Address Codex adversarial review #558 — three high-priority findings

Codex flagged three ways the recent subclass-plumbing work would
poison the merged-kg with semantically wrong relationships. All three
fixes ship together because they are interdependent (the loader
trust policy interacts with the placeholder fallback emit, and the
narrowMatch filter interacts with the get_parents() index).

Finding 1 [HIGH] — manual closeMatch rows promoted to canonical nodes
=====================================================================
File: kg_microbe/utils/isolation_source_mapping_utils.py
       mappings/validate_isolation_source_mappings.py

The loader's _row_is_trusted() accepted any row tagged
``semapv:ManualMappingCuration`` regardless of predicate. That admitted
41 manually-curated ``skos:closeMatch`` rows, including:
   * Catheter → NCIT:C50344 (Catheter Device)  — device, not source
   * Child → PATO:0001190 (juvenile)            — quality, not source
   * Humid → NCIT:C88206 (Humidity)             — quality, not source
   * Psychrophilic-<10°C → METPO:1000614        — phenotype class, not source
   * Boreal → ENVO:01000174 (forest biome)      — biome name mismatch

Tightened trust policy: substitution into the BacDive graph requires
``skos:exactMatch`` regardless of curator. closeMatch rows fall back
to placeholder isolation_source:* nodes. Two acceptable trust paths
within exactMatch: high-confidence auto-match OR manual curation.
Net effect: 207 → 158 trusted mappings; 49 closeMatch rows correctly
drop instead of poisoning the graph.

The standalone validator's _row_is_trusted() is updated to match
(test_validator_rules_match_loader enforces the parity).

Finding 2 [HIGH] — bad MIM narrowMatch rows generate false subclass edges
==========================================================================
File: scripts/consolidate_chemical_mappings.py

MIM's auto_classify_ingredient_type pipeline produced 5 narrowMatch
rows where the chemistry on both sides is unrelated:
   * MIM:Kh2po4 → CHEBI:32583 (KH2PO4 vs calcium sulfate dihydrate)
   * MIM:Mncl2_X_2_H2o → CHEBI:30200 (MnCl2 vs kaempferol glycoside)
   * MIM:Mncl2_X_4_H2o → CHEBI:30200
   * MIM:Mncl2_anhydrous → CHEBI:30200
   * MIM:D-Maltose_Monohydrate → CHEBI:233428 (maltose vs amiloride analog)

Without this filter, get_parents() exposed those rows to MediaDive's
new biolink:subclass_of emit path (commit f3a8199), which would
have made the maltose ingredient a subclass of an unrelated amiloride
analog in the merged-kg.

Added KNOWN_BAD_NARROWMATCH set in load_mediaingredientmech_sssom()
that drops these specific (subject_id, object_id) pairs at row-load
time. The filter is idempotent — when MIM upstream removes the rows
it becomes a no-op for us. Verified: regenerated unified file has
``cas:6363-53-7 parents []`` and the parallel cases for KH2PO4
and MnCl2 hydrates.

Finding 3 [MEDIUM] — blanket ENVO subclass_of for all isolation_source placeholders
====================================================================================
File: kg_microbe/transform_utils/bacdive/bacdive.py

The previous commit (959baa6) emitted
``isolation_source:* biolink:subclass_of ENVO:01000254`` for every
unmapped isolation_source placeholder. But the table intentionally
leaves labels like 'Human', 'Leaf-Phyllosphere', and
'host_animal_endotherm_intratissue' unmapped, and those are NOT
environmental materials — they're hosts / anatomy / niches. A blanket
ENVO parent would poison downstream reasoning over source type.

Removed the blanket subclass_of edge. Placeholders stay unparented
until a vetted host/anatomy/environment mapping lands in
mappings/isolation_source_to_ontology.tsv. The mediadive.solution →
CHEBI:60004, kgmicrobe.assay → MICRO:0000903, kgmicrobe.pathway →
GO:0008152 emits all stay (those are correct single-parent types).

Verified
========
* python mappings/validate_isolation_source_mappings.py → OK
* poetry run pytest tests/test_isolation_source_mapping_utils.py
  tests/test_chemical_mapping_utils.py
  tests/test_consolidate_chemical_mappings.py
  tests/test_metatraits.py → 110 passed
* Consolidator regenerates unified_ingredient_mappings.sssom.tsv.gz
  cleanly: 5 known-bad narrowMatch dropped at MIM load.
* test_loader_honors_manually_curated_fixes updated to match new
  policy (Plant→Viridiplantae was a closeMatch row that no longer
  qualifies; Mammals→Mammalia is exactMatch and still honored).

