseline.tsv is untouched, so the 181 baseline stands. All 7 completed checks green.

One factual note, not a defect: the PR comment says "the two remaining blocked_by_node_type rows", but the report has three such rows (causally upstream of 13, confers 3, has output 3) — the two named are edge examples, not the whole blocked set. Nothing committed is wrong; it just understates how much the gates are refusing.

The history/ record is still absent, still advisory, and #325 is the right place for it.

Verdict: approve. Both bookkeeping defects are properly fixed, and the fix for the second one changed data, so it was worth doing here rather than deferring.

---

Note: Python was not available to me in this run, so the count verification above was done with git, ripgrep, and file reads rather than by re-running the repo's own audit scripts. The 207/172 figure and the CAPACITY-is-1 finding are direct corpus counts; the nine other per-type subject counts I checked by arithmetic consistency and spot-check (neutrophilic, free_living, ph_delta, cell_width_small), not by exhaustive recount.
· branch feat/301-migrate-mechanical-families

[culturebot-reviewer]

Both requested changes verified against the corpus, not just the commit message.

The confers row now agrees with itself (10 types / 207 edges) and records that CAPACITY entered via #328. I counted independently: METPO:2007700 appears 207 times across 172 files under data/traits/**, an exact match for the v8 proposal's updated traits_addressed. The new subject-type enumeration sums to 207, and its delta against the old one (BIOLOGICAL_PROCESS +58, CHEMICAL +2, QUALITY +1, CAPACITY +1) is exactly the 62 family-A edges this PR migrated. CAPACITY 1 is exact: twelve files contain both a CAPACITY node and a confers edge, and only cell_width_small.yaml's rod_diameter_control is actually a confers subject.

The residual/corpus disagreement is closed - biological_process.yaml:136 now carries predicate_id RO:0002234, its subject node is MOLECULAR_FUNCTION so the new gate holds on its own terms, and upper/biological_process.yaml has dropped out of the report's has-output example list, so the 3 blocked edges are genuinely blocked rather than one being silently miscounted.

Outside those three files the commit touches only pages/**, all of it the regenerated timestamp line. The 181 baseline is untouched. All checks green.

Minor, non-blocking: the PR comment says 'the two remaining blocked_by_node_type rows' but the report has three (causally upstream of 13, confers 3, has output 3) - the two named are edge examples. Nothing committed is wrong.