02 knowledge graph lab

Knowledge Gravity Lab

Use this skill to turn a messy folder of notes into a practical knowledge map:

• center notes and themes,
• possible contamination/noise,
• overgrown topic clusters,
• orphan notes,
• cleanup actions,
• research or invention follow-up candidates.

This is a practical public knowledge-hygiene workflow, not a patent filing package and not a legal freedom-to-operate opinion. Keep the implementation framed as heuristic note analysis and avoid claiming that it implements or bypasses any protected system.

Quick Start

Run the bundled analyzer on a folder of Markdown/text files:

python scripts/analyze_corpus.py --input "C:\path\to\vault" --output "D:\knowledge-gravity-output"

Optional: skip folders that should not affect the map.

python scripts/analyze_corpus.py --input "C:\path\to\vault" --output "D:\knowledge-gravity-output" --skip-dir "archive" --skip-dir "templates"

The script creates:

• knowledge_gravity_nodes.csv
• knowledge_gravity_edges.csv
• knowledge_gravity_report.md
• knowledge_gravity_action_sheet.md
• knowledge_gravity_data.json

Then summarize the report and action sheet for the user, focusing on 3-5 concrete cleanup or research actions.

Output Quality Checks

After running the analyzer, check whether the output is useful:

• If one topic group contains most notes, tell the user the corpus needs stronger subtopic links/tags.
• If the review queue is mostly templates or daily notes, suggest adding --skip-dir or excluding low-value folders.
• If the center notes are all index files, explain that the map is showing navigation hubs rather than content hubs.
• If there are fewer than 20 notes, frame the result as a smoke test.

Workflow

1. Identify the corpus folder.
2. Run scripts/analyze_corpus.py.
3. Read knowledge_gravity_report.md.
4. Read knowledge_gravity_action_sheet.md when cleanup actions are requested.
5. Explain:
   • what the knowledge center is,
   • what looks noisy or low-signal,
   • what clusters are too large,
   • what should be split, merged, archived, or reviewed.
6. If the user wants publication, read references/public-boundary.md and avoid legal or protected-system claims.
7. If the user asks whether this collides with existing rights, read references/existing-rights-check.md and frame the answer as a risk checklist, not legal advice.

Interpretation Rules

Treat the analyzer output as decision support, not ground truth.

• High center score means a note is central in the local corpus.
• Review score means a note needs human attention, not that it is false.
• The action sheet is a temporary cleanup workspace. It should make cleanup choices easy, not replace human judgment.
• The 80-point feedback loop is an organization heuristic based on links, tags, headings, size, and repository share. Scores are capped to a practical 15-100 range; 80 is a cleanup target, not a natural breakpoint.
• Noise/contamination candidates are items with weak text signal, suspicious names, excessive boilerplate, poor linkage, or signs that they should not be trusted as a core source yet.
• Overgrown clusters are candidates for splitting into subtopics.
• Orphans are candidates for linking, archiving, or merging.

Public-Safe Language

Prefer:

• "knowledge hygiene"
• "center topic"
• "topic gravity"
• "attention-weighted note map"
• "possible noise"
• "possible contamination"
• "review queue"

Avoid:

• legal conclusions about existing rights,
• statements that the skill implements a protected internal system,
• "protected mechanism proves...",
• "guaranteed detection",
• "hallucination-proof".

When Results Are Weak

Say so directly. A useful result may be:

• "The corpus is too small."
• "The clusters are too clean to stress the method."
• "This proves only smoke-test scalability."
• "This needs baseline comparison."

Then propose the next experiment or corpus improvement.