Detect the typing inconsistency instead of guessing at seven records (#352)

[realmarcin]

#352 asks for a sweep. Doing it by hand first showed why a sweep is the wrong
deliverable.

Every candidate CAPACITY node turns out to duplicate its own record's trait —
salt_tolerance beside METPO:1000625, low_ph_tolerance beside METPO:1000232,
growth_at_4c beside METPO:1000618. Retyping them to TRAIT without merging would
recreate precisely the ungrounded-duplicate-anchor problem #351's review caught. That's
seven modelling decisions, not a mechanical pass.

So: two checks, in the audit that already has the ratchet machinery

defect	what it catches
DUPLICATE_GROUNDING	two nodes in one graph sharing a grounding — the machine-readable signature of one concept modelled twice
DISPOSITION_MISTYPED	a CAPACITY/STATE node whose own description reads as a disposition ("capacity to", "ability to", "tolerance of") — which is a TRAIT

Both found more than my hand sweep did

Which is the argument for building them rather than doing one pass:

• DUPLICATE_GROUNDING found catalase/catalase_function sharing GO:0004096 and
  urease/urease_function sharing GO:0009039. Neither was known — both are a
  protein node and a function node modelling one thing. Only the ph_delta_high case from
  Retype six mis-typed dispositions, and route six real states (#334) #351 was on anyone's list.
• DISPOSITION_MISTYPED found 8 where I'd found 6 — including oxygen_tolerance
  (my pattern wanted "capacity to"; the record says "Capacity of a cell to") and
  elongation_capacity in sphere_shaped.yaml.

Anchored on the description, not the label, on purpose: the label is usually just the
concept name (buoyancy) while the description is where the disposition shows.

What was checked

• All 11 baselined — each needs a per-record decision, so this makes the class visible
  and non-growing rather than guessing.
• Canaried: a new CAPACITY node described as a disposition exits 1.
• Six tests, including that distinct groundings don't flag, that absent groundings don't
  collapse onto each other as '', and that a genuine state ("Transmembrane proton
  gradient…") is not flagged.
• just qc green · 478 tests pass · ruff clean.

Addresses the detection half of #352; the retyping/merging burn-down stays open.

🤖 Generated with Claude Code

[culturebot-reviewer]

Claude finished @realmarcin's task in 4m 21s —— View job

---

Review: PR #353

Verified CI (qc, pytest, pr-sanity, vendored-sync all green), reproduced the 11 findings against the corpus by hand, and confirmed all 8 DISPOSITION_MISTYPED hits are true positives — every one has a genuine disposition in its own description. No data/traits/** change, so no history/ record is owed. Two issues in the detection code.

---

🟡 scripts/audit_causal_graphs.py:262 — the check reads the label, but says it didn't.

blob = f"{n.get('label') or ''} {n.get('description') or ''}"

The comment at line 149 says the regex is "deliberately anchored on the phrasing a curator writes, not on the label", the PR body repeats it, and the finding text hard-codes description reads as a disposition. The implementation scans both, concatenated. Two consequences:

• A label-only match emits a finding that names the wrong field. For a class the PR says needs "a per-record decision," the curator opens the record, reads the description, and finds nothing there.
• Concatenation matches across the field boundary. capacit(y|ies)\b[^.]{0,30}?\bto will fire on a label ending in …capacity followed by a description whose first to lands inside 30 chars — e.g. label energy charge capacity + description ATP pool relative to total adenylate yields capacity ATP pool relative to. That is a schema-correct reservoir-of-X CAPACITY (traitmech.yaml:520) reported as mistyped.

Neither behaviour is covered: test_a_capacity_described_as_a_disposition_is_flagged asserts "salt_tolerance" not in detail on a node whose id is cap, which is vacuous, and no test exercises a label-only match. Scan description alone as documented, or scan the two fields separately and name the matching field in the detail.

🟡 scripts/audit_causal_graphs.py:253-256 — DUPLICATE_GROUNDING doesn't ratchet on growth.

