From reading notes to Research Graph

This example asks whether short breaks are associated with immediate recall. All material is invented. DemoA and DemoB are demonstration identifiers, not real papers or findings to cite. Open the generated page without installing Qiongli, Python or Node, or reproduce the actual CLI run from the repository.

Open the full interactive example · Download offline HTML · Download snapshot JSON

The embedded page scrolls independently; use the full-page link for more room. It uses the CLI's actual HTML template and generated data. It is a saved snapshot and does not follow later changes to the project.

Try three paths

1. Select CLM-1 and inspect its two supports relations. Each retains a limit: no random assignment, one reading task or no delayed measurement. A cites relation records a citation; it does not establish support.
2. Select DEC-1 to inspect a reviewed informs relation, then DEC-2 for a proposed one. Decisions shape the research without becoming supporting evidence.
3. Select CLM-2. Lasting improvement remains untested, with no supports relation. The source checks retain that gap.

The actual output contains 9 research records and 8 non-structural relations, including 2 support relations. Other unfilled canonical sources remain visible. This is an incomplete starter project, not a finished paper.

Follow the workflow

Step	Material or operation	Responsibility
Supplied notes	DemoA / Table 1 and DemoB / Table 2 in notes/synthetic-reading-notes.md	Invented teaching material
Normalized records	evidence/claim-evidence-ledger.csv retains claims, locations and limits	Checked normalization supplied with the example
Research decisions	context/decision_log.md and manuscript/claims_evidence_map.md	Accepted decisions, proposals and manuscript citations stay distinct
Registration and projection	CLI preview, matching digest-bound apply, then Graph reads	Existing native CLI
View and source inspection	HTML from the same snapshot; a source command returns the anchored ledger excerpt	Existing Graph view and source reader

The CLI has not interpreted arbitrary prose or PDFs. In real work, you or the model must check the material and save normalized records through the existing approval flow. This example demonstrates projection and traceability, not reliable automatic extraction by a model.

Reproduce locally

These repository development commands require Rust build tools and Python 3. Viewing the supplied HTML requires neither.

cargo build --manifest-path packages/qiongli-native/Cargo.toml --locked -p qiongli
python3 tooling/scripts/build_research_graph_example.py \
  --cli packages/qiongli-native/target/debug/qiongli \
  --destination /tmp/qiongli-graph-demo

On Windows, use the corresponding qiongli.exe and a new output directory whose parent already exists. The script refuses existing destinations. Choose a new directory for another run; previous results are never removed automatically.

The script copies tests/fixtures/research_graph_example/, registers that copy using isolated Qiongli configuration, and confirms preview/apply only for this new demonstration project. It leaves your normal registration and the supplied fixture files untouched, and launches no browser.

Output	Purpose
research-graph.html	Standalone offline view
snapshot.json	Complete Graph and readiness data
source-excerpt.json	Actual source read, anchored at ledger line 2
proof.json	Binary/input hashes, relationship counts and completed checks
commands.json	Actual CLI arguments and exit codes
project/, config/	Retained demonstration project and isolated configuration

The website's source commands bind to the project and snapshot used to generate this example. They cannot be applied to your own project. After reproducing it, use your new HTML and the same CLI, with that process's QIONGLI_CONFIG_HOME set to the configuration path printed by the script. You can also read the saved source-excerpt.json without executing another command.

Check the evidence

View this run's verification record · View the actual source response

The generator checks evidence limits, unsupported claims, decision relations, source anchors, revision-mismatch refusal, repeatable projection, HTML/JSON parity, and unchanged fixture/project files after reads. CLI calls run with an empty PATH, without an external model or online service. Repository checks also detect drift between the public example, current template and supplied inputs.

These are synthetic-data and program-behavior checks. Actual browser rendering, clipboard permissions and file downloads still need observation; successful generation does not establish those outcomes or research quality.