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Author: realmarcin

kg_microbe/transform_utils/bacdive/bacdive.py

@@ -2890,30 +2890,23 @@ def run(self, data_file: Union[Optional[Path], Optional[str]] = None, show_statu
                             subject_id = ISOLATION_SOURCE_PREFIX + isol_source.lower()
                             # Only write a placeholder node when no ontology mapping exists;
                             # mapped CURIEs get their canonical node from the ontologies transform.
+                            # Note: we deliberately do NOT add a blanket
+                            # ``biolink:subclass_of ENVO:01000254`` edge for these
+                            # placeholders. Many unmapped labels are NOT environmental
+                            # materials (e.g. 'Human', 'Leaf-Phyllosphere',
+                            # 'host_animal_endotherm_intratissue' are hosts / anatomy /
+                            # niches), and a blanket parent assertion would poison the
+                            # hierarchy for downstream reasoning over source type
+                            # (Codex review #558). Placeholders stay unparented until a
+                            # vetted host/anatomy/environment mapping is added to
+                            # mappings/isolation_source_to_ontology.tsv.
                             node_writer.writerow(
                                 self._create_node_row(
                                     subject_id,
                                     ISOLATION_SOURCE_CATEGORY,
                                     isol_source,
                                 )
                             )
-                            # Type the residual placeholder under ENVO:01000254
-                            # (environmental material) so OBO-aware reasoners can navigate
-                            # from any unmapped isolation source back to the canonical
-                            # ENVO hierarchy. Curated mappings (handled in the `if`
-                            # branch above) get their canonical parent from the ontologies
-                            # transform; only the kg-microbe-minted placeholders need this.
-                            edge_writer.writerow(
-                                [
-                                    subject_id,
-                                    "biolink:subclass_of",
-                                    "ENVO:01000254",
-                                    "rdfs:subClassOf",
-                                    self.source_name,
-                                    "knowledge_assertion",
-                                    "manual_agent",
-                                ]
-                            )
                         # Write edge from the isolation source to organism
                         knowledge_level, agent_type = self._add_edge_metadata(
                             NCBI_TO_ISOLATION_SOURCE_EDGE, LOCATION_OF, organism_id

kg_microbe/utils/isolation_source_mapping_utils.py

@@ -102,16 +102,38 @@ def _row_passes_family_check(row: Dict[str, str]) -> bool:
 
 
 def _row_is_trusted(row: Dict[str, str]) -> bool:
-    """Apply the trust policy described in the module docstring."""
+    """
+    Apply the trust policy described in the module docstring.
+
+    Substitution into the BacDive graph requires ``skos:exactMatch`` —
+    i.e. the curator (or the auto-matcher's high-confidence pass) has
+    asserted that the BacDive label and the ontology term denote the
+    SAME entity. ``skos:closeMatch`` rows are NOT trusted for canonical
+    node substitution because they only assert similarity; promoting
+    them would connect organisms to devices, qualities, phenotype
+    classes, etc., as if those were the source entity itself
+    (Codex adversarial review #558 found 41 such bad-substitution
+    candidates in the table — Catheter→NCIT 'Catheter Device',
+    Child→PATO juvenile, Humid→NCIT humidity quality, etc.).
+
+    Two acceptable trust paths, both requiring exactMatch:
+      1. Auto-matcher hit with high confidence
+         (``skos:exactMatch`` + ``confidence == 'high'``).
+      2. Manual curation
+         (``skos:exactMatch`` + ``mapping_justification ==
+         'semapv:ManualMappingCuration'``).
+
+    Anything else — closeMatch under any justification, low/medium
+    auto-matcher confidence — is dropped, leaving the BacDive transform
+    to emit its placeholder ``isolation_source:*`` node.
+    """
     predicate = (row.get("predicate_id") or "").strip()
+    if predicate != "skos:exactMatch":
+        return False
+
     confidence = (row.get("confidence") or "").strip().lower()
     justification = (row.get("mapping_justification") or "").strip()
-
-    if predicate == "skos:exactMatch" and confidence == "high":
-        return True
-    if justification == "semapv:ManualMappingCuration":
-        return True
-    return False
+    return (confidence == "high") or (justification == "semapv:ManualMappingCuration")
 
 
 def load_isolation_source_mappings(

mappings/unified_ingredient_mappings.sssom.tsv.gz

@@ -60,21 +60,17 @@ def _row_is_trusted(row: Dict[str, str]) -> bool:
     """
     Mirror of the loader's trust policy.
 
-    The validator only fails on rows that *would* be loaded at runtime — a
-    low-trust ``ols4_auto`` ``skos:closeMatch`` row is already dropped by the
-    loader, so flagging it here would be noise. Promoting such a row by
-    setting it to ``skos:exactMatch`` / high confidence / manual curation is
-    what makes a family-mismatched mapping actually dangerous, and that is
-    the case the validator is designed to catch.
+    Both paths require ``skos:exactMatch``: closeMatch is never trusted for
+    canonical node substitution because it only asserts similarity, not
+    equivalence (see Codex review #558). Two trust paths within exactMatch:
+    high-confidence auto-match, or manual curation.
     """
     predicate = (row.get("predicate_id") or "").strip()
+    if predicate != "skos:exactMatch":
+        return False
     confidence = (row.get("confidence") or "").strip().lower()
     justification = (row.get("mapping_justification") or "").strip()
-    if predicate == "skos:exactMatch" and confidence == "high":
-        return True
-    if justification == "semapv:ManualMappingCuration":
-        return True
-    return False
+    return (confidence == "high") or (justification == "semapv:ManualMappingCuration")
 
 
 def iter_validation_failures(

mappings/validate_isolation_source_mappings.py

@@ -1198,8 +1198,27 @@ def load_mediaingredientmech_sssom(self, filepath: Path):
         # seeding carried forward obsolete MIM xrefs).
         current_mim_subjects: dict[str, str] = {}
 