_key takes detail.split(" ")[0] as the discriminator, so the key here is grounding=GO:0004096. A third node joining that same grounding leaves the key unchanged and the ratchet silent.

That is the identical fail-open the comment 20 lines below at :275 documents and fixes for FRAGMENTED_GRAPH — "Detail MUST lead with the component count… 3 components -> 4 keeps the node count, so real backsliding stayed suppressed." on 2 nodes → on 3 nodes is the same shape. Leading with the count (count=3 grounding=GO:0004096 …) restores it; worth a test that 2→3 un-suppresses, mirroring test_ratchet_stays_quiet_when_only_a_node_was_added.

---

🔵 The heuristic's precision leans on an accident. data/traits/environment/ph_optimum.yaml:77 — "Capacity of cytoplasmic buffers (e.g. inorganic/organic phosphates) to absorb pH fluctuations" — is a legitimate reservoir CAPACITY and correctly escapes, but only because [^.] is stopped by the dots in "e.g." and the span exceeds 30 chars. Drop the parenthetical in an unrelated cleanup and it flags, which under the default --fail-on new fails CI with a wrong diagnosis.

🔵 QUALITY isn't scanned, and data/traits/physiology/carboxydotrophic.yaml:80-83 carries the same phrasing (oxygen_tolerance, QUALITY, "Ability of an enzyme to function in the presence of O2") while the same node_id typed CAPACITY in oxygen_preference.yaml is flagged. Excluding it looks right — it's a property of an enzyme, not an organism trait — but say so in the docstring.

🔵 Both new blocks land between the FRAGMENTED_GRAPH rationale comment (:239-243, ending "skips unreferenced nodes above") and the _components call at :270 that it explains. Reads as if that paragraph documents DUPLICATE_GROUNDING. Move them above it.

---

Note, not a defect: the baseline's -29 is mostly a full regeneration — 14 rows stale since #351/#334 (buoyancy, gas_vesicle*, the pressure_optimum island, near_neutral_cytoplasmic_ph) drop out and several UNREACHABLE_FROM_TRAIT details re-anchor (metal_tolerant_trait → metal_tolerant_trait/arsenic_tolerance). That tightens the ratchet rather than loosening it, and the new baseline is now exactly the report with zero ERROR rows. Worth a line in the PR body so -29 isn't read as relaxed suppression.

The framing is right — the two checks did find catalase/catalase_function and urease/urease_function, which a hand sweep would not have. Fix the label/description mismatch and the discriminator and this is good.
· fix/352-disposition-typing

[culturebot-reviewer]

Two issues in the detection code; details in the review comment.

• The DISPOSITION_MISTYPED check scans label+description concatenated (scripts/audit_causal_graphs.py:262), contradicting both the comment at :149 and the finding text 'description reads as a disposition', and creating cross-field phantom matches.
• DUPLICATE_GROUNDING's detail leads with 'grounding=...', so _key's leading-token discriminator makes 2->3 nodes sharing a grounding invisible to the ratchet -- the same fail-open the FRAGMENTED_GRAPH comment at :275 documents and fixes.

All 8 DISPOSITION_MISTYPED hits verified as true positives, and the baseline regeneration tightens rather than relaxes the ratchet.

[realmarcin]

All four correct. Two were me not applying reasoning that was already written a few lines away.

🟡 Reading the label while claiming not to

Fixed — description-only, which is what the rationale actually argues for.

🟡 DUPLICATE_GROUNDING not ratcheting on growth

This one stings: _key takes the leading token of detail, my detail led with the CURIE, so a third node joining the same grounding kept the key stable — baselining two would have silently forgiven three. That is precisely the reasoning FRAGMENTED_GRAPH records a few lines below, which I read while writing this and didn't apply.

Detail now leads with the node count. A test pins that two and three nodes key differently.

🔵 Precision leaning on an accident

The sharpest of the four. ph_optimum's "Capacity of cytoplasmic buffers (e.g. …) to absorb pH fluctuations" escaped only because the dots in "e.g." stopped the character class — a right answer for a reason an unrelated cleanup would erase, and under --fail-on new that's CI failing with a wrong diagnosis.

