---

A research package connecting a Jev decision study with source-bound graph-synthesis experiments. It includes manuscript drafts, recorded model responses, methods, figures, and executable verification.

What the original study found

Task	Held-out evidence	Interpretation
Entity matching	413 DBLP–ACM pairs; macro-F1 0.9605 → 0.9859; false merges 8 → 2	A narrow improvement on the fixed split, with one additional missed match
Relation verification	339 SciFact examples; paired macro-F1 interval includes zero	No resolved improvement

The intervention combines question formulation and demonstrations. These findings do not establish general model superiority or independently attribute the effect to either component. See the methods and submission-readiness checklist.

Choose a reading path

Goal	Start here
Read the original decision study	Manuscript and tables
Audit an empirical claim	Claim-to-evidence map and frozen run
Explore graph synthesis	Extension guide, illustrated paper, and figure gallery
Understand data provenance	Source and data notes
Inspect the original package	Archival README and reproduction notes

Reproduce and verify

From the repository root, use Python 3.12 and an isolated environment:

python -m pip install -r graph_synthesis/requirements-verification.txt
python -B scripts/download_datasets.py
python -B -m graph_synthesis.verify --replay --tests

Dependency installation and the pinned dataset download need network access. The subsequent replay and original tests use a temporary runtime with network connections blocked and make no live model calls. The extension guide explains cross-platform numerical tolerances and further experiments.

Use the extension verifier for this expanded repository. The original scripts/verify_package.py expects the original closed inventory and rejects the additional extension files.

Archive integrity

This GitHub landing page is separate from the original root README: that file and the baseline MANIFEST.json are immutable inputs to the research verifier. The archived evidence remains byte-exact.

Public datasets are downloaded and hash-checked on demand. Dataset rights, attribution, and research limitations remain documented in the source supplement; this landing page grants no additional license.