+        # Known-bad (subject_id, object_id) narrowMatch pairs from MIM's
+        # auto_classify_ingredient_type pipeline. These were flagged in the
+        # 2026-05-02 Codex adversarial review of PR #558 — the chemistry on
+        # both sides is unrelated (e.g. KH2PO4 → calcium sulfate dihydrate;
+        # MnCl2 → kaempferol glycoside; D-Maltose → amiloride analog).
+        # Without this filter the runtime emits demonstrably false
+        # ``biolink:subclass_of`` edges via MediaDive's get_parents() path.
+        # Drop these specific rows here; the proper upstream fix is to
+        # remove them in MIM (sibling repo). The filter is idempotent — a
+        # later MIM update that removes the rows is a no-op for us.
+        KNOWN_BAD_NARROWMATCH = {
+            ("MIM:Kh2po4", "CHEBI:32583"),                # KH2PO4 vs CaSO4·2H2O
+            ("MIM:Mncl2_X_2_H2o", "CHEBI:30200"),         # MnCl2·2H2O vs kaempferol glycoside
+            ("MIM:Mncl2_X_4_H2o", "CHEBI:30200"),         # MnCl2·4H2O vs kaempferol glycoside
+            ("MIM:Mncl2_anhydrous", "CHEBI:30200"),       # MnCl2 vs kaempferol glycoside
+            ("MIM:D-Maltose_Monohydrate", "CHEBI:233428"),  # maltose vs amiloride analog
+        }
+
         added = 0
         skipped_unsupported = 0
+        skipped_known_bad = 0
         curator_tags_seen: set[str] = set()
         for _, row in df.iterrows():
             object_id = row.get("object_id", "").strip()
@@ -1219,6 +1238,17 @@ def load_mediaingredientmech_sssom(self, filepath: Path):
 
             predicate = row.get("predicate_id", "").strip()
             subject_id = row.get("subject_id", "").strip()
+
+            # Drop the specific narrowMatch rows MIM's auto-classifier got
+            # wrong (see KNOWN_BAD_NARROWMATCH definition above). Done at
+            # row-load time rather than at export so the rows never enter
+            # the parent_relations capture in the first place.
+            if (
+                predicate in {"skos:narrowMatch", "skos:broadMatch"}
+                and (subject_id, object_id) in KNOWN_BAD_NARROWMATCH
+            ):
+                skipped_known_bad += 1
+                continue
             subject_label = row.get("subject_label", "").strip()
             object_label = row.get("object_label", "").strip()
             other = row.get("other", "").strip()
@@ -1326,7 +1356,8 @@ def load_mediaingredientmech_sssom(self, filepath: Path):
 
         print(
             f"  Loaded {added} MIM SSSOM entries "
-            f"(skipped {skipped_unsupported} unsupported object_id prefix)"
+            f"(skipped {skipped_unsupported} unsupported object_id prefix"
+            f", {skipped_known_bad} known-bad narrowMatch)"
         )
         if curator_tags_seen:
             print(

scripts/consolidate_chemical_mappings.py

@@ -61,11 +61,22 @@ def test_loader_drops_family_mismatched_rows(mappings):
 
 
 def test_loader_honors_manually_curated_fixes(mappings):
-    """Rows promoted by manual curation (family_mismatch_fix and other curator tags) are honored."""
-    # 'Plant' was wrongly mapped to ENVO:01001255 (a process); manual fix points at Viridiplantae taxon
-    assert mappings.get("plant") == ("NCBITaxon:33090", "Viridiplantae")
-    # 'Mammals' was wrongly mapped to NCBITaxon:32525 (Theria, a subgroup); manual fix points at Mammalia
+    """Rows promoted by manual curation are honored — but ONLY when predicate is skos:exactMatch.
+
+    The 2026-05 Codex adversarial review tightened the trust policy so that
+    ``skos:closeMatch`` rows are no longer trusted for canonical node
+    substitution, even when manually curated, because closeMatch only
+    asserts similarity (not equivalence) and would connect organisms to
+    devices, qualities, and phenotype classes as if those were the source
+    entity itself.
+    """
+    # 'Mammals' was wrongly mapped to NCBITaxon:32525 (Theria, a subgroup); manual fix points
+    # at Mammalia with skos:exactMatch — honored under the new policy.
     assert mappings.get("mammals") == ("NCBITaxon:40674", "Mammalia")
+    # 'Plant' was promoted to NCBITaxon:33090 (Viridiplantae) via skos:closeMatch — NOT honored
+    # under the tightened policy (manual curation alone isn't enough; needs exactMatch).
+    # The BacDive transform falls back to the isolation_source:plant placeholder.
+    assert mappings.get("plant") is None
 
 
 def test_loader_rejects_low_trust_lexical_close_matches(mappings):