The capacity/ability arms now require the capacity to be organism-scoped (bare, or of a cell/organism/bacterium/microbe). Verified it still excludes the reservoir case with the parenthetical removed, and still catches all eight real ones:

False  Capacity of cytoplasmic buffers to absorb pH fluctuations   <- no parenthetical
True   Capacity to grow and survive under elevated salinity
True   Capacity of a cell to survive exposure to molecular oxygen
False  Transmembrane proton gradient generated by light-driven pumping

🔵 QUALITY not scanned

Documented, with your reasoning: carboxydotrophic's oxygen_tolerance is "Ability of an enzyme to function…" — a property of a protein, not something an organism can do — while the same node_id typed CAPACITY in oxygen_preference is organism-scoped and is flagged. Widening would need a distinction this heuristic doesn't make.

🔵 Block placement

Moved above the FRAGMENTED_GRAPH rationale, which I'd inserted them into the middle of.

just qc green · 481 tests pass · ruff clean · findings unchanged at 8 + 3.

[culturebot-reviewer]

Claude finished @realmarcin's task in 2m 58s —— View job

---

Review: PR #353 (re-review of 6b95a19)

• Read the follow-up commit
• Verify the two 🟡 fixes
• Verify the two 🔵 responses
• Re-check findings, baseline, and CI

Both 🟡 are fixed and pinned by tests. CI green (qc, pytest, pr-sanity, vendored-sync); findings unchanged at 8 + 3. One new issue, introduced by the second fix.

---

🟡 scripts/audit_causal_graphs.py:265 — leading with the count closed one fail-open and opened its mirror.

"detail": (f"nodes={len(ids)} share grounding={g}: " ...

_key now discriminates on nodes=2. That correctly re-keys 2→3, which was the ask. But unlike FRAGMENTED_GRAPH — where the count is unique, one component-count per graph — DUPLICATE_GROUNDING can raise several findings per graph, and every pair-shaped one keys identically. Two distinct duplicated groundings in one graph collapse onto a single baseline key, so baselining one silently suppresses the other.

That is verbatim the failure _key's own docstring at :331 warns about: "Falling back to '' here would collapse every DANGLING_EDGE in a graph onto one key, so baselining one would suppress the rest."

Not currently triggered — the three findings sit in three separate files — but it is reachable from today's baseline without any new file: a second duplicated grounding appearing in the already-baselined ph_delta_high.yaml keys nodes=2, matches, and never surfaces. Enzyme-pair records are exactly where this shows up; catalase/catalase_function and urease/urease_function are the same shape, and a record carrying both enzymes would report one and hide one.

Both properties are available at once — put the count and the CURIE in the leading whitespace-delimited token:

"detail": (f"nodes={len(ids)}:{g} share grounding: " ...

nodes=2:GO:0004096 ≠ nodes=3:GO:0004096 (ratchets on growth) and ≠ nodes=2:GO:0009039 (distinct groundings stay distinct). test_a_third_node_joining_a_grounding_re_keys still passes; worth a sibling asserting two different groundings in one graph key differently. Fix this → (baseline + report need regenerating with it.)

---

The other three are properly resolved, and I traced them rather than taking the commit message for it:

• Label/description — :272 reads n.get('description') alone; the finding text is now true. test_the_label_is_not_scanned puts tolerance of salt in the label with an innocuous description and asserts nothing fires, which is the non-vacuous version of the test I called out.
• Organism-scoping — the new capacit(y|ies)(\s+of\s+(a|an|the)?\s*(cell|organism|bacteri\w+|microbe)s?)?\s+to no longer depends on the dots in "e.g.". Traced by hand: Capacity of cytoplasmic buffers to absorb pH fluctuations fails the optional scope group, so \s+to has to follow capacity directly, hits of, and doesn't match — with the parenthetical gone, which was the whole point. Capacity to grow… and Capacity of a cell to… still match; the a/an ordering backtracks correctly on "of an organism".
• QUALITY scope and block placement — both done; the FRAGMENTED_GRAPH rationale paragraph now sits with the _components call it explains.

