Five records where the growth medium is defined by what it omits, and naming its contents would be wrong

[realmarcin]

Filed as curation guidance rather than a defect. The #183 sweep has now hit this five times independently, which makes it the most reliable single failure mode found in that work — and one no validator can see.

The pattern

A community's medium is chosen for what it lacks. The omission is what makes one member depend on another, so it is the experiment. Recording the medium by its contents produces a block that is verbatim-sourced, schema-valid, and describes a system in which the interaction under study is unnecessary.

record	medium	what is omitted	consequence of omitting the omission
Suillus_Bacillus_Thiamine_Ectomycorrhizal_SynCom	P20	thiamine	S. clintonianus is a thiamine auxotroph; with thiamine it needs no partner
Mesorhizobium_Synechococcus_B12_Synthetic_Consortium	BG11	vitamin B12	Syn7002 lacks the B12 cluster; the paper's control adds it back at 4 µg/L
Engineered_Gut_Amino_Acid_CrossFeeding_Consortium	M9	the cross-fed amino acids	the auxotrophs stop depending on each other; the paired condition adds them at 1 mM
Emiliania_Phaeobacter_Dynamic_Interaction	L1-Si	(nothing added — it is simply non-permissive)	Phaeobacter cannot grow in it alone; its growth in coculture is the measurement
Syntrophobacter_Methanobacterium_Syntrophy	basal medium	fumarate	fumarate lets S. fumaroxidans grow without a partner, abolishing the syntrophy

The last is the inverse case and worth keeping in the table: there the error would be including something the coculture did not get, because the source lists it for the pure cultures.

Why no gate catches it

• linkml-validate sees a valid string.
• validate-references sees a verbatim snippet — "grown in P20 medium" is in the source.
• The A cultivation_setup can be sourced from a paper that never studied the community #529 member-coverage check passes: the paper is about this community.
• The #262's DIET re-scoping missed one occurrence: same snippet still cited as SUPPORT #295 truncation gate passes: nothing is cut mid-word.

Every existing check is satisfied by the wrong answer. What is missing is not in the text; it is the absence the text names, and absences are only visible if you already know the biology enough to look for them.

What would help, in rough order of cost

1. A curation note in the add-growth-conditions skill: when a record's interactions describe auxotrophy, vitamin exchange, or cross-feeding, check whether the medium is defined by an omission before recording it. Cheapest, and it targets the moment the mistake is made.
2. A weak heuristic gate — flag any growth_media or cultivation_setup on a record whose ecological_interactions include CROSS_FEEDING or SYNTROPHY where neither notes nor instrument_detail contains a negation ("without", "-free", "lacking", "omitted", "deficient"). Would have fired on all five before they were curated. High false-positive rate expected — plenty of cross-feeding communities have unremarkable media — so it would need to be a warning, not a gate.
3. Nothing — accept it as a thing curators have to know.

I lean to 1, with 2 worth prototyping to see how noisy it actually is. Option 3 is defensible only while one person is doing all the curation.

Related: #183 (where all five were found), #529 (the sibling failure mode — conditions belonging to a member rather than the community).

[realmarcin]

Prototyped option 2. It does not work — dropping it from the recommendation.

The proposal was: flag any record with a cultivation_setup whose ecological_interactions include cross-feeding, syntrophy or mutualism, where neither notes nor instrument_detail contains a negation ("without", "-free", "lacking", "omitted", "deficient", "absence of", "non-permissive").

Measured against the corpus:

records with a setup + cross-feeding/syntrophy/mutualism:  49
  mention an omission:                                     13
  do NOT (would be flagged):                               36

A 73% flag rate. The flagged set includes Clostridium_Autoethanogenum_Kluyveri_Syngas_Coculture, BioRock_ISS_Basalt_Biomining_Consortium, Caldicellulosiruptor_TwoSpecies_Hydrogen_Coculture and most of the syngas and biomining records — communities whose media are entirely unremarkable and whose interactions are real. Cross-feeding on syngas or on mineral surfaces simply does not require a medium defined by subtraction.

A check that fires on three quarters of its population is not a signal. It would be ignored within a week, and worse, its presence would imply the class was covered.

Revised recommendation

Option 1 only — a curation note in the add-growth-conditions skill, at the point the medium is recorded: if the community's interactions turn on auxotrophy, vitamin exchange or a metabolic dependency, check whether the medium is defined by what it lacks before writing down what it contains.

That targets the moment the mistake is made, costs nothing, and does not pretend to mechanical coverage the problem does not admit.

Worth stating plainly: I suggested option 2 and it was wrong. The five instances share a biological property — one member cannot grow without the other, and the medium enforces it — and biological properties are not reliably visible in the shape of a record. The same reason #529 has no mechanical test either.