DISCONNECTED rewards unsourced metadata: removing a fabricated abundance_level creates findings

[realmarcin]

Found while fixing #300. Concrete evidence for the DISCONNECTED policy question already raised in #273.

Two interacting rules

1. COMMUNITY_LEVEL interactions connect nothing. audit_community builds connected_taxa purely from source_taxon / target_taxon (network/auditor.py:189,229). A COMMUNITY_LEVEL interaction has neither by design, so it contributes no connections. 107 of the 302 records with interactions are entirely COMMUNITY_LEVEL, meaning a third of the KB has, structurally, zero connected taxa.

2. abundance_level or functional_role exempts a taxon from being reported (auditor.py:258):

if taxon_data.get("abundance_level") or taxon_data.get("functional_role"):
    continue

That exemption is why rule 1 does not produce hundreds of findings — most taxa carry one of those slots.

The perverse incentive

Combined, they mean removing unsourced metadata creates audit findings.

In #298 I had invented abundance_level values for SynCom_ARC_Peanut_Aflatoxin_Nodulation — AbundanceEnum is quantitative (DOMINANT = ">1% relative abundance") and the paper reports no abundances, so all of them were fabricated. Review removed them, which is unambiguously the right call. The immediate consequence was five new DISCONNECTED findings for that record, and #300's fix (dropping a misleading pairwise edge in favour of a COMMUNITY_LEVEL one) added two more: 34 → 36 repo-wide.

So the audit currently scores a record better for carrying invented abundances than for honestly omitting them. A curator optimising the audit would fill those slots with guesses.

Why this matters for #273

#273 has to decide what DISCONNECTED means before the gate can be restored. This sharpens it: as implemented, DISCONNECTED is not "this taxon has no curated interaction" but "this taxon has no pairwise interaction and no membership metadata" — a compound condition that mixes graph connectivity with slot completeness, and which the exemption clause makes non-obvious.

Options

• Credit COMMUNITY_LEVEL interactions. Treat every taxon in a record as connected by a community-level interaction, which is what such an interaction asserts. Removes the structural penalty on 107 records and makes DISCONNECTED mean what its name says.
• Drop the metadata exemption and let DISCONNECTED be purely about connectivity — cleaner semantics, but would surface a large backlog at once.
• Split the two ideas into separate issue types: no-interaction versus no-membership-metadata. They are different curation gaps and currently share a label.

The first looks correct on its own merits regardless of what #273 decides, since a COMMUNITY_LEVEL interaction genuinely does relate its members.

[realmarcin]

A written proposal covering this issue together with the other two in its theme is open as PR #309 (proposals/absence_semantics.md).

The three share a diagnosis — the schema can say "has value X" or "is empty", so empty is overloaded across not-done, impossible, undecidable, and deliberately-excluded — but the proposal argues they need three different remedies, and can be decided independently.

Simulating all four rule configurations over the current KB changed the recommendation:

configuration	DISCONNECTED reported
today	32
drop the exemption only	390
credit COMMUNITY_LEVEL only	1
both together	19

Crediting COMMUNITY_LEVEL alone is safe but makes the rule vacuous — one finding across 307 records. Dropping the exemption alone reports 390. The halves must ship together, which the proposal now says explicitly. This is the smallest of the three remedies and the one that unblocks #273.