Implicit Thesis (Na)
Element: Sodium (Na · 11) · alkali metal, foundational.
What it does
Implicit Thesis finds tight neighborhoods of notes that pull toward a single position. Two variants ship:
- v0 (default, no LLM): returns the cluster — member notes, a centroid note, member quotes, a similarity score, and a folder-diversity gate.
- v1 (
--llm <provider>on the CLI, configured-LLM on Desktop and API): also calls the LLM to synthesize a one-sentence named thesis in the author’s voice. See What’s new in v1.1.0.
Not a cluster label. Not a tag. An actual thesis sentence in your voice.
Algorithm sketch
- Build a link graph over the indexed notes
- Apply hub-penalty smoothing (heavy hubs like daily-notes get down-weighted)
- Detect tight near-cliques (density ≥ 0.5)
- For each clique, pull representative quotes
- Synthesize a thesis sentence via the configured LLM (or skip with
--no-llm)
Thresholds
| Setting | Default | Where to override |
|---|---|---|
HUB_DENSITY_HARD | 1.5 | SPEC.md §9.2 |
HUB_DENSITY_SOFT | 0.5 | SPEC.md §9.2 |
| Min clique size | 3 | --min-clique |
| Confidence floor | 0.55 | --confidence-floor |
False positives
The most common failure mode is diary clusters: 12 daily notes all mentioning “the build” don’t constitute a thesis. The hub-penalty layer addresses this, but if your vault has unusually structured daily-notes the penalty may not fire. Add a project-root tag to the daily folder to make it visible.
Confidence
Confidence is computed via the falsification audit (see audit). A finding with confidence ≥ 0.65 is suitable for promotion; below 0.45, review by hand first.