diff --git a/kb/communities/Clostridium_Autoethanogenum_Kluyveri_Syngas_Coculture.yaml b/kb/communities/Clostridium_Autoethanogenum_Kluyveri_Syngas_Coculture.yaml
index 87b0937d..ab1d90f8 100644
--- a/kb/communities/Clostridium_Autoethanogenum_Kluyveri_Syngas_Coculture.yaml
+++ b/kb/communities/Clostridium_Autoethanogenum_Kluyveri_Syngas_Coculture.yaml
@@ -232,6 +232,53 @@ ecological_interactions:
     snippet: continuous removal of ethanol by C. kluyveri
     explanation: Supports the inferred thermodynamic pull mechanism caused by C. kluyveri ethanol
       consumption.
+cultivation_setup:
+- cultivation_mode: CHEMOSTAT
+  system_type: STIRRED_TANK_BIOREACTOR
+  operating_temperature: 37.0
+  operating_temperature_unit: °C
+  ph_controlled: true
+  temperature_controlled: true
+  instrument_detail: >-
+    Continuous stirred-tank reactors, CO/syngas-fed, chosen so pH, temperature and medium
+    composition could be held constant while growth rate was set by dilution. The coculture
+    was assembled inside the reactor rather than co-inoculated: C. autoethanogenum ran alone
+    on CO for 20 days, then C. kluyveri was added at 5% v/v under the same conditions.
+  controls_notes: >-
+    pH and temperature under active control from the control tower. Reactors were flushed
+    with N2 at 20 mL/min for 3 hours before inoculation to establish anaerobic conditions.
+  notes: >-
+    This is the coculture's own setup, not a member's: the source runs both organisms
+    together in the reactor and the chemostat data it reports are co-cultivation data. Worth
+    stating because the neighbouring #183 candidate,
+    Methylacidiphilum_Galdieria_Thermoacidophilic_Coculture, was rejected for the opposite
+    reason - its cited paper never mentions the second member at all, so its rich bioreactor
+    methods describe a monoculture (#529).
+
+    Dilution rate is not recorded here. The paper varies it across runs and reports the
+    conditions per run in Table 1, which the cached text carries only as a caption; a single
+    value would assert a steady state the paper does not claim for the whole study.
+  evidence:
+  - reference: PMID:31792266
+    supports: SUPPORT
+    evidence_source: IN_VITRO
+    snippet: Experiments were performed in continuous stirred-tank reactors (CSTR) to accurately
+      control environmental conditions (pH, temperature, stable medium composition) and microbial
+      growth rates
+    explanation: The vessel, and that pH and temperature were controlled rather than merely measured.
+  - reference: PMID:31792266
+    supports: SUPPORT
+    evidence_source: IN_VITRO
+    snippet: constant pH of 6.2 and 37 °C, for 20 days. After this period, the reactor was inoculated
+      (5% v/v) with a culture of C. kluyveri for the establishment of the synthetic co-culture
+    explanation: The operating temperature, and that the coculture was built up inside the reactor.
+  - reference: PMID:31792266
+    supports: SUPPORT
+    evidence_source: IN_VITRO
+    snippet: reactors were connected to the control tower, initiating temperature (37 °C) and pH
+      control
+    explanation: Sources ph_controlled and temperature_controlled.
+
 environmental_factors:
 - name: Gas substrate
   value: carbon monoxide or syngas
diff --git a/tests/test_cultivation_source_studied_the_community.py b/tests/test_cultivation_source_studied_the_community.py
new file mode 100644
index 00000000..7e3a71dd
--- /dev/null
+++ b/tests/test_cultivation_source_studied_the_community.py
@@ -0,0 +1,177 @@
+"""A cultivation_setup's source must be about *this* community (#529).
+
+`Methylacidiphilum_Galdieria_Thermoacidophilic_Coculture` was the strongest
+remaining #183 candidate by every mechanical measure: 47k characters of cached
+full text and a methods section with a LABFORS 3 bioreactor, 2 L working volume,
+45 °C, 450 rpm, fed-batch switched to chemostat. Its cited reference mentions
+*Galdieria* **zero times** and "coculture" zero times — it is a
+*Methylacidiphilum* monoculture study.
+
+Curating those numbers would have been wrong in the way hardest to catch later:
+every value individually correct, every snippet verbatim, `just validate` and
+`just validate-references` both clean, and the record asserting that a two-member
+coculture ran under conditions never applied to it.
+
+`Soil_Corrinoid_B12_Reservoir_Community` is the same shape — scored 5/5 on
+conditions belonging to an *E. coli* reporter strain, in a community that is
+metagenomic and was never cultured. Only reading the record caught it.
+
+What no existing gate can see: `validate-references` confirms a snippet is a
+substring of the cited text. Nothing confirms the cited text is *about this
+community*. This is the cheap discriminator — how many members the source names —
+and it is deliberately a **heuristic with an escape hatch**, not a proof. A paper
+can name every member and still report monoculture conditions, and a legitimate
+apparatus paper may name none. The point is that the ratio is looked at once, by
+someone, rather than never.
+"""
+
+from __future__ import annotations
+
+import pathlib
+import re
+
+import pytest
+import yaml
+
+REPO = pathlib.Path(__file__).parent.parent
+COMMUNITIES = REPO / "kb/communities"
+CACHE = REPO / "references_cache"
+
+# (record, reference) pairs where a low ratio is correct and understood. Each
+# needs a reason: an empty allow-list entry is how the next Methylacidiphilum
+# gets waved through.
+_ACCEPTED: dict[tuple[str, str], str] = {}
+
+# At or below this share of members named in the source, the pairing is suspect.
