Journal — 2026-07-03
Promotion report — gap-detection capture
Promoted 10-inbox/raw/2026-06-28-how-should-seekvault-detect-gaps-in-its-own-knowledge.md (dated 2026-06-28, batch origin, 5 days past the 14-day inbox floor with time to spare).
Notes created
Four claim-notes, all in 30-notes/:
-
claim-obligatory-attributes-as-gap-signal — The Razniewski et al. 2023 (arXiv:2305.05403) obligatory-attributes mechanism: when entities in a class lack expected predicates, that absence is a machine-readable incompleteness signal. Applied to the vault: missing frontmatter keys are structural gap markers detectable by lint. This is the cheapest and most immediately implementable of the four mechanisms.
-
claim-query-failure-clustering-as-gap-signal — Medhat et al. 2026 (arXiv:2606.10736): repeated query failures on a topic, logged over time, are a behavioral signature of absent knowledge. Quantitative grounding: ρ = 0.491, p = 0.008 across 1,340 questions, Tier 1 source. Vault implication: batch-run search logs would constitute the interaction log the method requires. We don't have such a log yet.
-
claim-llm-explicit-implicit-gap-detection — Salem et al. 2025 (arXiv:2510.25055) GAPMAP system: LLMs can reliably distinguish explicit gaps (uncertainty lexical cues) from implicit gaps (inferable from section context). Applied to the vault: explicit gaps are greppable; implicit gaps require LLM inference over the link graph. Distinct from the structural method (C1) and the behavioral method (C2).
-
claim-llm-neural-metacognition-incomplete — Ji-An et al. 2025 (arXiv:2505.13763): LLMs monitor only a subset of their neural activations; their metacognitive space is substantially lower-dimensional than the full neural space. The practical implication: Claude's volunteered uncertainty statements are insufficient as a primary gap-detection mechanism. This note explicitly distinguishes the Ji-An dimensionality argument from the structurally different claims in introspection-access-problem and cot-faithfulness-anthropic-biology.
Questions routed
None created. All four claims are clean Tier 1 with no [unverified-*] flags. The existing question-gap-detection in 50-questions/ is the question this capture was built to answer; all four notes wikilink to it. I did not retire the question — it has broader scope (detection algorithms, integration with hop protocol, gap taxonomy) that the four claims address only partially.
What was skipped and why
Five further leads left in the capture, none promoted:
- Azuma Obsidian plugin — forum post, Tier 4. The mechanism (BFS, orphan detection against
metadataCache.resolvedLinks) is interesting, but forum posts don't clear the sourcing floor for technical-mechanism claims. Would need an academic source before promoting. - GRADE gradient detection (Wang et al. 2026) — Relevant to LLM internals but not to vault structure. The capture notes it's applicable only "if SeekVault ever uses an embedded model as an index." Not a claim about how to detect vault gaps now.
- Curriculum prerequisite graphs as vault ontology — Speculative extension of Claim 2's source. It's an application idea, not a claim the Medhat et al. paper makes. Worth returning to on a dedicated ontology run.
- InfraNodus plugin — Commercial product, Tier 3 product documentation. The modularity-as-gap-signal idea is interesting but has no mechanism paper backing it yet.
- Cardinality assertions and No-Change Assumption (Razniewski et al. 2023) — Same paper as Claim 1, richer toolkit. The capture explicitly flags this as needing a dedicated run. Not worked up.
Retrieve-before-write collisions
- Claim 4 had partial overlap with two existing notes: introspection-access-problem and cot-faithfulness-anthropic-biology. I promoted it anyway because the Ji-An et al. empirical basis (metacognitive dimensionality argument) is genuinely distinct from the SKBench results and the Anthropic attribution-graph work in those notes. The three notes are now explicitly cross-linked.
- No other collisions found.
Anything that felt off
The capture is well-sourced and honestly narrow — the batch worker did good work here. The four main claims are all Tier 1, with exact quotes. One flag: the capture's introduction says "narrow capture. Four core claims directly answering the topic." That's accurate but I'd note the capture is unusually clean for batch output. No tier inflation, no suspicious paraphrasing. The grounding quotes are precise and traceable.
The only sourcing question I'd want Cali to know about: the Ji-An et al. paper (2505.13763) was revised 2025-10-24 from its original 2025-05-19 date. The capture cites the May date. I used the May date in the note frontmatter as the original submission date, which is standard for arXiv, but the revised version may have updated findings. Worth a verify pass if this claim becomes load-bearing for any spec or draft.
— Seek
Note on commit: All files are written and staged-ready but git add requires interactive approval that cannot be given in a headless run. Per spec §9 (failure handling) I retried once and logged here. Per spec §8, the auto-commit timer will pick up uncommitted work within 15 minutes. No data loss risk — all files are on disk.