Alexandria

A citation-integrity-first academic writing system for Claude Code and Obsidian.

Large language models fabricate citations. A recent cross-model audit of ten models measured reference-fabrication rates between 11.4% and 56.8%, and most tooling attacks the problem after the text is written, by detecting hallucinated references. Alexandria attacks it before: it is a six-role writing system whose workflow makes fabrication structurally difficult. Every verbatim quote must be read back from the source PDF at a real page number, pass three verification layers, survive a blind review, and be audited line-by-line before a draft is allowed to call itself done.

Alexandria is not a library or a server. It is a set of role definitions, methods, and governance files you drop into Claude Code, pointed at your own Obsidian vault.

Where it comes from. I am a graduate student in Taiwan working at the intersection of political philosophy and AI ethics. Alexandria is the system that carries that research in production — every gate in it was added because something actually went wrong. I am sharing the skeleton so that students in neighbouring humanities fields have a working reference for building their own. It is tuned for interpretive work: close reading, verbatim quotation, page-anchored citation of books and articles. If your research is quantitative or in the natural sciences, the role architecture may still serve you, but the evidence layer assumes texts rather than datasets — expect to study the framework and rework that layer yourself.

How it grew. Alexandria was assembled at the end of May 2026 and has been in daily production ever since — two months of real research at the time of this snapshot. Nothing here was designed on a whiteboard: every gate traces to a documented failure, every default to a measured comparison. The role files carry dated changelogs (the Librarian's rulebook is eighty revisions into its third generation), and the overhauls that mattered most are told inside the files where they happened: the source-tier rule (after the founding incident of quotes copied from notes), the ban on "reconstructing from general knowledge" (root cause of every serious fabrication), the full-text corpus layer, per-family parallel discovery (adopted after a head-to-head experiment), and the council's redesign from self-played review to independent blind seats.

Influences. Two public projects left direct marks: academic-research-skills (quote anchors, the blind-review pre-commitment, anti-sycophancy) and everything-claude-code (the topic-report, council, and scholar-evaluation modes began as adaptations of its method prompts). What was evaluated and deliberately not adopted shaped the system just as much.

The wider system. Alexandria is one domain of a larger personal multi-agent system, internally called Chaos. Around its members Chaos maintains a constitutional layer that every agent re-reads before regulated actions, file-based messaging that lets agents cooperate across sessions, mechanical drift detection, security vetting for anything external, a survival protocol for context compaction, and a standing habit of turning incidents into new gates. Much of what makes Alexandria dependable in production is this reinforcement from above: the roles supply the discipline, and the wider system keeps the discipline honest. Everything published here stands on its own without it — and what is published is Alexandria alone.

What it is, and is not

It is:

• A role architecture: six specialised roles with strict separation of duties and write permissions.
• A citation integrity pipeline: the discipline that runs through every role, from ingestion to final audit.
• An Obsidian-native workflow: your notes and your sources form a mirrored pair the system maintains and verifies.
• A governance layer: mechanical defences against the slow drift that kills every complex prompt system.

It is not:

• A RAG server or embedding database (semantic recall is an optional integration, with the interface documented).
• An autonomous researcher that writes papers while you sleep. What matters most to me is the preservation of human agency — full participation in the thinking and in the work. The finished text represents you; that is why I refuse to build a fully automated text-production system. The human is a working part of this system, not its audience.
• A citation manager. It complements Zotero/BibTeX-style tools; it does not replace them.

The six roles

One permission rule anchors the whole system: only the Librarian writes to the vault. The Writer drafts in its own project folder; the Deep-reader adds only its own notes to one dedicated folder; the Searcher and Critic are read-only. This prevents working drafts from contaminating your source of truth, and prevents the echo chamber where a model ends up citing its own earlier output.

One article, one Writer, one folder. Every project gets a scaffold whose front door orients anyone who opens it — you after a week away, a review seat, the Critic — in thirty seconds. The rule underneath: a product that lives only in the conversation does not exist.

The citation integrity pipeline

This is the spine of the system, and the reason it exists.

