Causal graphs are 62% fragmented and audit-graphs cannot see it

[realmarcin]

scripts/audit_causal_graphs.py reports DANGLING_EDGE (an edge naming a node that does not exist) and ORPHAN_NODE (a node no edge references at all). Neither fires when every node has at least one edge but the graph still splits into mutually unreachable components — which is the common case here, because a node picks up an edge to a neighbour long before it is wired back toward the trait node.

So the audit-graphs gate is green while most of the corpus is fragmented.

Measured

Over the full corpus on 2026-07-30 — 353 causal graphs, 4,136 nodes, treating edges as undirected for the connectivity test:

• 220 graphs (62%) have more than one connected component.
• 1,264 nodes (30%) sit outside their graph's largest component.

Worst cases, all with 12 nodes stranded outside the main body:

trait	nodes	components
physiology/lithotrophic.yaml	19	7
environment/temperature_range_very_low.yaml	15	7
environment/ph_delta_mid2.yaml	15	7
morphology/black_pigmented.yaml	18	6
environment/temperature_range_mid4.yaml	17	6
environment/temperature_delta_high.yaml	15	6

The generated environment/physiology traits dominate the tail.

This is a content gap — missing edges — not a schema defect. That is precisely why nothing currently reports it.

Two independent pieces of work

1. Add a connectivity defect to the audit. A FRAGMENTED_GRAPH and/or UNREACHABLE_FROM_TRAIT check alongside the existing two, reported in reports/causal_graph_audit.tsv.

It must land non-blocking first. At 62% it would fail just qc on day one, so shipping it as a hard gate is not an option until the backfill is well advanced.

A design decision to make explicitly: is the invariant "the graph has exactly one component", or the stronger "every node is reachable from the trait node"? The second is what actually matters for interpretation — a well-connected subgraph that never links back to the trait is still useless — but it is also the harder bar.

2. Backfill the missing edges, per trait, with real evidence. This is deep-research work (research-causal-graphs / deep-research-trait), not something to mass-generate: every added edge needs a reference and a snippet matching the existing ones. Mass-generating plausible edges would defeat the point of the graphs.

Worked example already in hand

data/traits/metabolism/cellulolysis.yaml was fixed as the first instance and is where this was found. Its graph split into 4 components with 9 of 14 nodes unreachable from the trait node. The repair added 7 evidence-backed edges (cellulose → cellobiose, cellobiohydrolase → cellobiose, endoglucanase / cellobiohydrolase / beta_glucosidase part_of cellulase, cellulosome → enables → trait, cellulolytic_genes → encodes → cellulase), grounded cellobiose to CHEBI:17057, and retyped cellulosome from CELLULAR_LOCALIZATION to GENE_OR_PROTEIN (it is a complex, not a location; its GO:0043263 grounding is unchanged).

Its own deferred remainder, worth carrying forward as the template for what "done" looks like per trait: the LPMO / CBM / transport / phosphorolysis branches, splitting cellobiohydrolase by chain end, and replacing the generic catabolite-repression edge with taxon-specific cip-cel evidence.

Reproducing the numbers

Load each causal_graphs block, build an undirected adjacency over nodes[].node_id from edges[].subject/object, and count components per graph. Recorded here so the baseline can be re-measured after any backfill batch.

[realmarcin]

Detection half shipped in #185; this issue now tracks the backfill

audit-graphs can now see this. #185 (merged) adds:

• UNREACHABLE_FROM_TRAIT — a node that has edges but sits in an island with no undirected path to a TRAIT node
• NO_TRAIT_NODE — a graph with no anchor to measure reachability from

One deliberate difference from the measurement above

This issue measured "nodes outside the graph's largest component" (220 graphs, 1264 nodes). The shipped check measures "nodes not reachable from the trait node" (219 graphs, 1314 nodes), which is the stricter and more meaningful invariant: a large well-connected cluster that happens to outnumber the trait's own component is still not part of the trait's mechanism. The two metrics agree closely, and the small delta is exactly the graphs where the biggest component is not the trait's.

Reachability is treated as undirected, as this issue assumed — curated predicates legitimately mix direction (cellulase -enables-> trait inward, trait -produces-> glucose outward), so a directed walk would flag correctly-modelled graphs.

It landed as a ratchet, not a warning

conf/causal_graph_audit_baseline.tsv freezes the 1314 known findings and --fail-on defaults to new. Pre-existing fragmentation never fails; any newly introduced island exits 1. So the corpus cannot get worse while this backlog is worked, and just qc stayed green on day one.

What remains — the content work

Backfilling the missing edges, trait by trait, with real sources. Every added edge needs a reference and a verbatim snippet; CausalEdge.evidence is schema-required precisely because mechanism assertions carry no METPO provenance of their own. The research-causal-graphs / deep-research-trait skills are the intended route.

data/traits/metabolism/cellulolysis.yaml is the worked example — 4 components, 9 of 14 nodes unreachable, fixed with 7 evidence-backed edges. 1 of 220 done.

After each batch, regenerate the frozen set with just audit-graphs --write-baseline so the ratchet tightens. At zero, switch to --fail-on any and drop the baseline.

Caution: do not satisfy the check with plausible-sounding unsourced edges. An edge added only to connect a component launders a guess into the knowledge base, which is worse than a visible gap. If a mechanism genuinely has disconnected parts in the literature, that is a finding worth recording.

Closing #189 as a duplicate of this — it covered the same backfill and this issue was filed first. Tracked in NEXT_TASKS.md section 5, item 2.