i.2Wiki
Agentic AI Atlas · Coverage
qa/03-coveragea5c.ai
I.
Wiki article

qa/03-coverage

Reading · 2 min

Coverage reference

Page nodewiki/qa/03-coverage.mdNearby pages · 4Documents · 0

Continue reading

Nearby pages in the same section.

Coverage

Phase-5 deliverable. Five distinct coverage measures. Each must be reportable as a percentage and tracked across releases. A regression in any of them fails CI gate 7 unless explicitly waived.

NodeKind coverage

"What fraction of the schema's NodeKinds have real-world entities populated in the graph?"

Code
NodeKindCoverage = |{ kinds with ≥1 non-fixture node }| / |all NodeKinds|

Fixture-only kinds count as 0. Excluded by waiver: kinds explicitly marked populationDeferred in the coverage checklist.

Edge coverage

"What fraction of declared EdgeKinds are exercised by at least one edge in the graph?"

Code
EdgeCoverage = |{ edge kinds with ≥1 instance }| / |all EdgeKinds|

Edges in ontology-schema.yaml that have zero instances are flagged for either population or removal.

Capability coverage

"What fraction of named Capability nodes have at least one CapabilitySupport binding?"

Code
CapabilityCoverage = |{ capabilities with ≥1 CapabilitySupport edge }| / |all Capability nodes|

A capability with no support binding is dead text. The threshold for new capabilities is 1 binding within 30 days of creation; missing the threshold opens a Gap.

Evidence freshness coverage

"What fraction of fact-bearing claims have evidence within the freshness window?"

Code
FreshnessCoverage = |{ claims with all evidenceSources in-window }| / |all claims|

Claims with stale evidence count as not-fresh until refreshed. Long-lived stable facts (e.g. "Claude Code is published as @anthropic-ai/claude-code") get a longer freshnessWindowDays from their policy; volatile facts (pricing, model lineup) get shorter windows.

Generator coverage

"What fraction of declared Generator nodes have an integration test and a passing run on the current graph?"

Code
GeneratorCoverage = |{ generators with passing integration test ∧ green run on current graph }| / |all Generator nodes|

A generator without a test is a future regression. A generator that fails on the current graph is a current regression. Both reduce coverage.

Composite

A single weighted score is reported on the home page of the wiki's /meta/coverage/:

Code
Composite = 0.20·NodeKind + 0.20·Edge + 0.20·Capability + 0.25·Freshness + 0.15·Generator

The weights are chosen so that **freshness** dominates — drift between the world and the graph is the single most expensive failure mode. Composite is informational; the per-axis percentages are the real gates.