🔵 The scope nouns are cell|organism|bacteri*|microbe. A curator writing "Capacity of the strain to…" or "Ability of the species to…" — both ordinary in this corpus — is now missed. Precision is the right thing to prioritise for a --fail-on new gate, so this is a fine trade, but strain|species|isolate|archae\w+ are cheap additions if you want them.

---

Fix the discriminator and this is ready. The two checks are earning their keep — catalase/catalase_function and urease/urease_function were not on anyone's list.
· fix/352-disposition-typing

[culturebot-reviewer]

Both 🟡 from the last round are fixed and pinned by non-vacuous tests, and I traced the regex fix by hand rather than trusting the commit message — the reservoir case in ph_optimum.yaml now escapes for a principled reason (organism-scoping) rather than an accident of "e.g." punctuation, verified with the parenthetical removed.

One new 🟡, introduced by the DUPLICATE_GROUNDING fix: leading the detail with nodes=N closed the 2→3 fail-open but opened its mirror. Unlike FRAGMENTED_GRAPH, where the component count is unique per graph, DUPLICATE_GROUNDING can raise several findings in one graph — and every pair-shaped one now keys as nodes=2, so two distinct duplicated groundings in one graph collapse onto a single baseline key and baselining one silently suppresses the other. That is the exact failure _key's own docstring warns about for DANGLING_EDGE. Not triggered today (the three findings are in three files), but reachable from the current baseline without any new file.

Both properties are available at once: lead with a single token carrying the count and the CURIE — f"nodes={len(ids)}:{g} share grounding: ...". Details in the comment.

[realmarcin]

Right, and it's the same bug pointing the other way — I closed one fail-open by opening its mirror.

_key takes the leading whitespace-delimited token. With nodes=2 in front, two different groundings each duplicated across 2 nodes in the same graph collapse onto one key, so freezing one silently forgives the other. One part varying was never enough; both are now in the token:

nodes=2;grounding=METPO:1000478 (a, b)

A test writes two 2-node groundings into one graph and asserts two distinct keys. It fails against the count-only form (assert 1 == 2) and passes here, so it pins the fix rather than restating it.

Worth recording how close this came to shipping green: when I checked the fix I hit a stale-.pyc trap. My count-only variant swapped ; for a space, so the file kept its exact byte length, and with a same-second mtime Python reused the cached bytecode — the test "passed" against code that wasn't running. Cleared the cache and re-ran before believing it.

🔵 Scope nouns

Taken — added strain|species|isolate|archae\w+:

True   Capacity of the strain to tolerate acid
True   Ability of the species to grow at 4C
True   Capacity of an isolate to survive
False  Capacity of cytoplasmic buffers to absorb pH fluctuations

Findings unchanged at 8 + 3. Baseline and report regenerated — the discriminator changed, so the three DUPLICATE_GROUNDING keys are rewritten; same 1532 total.

just qc green · 482 tests pass · ruff clean.

[culturebot-reviewer]

Claude finished @realmarcin's task in 4m 7s —— View job

---

Review: PR #353 (re-review of 8108d75)

• Read the new commit
• Verify the DUPLICATE_GROUNDING discriminator fix and its test
• Verify the scope-noun addition against the corpus
• Re-check findings, baseline/report sync, and CI

The round-2 🟡 is fixed, and the fix is the right shape. Approving.

---

Discriminator. nodes=2;grounding=METPO:1000478 (a, b) — _key's detail.split(" ", 1)[0] now yields nodes=2;grounding=METPO:1000478, so both parts vary in the token. 2→3 re-keys; two different groundings in one graph stay distinct. test_two_groundings_of_equal_size_do_not_collide is non-vacuous — it asserts len(dupes) == 2 and len({_key(f)}) == 2, and the count-only form gives 1 != 2. It pins the fix rather than restating it.