1. Source tiers. Not everything that contains text is allowed to be a citation source:

Positioning layers tell you where to look. Only the source itself tells you what it says. Quotes copied from notes are second-hand and carry every error the note ever made; Alexandria forbids them in final drafts.

2. Three-layer quote verification. Every quote the Searcher returns must pass: (a) correspondence: the passage actually supports the claim it is attached to, not merely keyword-matches it; (b) not second-hand: the words are the author's own position, not the author quoting or summarising someone else; (c) settled position: the passage reflects the author's developed view, not a setup being torn down two pages later.

3. Blind review and anti-sycophancy. The Critic commits to its evaluation criteria before reading the draft, and operates under a standing rule: criticism is not softened to please, and a weak rebuttal may not dismiss a valid objection.

4. Final audit. Before any draft is delivered, the Searcher re-enters in audit mode and walks every citation-bearing claim, producing a support-status list: ✅ supported / ⚠️ needs adjustment / ❌ unsupported. Unsupported claims are fixed, explicitly re-labelled as the writer's own position, or cut. They are never quietly left in.

5. The human is the last line of defence. The pipeline reduces the error surface; it does not replace your eyes. Final verification against the source is a design assumption, not an afterthought. Every gate in this system was distilled from a real, post-mortemed failure, including entire fabricated summaries traced to "reconstructing from general knowledge", which is why that fallback is banned by name.

Obsidian integration

Alexandria treats your vault as a two-end mirror: a notes/ tree of literature notes and a sources/ tree of PDFs, with matched structure and mechanical verification that the two ends stay aligned.

• Note schema: a consistent literature-note format (metadata, structured summary, verified key quotes) that both humans and the roles can navigate.
• Wikilinks and MOCs: a four-step linking discipline plus Maps of Content, so the graph stays a map instead of becoming spaghetti.
• Vault map: a template for describing your taxonomy so the roles can navigate it (a synthetic example is included; bring your own).
• Hygiene tools: dead-link scanning, structure verification, and a rename-chain procedure so reorganisations don't silently break references.

Governance: how it survives its own growth

Prompt systems rot: definitions drift apart across copies, numbers go stale, "temporary" exceptions become permanent. Alexandria ships the counter-machinery it was built with:

• Single source of truth per rule, with everything else linking rather than copying.
• A sync matrix that lists every fact that lives in more than one place, and where its mirrors are.
• A health-check script that mechanically verifies version mirrors, forbidden-pattern usage, and structural invariants.
• A claims-and-evidence layer: "verified" must be bound to a trace that could only exist if the looking actually happened — rendered-page credentials on quotes, denominators on every passing summary, and no negative conclusion until the page has been read whole (claims-and-evidence.md).
• A corpus integrity layer: your grep can lie — three file-level failure families each invisible to the others' detector, three-layer extraction QA, corpus-statistics-driven repair, and an honest degradation registry the search index itself reports (librarian §8, degradation-registry.md).
• Exclusion-zone versioning: superseded rules are moved to a marked zone with a note on what replaced them, never silently deleted. The system remembers why it changed.

Getting started

Three adoption tiers, in GETTING-STARTED.md:

Start at A. Everything above it is additive.

Style modules

The skeleton's prompts are English and deliberately voice-neutral. Language- and style-specific rules (punctuation conventions, tone, idiom policies) plug in as style modules: a documented slot in the Writer's pipeline, with a synthetic example module included. Write your own; the upstream system this was extracted from runs a Traditional Chinese module with its own typography and idiom rules.

Optional integrations

Documented as interfaces, not shipped as dependencies: a full-text corpus layer (searchable text extracted alongside each PDF), semantic recall (an embedding index for cross-lingual candidate discovery, always recall-only, never a citation source), and an OCR escalation path for scanned sources. optional-integrations.md describes what each contributes, the contract it must satisfy, and where it plugs in. Equivalents are straightforward to build with common tools.

Status

Extracted from a live system (2026). The upstream continues to evolve; this skeleton is a curated snapshot, not a mirror. Issues and adaptations are welcome; the licence is MIT.