+# `<=`, not `<`: the case that motivated this is a two-member coculture whose
+# source names exactly one of them, which is 50% — a strict `<` let the guilty
+# record through while flagging five innocent ones.
+_THRESHOLD = 0.5
+
+# Only full-text caches are checked. An abstract names few members by nature, so
+# on abstract-only entries this ratio measures **cache depth, not relevance**:
+# the first version of this file flagged five records whose caches are 3 KB, and
+# missed the 47 KB one it was written for. 12 KB is the same bar used to pick
+# candidates worth reading in the first place.
+_FULL_TEXT_BYTES = 12_000
+
+
+def _cache_path(reference: str) -> pathlib.Path | None:
+    stem = reference.replace(":", "_").replace("/", "_")
+    for suffix in (".md", ".txt"):
+        candidate = CACHE / f"{stem}{suffix}"
+        if candidate.is_file():
+            return candidate
+    return None
+
+
+def _member_words(document: dict) -> list[str]:
+    """One distinctive word per member — the genus, usually.
+
+    Deliberately loose. Matching a full `preferred_term` would miss
+    "C. kluyveri", and matching any word would hit "sp." and "strain".
+    """
+    words = []
+    for entry in document.get("taxonomy") or []:
+        term = entry.get("taxon_term") or {}
+        name = term.get("preferred_term") or (term.get("term") or {}).get("label") or ""
+        head = name.split()[0] if name.split() else ""
+        if len(head) > 3 and head[0].isupper():
+            words.append(head)
+    return words
+
+
+def _pairs() -> list[tuple[str, str, float, int]]:
+    """(record, reference, share_of_members_named, member_count) for cached sources."""
+    found = []
+    for path in sorted(COMMUNITIES.glob("*.yaml")):
+        document = yaml.safe_load(path.read_text(encoding="utf-8")) or {}
+        setups = document.get("cultivation_setup") or []
+        if not setups:
+            continue
+        members = _member_words(document)
+        if not members:
+            continue
+        references = {
+            evidence.get("reference")
+            for setup in setups
+            for evidence in (setup.get("evidence") or [])
+            if isinstance(evidence, dict) and evidence.get("reference")
+        }
+        for reference in sorted(references):
+            cached = _cache_path(reference)
+            if cached is None:
+                continue  # nothing to check against; not this test's business
+            text = cached.read_text(errors="replace")
+            if len(text) < _FULL_TEXT_BYTES:
+                continue  # abstract-only: the ratio would measure the cache
+            named = sum(1 for word in members if re.search(rf"\b{re.escape(word)}", text))
+            found.append((path.name, reference, named / len(members), len(members)))
+    return found
+
+
+@pytest.fixture(scope="module")
+def pairs() -> list[tuple[str, str, float, int]]:
+    return _pairs()
+
+
+def test_there_are_pairs_to_check(pairs):
+    """Guard: with nothing cached the check below passes on an empty list."""
+    assert len(pairs) >= 1, (
+        f"only {len(pairs)} cultivation_setup/reference pairs have a cached "
+        f"source; this check is close to vacuous"
+    )
+
+
+def test_every_source_names_most_of_the_community(pairs):
+    """The discriminator that would have caught the Methylacidiphilum case."""
+    suspect = [
+        f"  {record}: {reference} names {share:.0%} of {count} members"
+        for record, reference, share, count in pairs
+        if share <= _THRESHOLD and (record, reference) not in _ACCEPTED
+    ]
+    assert suspect == [], (
+        "these cultivation_setup blocks cite a source that barely mentions the "
+        "community's members, which is how a monoculture's conditions get "
+        "recorded as a coculture's (#529):\n"
+        + "\n".join(suspect)
+        + "\n\nCheck whether the paper actually studied this community. If it "
+        "did and the names are simply written differently, add the pair to "
+        "`_ACCEPTED` with the reason."
+    )
+
+
+def test_the_accepted_pairs_have_reasons_and_are_current(pairs):
+    """An allow-list that outlives its entries stops meaning anything."""
+    blank = sorted(key for key, why in _ACCEPTED.items() if not (why or "").strip())
+    assert blank == [], f"accepted pairs need a reason: {blank}"
+    known = {(record, reference) for record, reference, _, _ in pairs}
+    stale = sorted(key for key in _ACCEPTED if key not in known)
+    assert stale == [], f"these accepted pairs no longer exist: {stale}"
+
+
+def test_the_check_can_actually_fail():
+    """Mutation check, without touching the corpus.
+
+    The assertion above passes today because every shipped record is fine —
+    which is also what it would do if `_member_words` returned nothing, or the
+    cache lookup silently missed every file. This runs the same arithmetic on
+    the case that was rejected, and requires it to score below the threshold.
+    """
+    document = yaml.safe_load(
+        (COMMUNITIES / "Methylacidiphilum_Galdieria_Thermoacidophilic_Coculture.yaml").read_text(
+            encoding="utf-8"
+        )
+    )
+    members = _member_words(document)
+    assert len(members) >= 2, "the rejected example no longer has two named members"
+
+    cached = _cache_path("PMID:33841379")
+    assert cached is not None, "the rejected example's source is no longer cached"
+    text = cached.read_text(errors="replace")
+    named = sum(1 for word in members if re.search(rf"\b{re.escape(word)}", text))
+    assert named / len(members) <= _THRESHOLD, (
+        "PMID:33841379 now names most of that record's members. If the record "
+        "or the cache changed, re-check whether the paper studies the coculture "
+        "— the whole point of #529 is that it does not"
+    )