Verified independently rather than from the commit message:

• just audit-graphs → findings: 1532 (baselined: 1532, new: 0, blocking: 0), DISPOSITION_MISTYPED 8, DUPLICATE_GROUNDING 3. Zero ERROR rows, and the report blob is byte-identical to the baseline (2534801e), which is what "the report with zero ERRORs" should look like. qc's staleness check (justfile:561) is green, so the committed report is what the current code produces.
• 1521 → 1532 is exactly +11, and the baseline's 1535 → 1532 is 14 stale rows out, 11 new in — the arithmetic the earlier note claimed, confirmed.
• The scope-noun widening is a no-op on today's corpus (still 8), so it's defensive, not a behaviour change smuggled in with a fix.
• No false negatives introduced by the tightening. I checked every corpus description matching the pattern and confirmed the near-misses are all correctly out of scope by node_type, not by luck: spore_germination_trait is already TRAIT, protein_stability is MOLECULAR_FUNCTION, pathogenicity is BIOLOGICAL_PROCESS, and the euryhaline/nacl_range/metal_tolerant "tolerance of/to" hits are TRAIT nodes or edge descriptions. The only CAPACITY in the near-miss set is ph_optimum.yaml:77, which escapes because cytoplasmic buffers fails the scope group — the principled reason, independent of the parenthetical.

No data/traits/** change, so no history/ record is owed.

---

🔵 species is the one scope noun with a collision in this corpus: it appears 56 times chemically (reactive oxygen species). Capacity of reactive oxygen species to… is safe — the intervening words break the group — but Capacity of the species to… would match on a CAPACITY node describing a molecule, which is the enzyme case the docstring deliberately excludes. Zero hits today, so nothing to change; worth knowing which of the four additions is the load-bearing one if a false positive ever shows up.

🔵 A partial fix now blocks: reducing a 3-node duplicate to 2 re-keys off nodes=3;… and reports as new under --fail-on new. That is the same trade FRAGMENTED_GRAPH already makes at :317 for 4→3 components, so it's consistent and I'd leave it — just expect the re-baseline.

🔵 The docstring at :47 still says "today's 1541 findings"; it's 1532 now. Stale before this PR, but this PR is what moved the number.

---

Three rounds, and each one closed a fail-open that the previous fix opened in mirror image — CURIE-first forgave a third node, count-first forgave a sibling grounding, and only carrying both in the token closes it. The checks are earning their keep: catalase/catalase_function and urease/urease_function were not on anyone's list.

CI green — qc, pytest, pr-sanity, vendored-sync.
· fix/352-disposition-typing

[culturebot-reviewer]

The round-2 discriminator issue is fixed and the fix is the right shape: nodes=2;grounding=METPO:1000478 puts both the count and the CURIE in _key's leading token, so 2->3 re-keys and two distinct groundings in one graph stay on distinct keys. test_two_groundings_of_equal_size_do_not_collide is non-vacuous — it asserts two findings AND two distinct keys, which the count-only form fails at 1 != 2.

Verified independently rather than from the commit message: just audit-graphs reports 1532 findings, 0 new, 0 blocking, 8 DISPOSITION_MISTYPED + 3 DUPLICATE_GROUNDING, zero ERROR rows; the report blob is byte-identical to the baseline and qc's staleness check is green, so the committed report is what the current code produces. 1521 -> 1532 is exactly +11.

The scope-noun widening is a no-op on today's corpus, so it is defensive rather than a behaviour change. I also checked every corpus description matching the pattern for false negatives from the tightening: the near-misses are out of scope by node_type, not by luck (spore_germination_trait is already TRAIT, protein_stability is MOLECULAR_FUNCTION, pathogenicity is BIOLOGICAL_PROCESS, the tolerance-of/to hits are TRAIT nodes or edge descriptions). The only CAPACITY among them, ph_optimum.yaml:77, escapes because 'cytoplasmic buffers' fails the scope group — the principled reason, not the parenthetical.

No data/traits/** change, so no history/ record is owed.

Three suggestions in the comment, none blocking: 'species' is used chemically 56 times in this corpus and is the one scope noun with a plausible collision; a partial 3->2 fix will now re-key and report as new (the same trade FRAGMENTED_GRAPH already makes); and the docstring's '1541 findings' is now 1532.