LLMs can distinguish explicit from implicit knowledge gaps when scanning a corpus

When directed to scan a body of text for knowledge gaps, LLMs can reliably identify two distinguishable classes. Salem et al.'s GAPMAP system, validated across approximately 1,500 biomedical documents, defines them precisely:

• Explicit gaps — "directly signaled by high uncertainty lexical cues (e.g., 'unknown,' 'further research is needed')." These are gaps the source authors themselves flagged.
• Implicit gaps — "not stated directly by the authors of the research papers, but can be inferred from section-level context." These are gaps visible only from outside the text, by reasoning about what the text implies but does not establish.

The harder case is implicit. GAPMAP addresses it with TABI (Toulmin-Abductive Bucketed Inference), which structures LLM reasoning into four components: Claim (the implied gap), Grounds (evidence span), Warrant (a one-sentence reasoning link), and Bucket (binary confidence classification). Evaluation found LLMs exhibit "robust capability in identifying both explicit and implicit knowledge gaps."

Applied to SeekVault: an LLM pass over the vault's claim-notes can flag two distinct gap types. Explicit: notes containing uncertainty markers ([unsourced], [unverified-quant], inline hedges) — these are detectable by lexical scan without any inference. Implicit: notes where expected supporting context is absent — a claim references a mechanism with no note explaining that mechanism, or links to a concept with no corresponding concept note in the vault. The explicit/implicit distinction maps onto the vault's own flagging system and link structure respectively.

Each gap type calls for a different detection pass. Explicit gaps are cheaper: a grep or frontmatter lint will surface them (see also claim-obligatory-attributes-as-gap-signal). Implicit gaps require LLM reasoning over the note graph — inference about what should be present but isn't, based on what is. This is the most expensive but highest-yield pass for a well-maintained vault where schema hygiene is already good.

The behavioral method in claim-query-failure-clustering-as-gap-signal detects topic-level coverage gaps from the outside; this method detects semantic structure gaps from the inside. Together the three methods cover the detection problem at different granularities. See question-gap-detection for where this fits in the full sequence.