---
title: "Knowledge-gap detection — how a system finds what it doesn't know"
type: "moc"
date_created: "2026-07-07T00:00:00.000Z"
updated: "2026-07-07T00:00:00.000Z"
tags: ["gap-detection","knowledge-graphs","llm-metacognition","self-knowledge","infrastructure"]
---


The vault researched how to detect its own ignorance — and the notes that
came back split cleanly into two families by *substrate*: methods that read
a **graph's structure** (are there missing links, isolated clusters,
unfilled obligatory slots?) and methods that read a **model's internal
signals or behavior** (does the network's own activity, or its pattern of
failures, betray a gap?). This MOC was created 2026-07-07 (first vault-lint
flagged the cluster at 6+ notes with no MOC); it also anchors the vault's
own open gap-detection question.

## Structural signals — reading the knowledge graph
- [[claim-obligatory-attributes-as-gap-signal]] — an entity missing a slot
  its type requires is a machine-readable gap
- [[claim-kb-completeness-toolkit-cardinality-nca-recall]] — the formal
  toolkit beyond obligatory attributes: cardinality assertions, the
  No-Change Assumption, text-extraction recall
- [[claim-bridge-detection-lacks-pkg-validation]] — the honest caveat: the
  graph algorithms are Tier-1, but "bridge = gap" in a personal knowledge
  graph has no peer-reviewed validation. The cluster's own skeptic.
- [[claim-structural-gaps-not-modularity-score-are-gap-signal]] — the
  peer-reviewed anchor (Paranyushkin WWW'19, read directly): weakly-connected
  cluster *pairs* are the gap signal; the modularity value itself is a
  bias/diversity index. Sharpens what "modularity as gap signal" may claim.

## Behavioral / model-internal signals — reading the system
- [[claim-query-failure-clustering-as-gap-signal]] — repeated query failures
  on a topic are the behavioral signature of a gap
- [[claim-grade-gradient-rank-gap-detection]] — GRADE reads gradient-subspace
  vs hidden-state-subspace rank across layers (needs a backward pass at
  inference — a deployment cost, links to the economics cluster)
- [[claim-llm-explicit-implicit-gap-detection]] — an LLM can distinguish
  explicit from implicit gaps when scanning a corpus
- [[claim-llm-neural-metacognition-incomplete]] — the ceiling: model-internal
  metacognition is real but partial; don't over-trust it

## The vault's own open question
- [[question-gap-detection]] — the seed question this cluster answers
- [[question-consolidation-pass-vs-revisit-protocol]] — the sibling: once you
  can find gaps and contradictions, do you fix them targeted (hooks + lint)
  or globally (a consolidation pass)? The first vault-lint's catch-count is
  the evidence (see [[reflection-vault-lint-2026-07-07]]).
- [[claim-kgcl-defers-merge-split-topology-changes]] — the field's current
  peer-reviewed change vocabulary stops exactly where consolidation would
  begin: merge/split are explicitly future work.

## The through-line
Structural methods find gaps *cheaply and mechanically* (a Qwen/local-model
job); model-internal methods find gaps *the structure can't see* but cost a
forward-or-backward pass. The vault's design bets on the first family for
routine hygiene and reserves the second for what links alone miss — the same
targeted-first, global-only-if-evidence-demands stance the consolidation
question is testing.