| name | concept-research-navigation |
| description | Use when asking about "concept research", "dossier", "research data", "verify claim", "check sources", "concept sources", "companion directory", "source extraction", "check images", "source quality", "research directory", "iter-*/sources", "Haiku paraphrase", "extraction quality", or when navigating concept research directories, evaluating source reliability, verifying quantitative claims against images, or assessing data sufficiency for a concept. Provides methodology for working with concept research in knowledge/concept_research/.
|
| allowed-tools | Read, Grep, Glob |
| user-invocable | false |
Concept Research Navigation
Methodology for finding, evaluating, and using concept research data in knowledge/concept_research/.
Core Principle
Research data is evidence, not truth. Every source has a quality tier and an extraction method. Dossiers synthesize across sources but are not authoritative for specific claims. Individual sources are authoritative for what they contain, but text extraction is lossy — images may hold data the text doesn't. Always trace quantitative claims to the most authoritative representation available.
Data Layout (Brief)
For full directory structure, R2 sync setup, and format details, see knowledge/concept_research/README.md. Minimal orientation:
- Concept directories:
knowledge/concept_research/{concept-id}/ (e.g., 11-magnetic-mirror/)
- Start with
{concept-id}/dossier.md for synthesized overview
- Evidence in
{concept-id}/iter-NN/sources/*.md — check ALL iterations
Source Discovery Protocol
To find research for a given concept:
- If you know the concept ID: go to
knowledge/concept_research/{concept-id}/
- If you don't: check
knowledge/concept_research/SOURCE_INDEX.md, or glob knowledge/concept_research/*/dossier.md
- Read
dossier.md for orientation — what the concept is, key parameters, which sources exist
- Find evidence in
iter-NN/sources/*.md — check ALL iterations, later ones may add new sources
- For each source, check if companion directory
iter-NN/sources/{name}/ exists. Its presence indicates richer data: images, original artifacts (raw.html/raw.pdf), and extraction metrics
Source Evaluation
Three quality tiers. How to identify and how much to trust:
| Tier | How to identify | Trust level |
|---|
| Direct extraction | YAML frontmatter with source:, backend:, content_hash_sha256: | Authoritative for what it contains. Text may be lossy — verify quantitative data against images. |
| Haiku paraphrase | No YAML frontmatter, starts with # Title directly (or is a .orig.md file) | Lossy summary by Haiku. Specific numbers and technical details may be wrong. Flag as unverified if citing values. |
| Dossier | dossier.md at concept root | Synthesized overview. Good for orientation. Do NOT treat as authoritative for specific numbers — trace claims to individual sources. |
Source authority hierarchy when sources disagree:
- Peer-reviewed paper > technical report > company website > news article > Haiku paraphrase
- Direct extraction > Haiku paraphrase (for the same underlying source)
- Later iteration > earlier iteration (for the same topic)
Image Inspection Protocol
When you MUST read images:
- You see
 in the text — the content (equation, figure) exists ONLY in the image
- You are extracting a number for analysis or modeling — cross-check against the table image
- The text references a table or figure by number ("see Table 3") — find the corresponding image
- Numbers don't add up or text seems garbled — the missing data is probably in a table/figure image
Path resolution: refs in a source .md like  are relative to the companion directory → iter-NN/sources/{name}/images/page_003_table_0.png.
If images/ is empty: binaries are R2-synced. Run ./scripts/sync_research.sh pull.
Cross-Referencing Methodology
To verify a claim:
- Find the claim in the dossier or analysis
- Trace to the cited source file
- Find the specific section/table/figure in the source
- For quantitative values: check the companion dir
images/ for the authoritative table/figure image
- If the source text seems incomplete, check
raw.html/raw.pdf in the companion dir or fetch the original URL from YAML frontmatter
To handle contradictions:
- Note both values with their sources
- Apply the authority hierarchy (peer-reviewed > company > paraphrase)
- If both are peer-reviewed, note the discrepancy — it may reflect different assumptions or time periods
- Flag for human review if the contradiction affects a modeling decision
Confidence Assessment
| Level | Criteria | Action |
|---|
| Well-sourced | Value appears in a direct-extraction source with image verification | Use directly, cite the source |
| Source-backed but unverified | Value appears in text extraction but not cross-checked against image | Use with note: verify against source images if precision matters |
| Paraphrase-only | Value appears only in Haiku paraphrase (no frontmatter source) | Flag as [unverified], seek better source |
| Inferred | Value derived from related data in sources | Flag as [inferred] with derivation reasoning |
| Estimated | Value based on analogy to similar concepts | Flag as [estimated] with analogue basis |
Related Skills
- For citation formatting and the DI→PR→model traceability chain: source-traceability
- For SysML doc comment syntax: sysml-conventions
- For project directory layout beyond concept research: project-structure