RESEARCH_INDEX

An LLM-maintained literature knowledge base. Drop in a paper. Claude Code reads the full text, writes a source page where every claim is anchored to a verbatim quote, cross-links it into what you already know, flags anything that contradicts your existing pages, and logs it. Ask a question six months later and get a cited answer instead of a list of PDFs to re-read.

Not a RAG index — the synthesis happens at ingest time and is written down as plain markdown you can read, edit, and check against sources. Which is why it works in Obsidian, and why it outlives whatever AI tooling you use next.

Pattern adapted from Andrej Karpathy's approach to LLM-powered knowledge management and originally scaffolded from claude-second-brain, then substantially reworked for scientific rigor: verbatim-quote anchoring, mandatory DOI/PMID provenance, full-text-only ingest, and a human-only verified promotion.

---

Start here

Read	For
SETUP_GUIDE.md	First-time setup, end to end, with the reasoning. Start here.
TASKS.md	The same procedure as an ordered task list — hand it to your own Claude Code session.
CLAUDE.md	The binding schema. What the agent reads every session. Read it before you rely on the wiki for anything that matters.

Two things that will bite you if you skip the guide:

cd /path/to/RESEARCH_INDEX && claude    # ALWAYS. Launching elsewhere silently
                                        # disables the guardrails and the skills.
mise exec -- qmd vsearch "…"            # ALWAYS. A bare `qmd` crashes on the
                                        # wrong node ABI rather than warning.

---

Commands

Command	Does
/brain-ingest <path or URL>	Full 9-step ingest of one source: read, discuss, source page, index, cross-link, contradictions, log
/brain-search <question>	Cited answer synthesized from the wiki; offers to file it as a permanent Q&A page
/brain-refresh	Re-index and re-embed after a batch of ingests
/lint	Health check — orphan pages, broken links, open contradictions, stale claims, thin coverage
/brain-rebuild	Redesign the search schema from scratch. Destructive; rarely needed

Layout

wiki/                 LLM-maintained. index.md, log.md, overview.md,
                      sources/ (one page per source), qa/, topic + entity pages
raw-sources/          Immutable evidence base — articles/, pdfs/, personal/
CLAUDE.md             The schema the agent follows
.claude/settings.json Permission guardrails (load ONLY when launched here)

The five rules that make it research-grade

1. Every claim carries the verbatim sentence from the source. No quote, no claim.
2. Full text only — a summary built from an abstract is not a summary of the paper.
3. Everything starts agent-drafted; only a human sets verified, after reading the source.
4. raw-sources/ is immutable. Add freely; never edit. A claim whose source changed is unverifiable.
5. Dates are compact YYMMDD; source pages are YYMMDD_full_title_with_underscores.md from the publication date.

Exactly one of those five — raw-sources immutability — is enforced by a permission rule. The other four are enforced by nothing: no rule can see a field inside a file, or tell a real quote from an invented one. CLAUDE.md tags every rule with which kind it is, and that distinction is the point — knowing which of your safeguards are real is what lets you trust the ones that are.