📂 Examples

Status: reference artefacts. Written 2026-09-02.

Written for you, alongside the documentation: every file under sample-project/ is a real output of the real pipeline, run against the five synthetic sample papers committed in sample-project/papers/, so when a document says "a dossier looks like this" it can show you one that actually passed the gate. The documentation quotes these files throughout; this page is the map.

Two notes, so the samples cannot mislead:

• The sources are synthetic. The five "papers" were written purely as sample sources (each carries a notice; all are CC0). Their claims are illustrative, not scholarship. The machinery is real: every artefact here came out of a real corpus sync, a real draft gate pass, real review aids and real renders.
• The drafts were written the way the pipeline writes drafts: from logged retrieval over this corpus, citing only citekeys the ledger holds, each under 1,000 words, with the dossier filled as drafting went. They show the shape of a draft; they are not scholarship.

A second directory sits beside sample-project/ and is not pipeline output: dossiers/ holds hand-written scope.md and outline.md files, one pair per genre, plus a book's spec.md. They are the files you fill in, shown filled in, for the five start-to-finish tutorials under Writing. Nothing in them was produced by a run, and they contain no citekeys.

Two more directories, codex/ and opencode/, each hold one survey drafted end to end by that harness driving a local model, over the same five papers. Each is self-contained: the inputs, the prompt, the script that ran, and everything the run wrote, unedited. They record what one real model did, including where it left the skill's path, so their dossiers are less tidy than the sample project's. LOCAL-MODELS.md compares them, and LLM-AGENTS.md says how to set up each agent.

🗺 The sample project, artefact by artefact

Path (under sample-project/)	What it is	The document that explains it
papers/bibliography.bib + papers/files/*.pdf	the whole input universe: five entries, five PDFs	ZOTERO.md
content/drafts/dt-overview/survey.md	a literature survey (gate-passed, referenced)	GENRE.md
content/drafts/dt-overview/staleness-tutorial.md	a hands-on tutorial; every step was executed for real before presenting	GENRE.md
content/drafts/dt-overview/staleness-chapter.md	an undergraduate textbook chapter	GENRE.md
content/drafts/dt-overview/trust-chapter.tex	a thesis chapter fragment (\citep, no preamble)	GENRE.md, RENDERING-FLOW.md
content/dossiers/dt-overview/<stem>/	each draft's dossier: scope, kept evidence with claim:/quote:, rejected candidates, logged retrieval, sections, steering, revisions	DOSSIER.md, DRAFT-ITERATION.md
content/review/dt-overview/<stem>.*.md (+ .json)	the review layer's reports for each draft: provenance, verbatim, coverage, synthesis, uncited, quotation, support, and the merged agenda	REVIEW.md, CITATION-PROVENANCE.md, PLAGIARISM.md
content/rendered/dt-overview/	rendered outputs: the survey as PDF and Markdown, the thesis fragment as \input-ready .tex	RENDERING-FLOW.md
content/specs/twin-basics/	a signed book outline, its sign-off record, and one accepted unit (units/ch-staleness.json); the second unit is truthfully marked unwritten	WRITE-A-BOOK.md
content/seed_topics.toml, content/topics.json, content/topic_seeds.json, content/topic_set.json, content/topic_graph.json	the topic artefacts, from hand-written phrases through clustering to the derived graph	TOPIC-MODELLING.md, TOPIC-DISCOVERY.md
content/topic_map.html	the whole topic map as one offline page	TOPIC-DISCOVERY.md
content/topic_gold.toml, content/topic_gold_results.json, content/discover_digital_twin.txt	a gold query set, its measured scores per resolution rung, and one corpus discover transcript	TOPIC-DISCOVERY.md

🔄 Regenerating the machine state

The ledger, parsed text, embeddings and caches are deliberately not committed, because the pipeline rebuilds them. From sample-project/:

bash regenerate.sh

After that, every command the documentation demonstrates runs here unchanged: chitragupta corpus ledger, chitragupta corpus discover "digital twin", chitragupta draft gate content/drafts/dt-overview/survey.md, chitragupta review agenda content/drafts/dt-overview/survey.md, and so on. This directory is also the cheapest safe playground: nothing in it is anyone's real research.

⚠ One caveat worth reading before you copy numbers

Five papers is deliberately tiny. Some behaviour differs from a real corpus at this size, and two differences are themselves instructive: the topic graph has no shared-member edges, because sharing a paper between topics of size 2 and 5 in a five-paper corpus is what chance predicts (the hypergeometric gate refuses it; see TOPIC-DISCOVERY.md); and the gold-set scores in topic_gold_results.json are visibly imperfect, which is the reason to measure them rather than assume them.