[good first issue] Add citation verification regression test

[federicodeponte]

Add a regression test that verifies citation accuracy does not degrade between releases.

What to do:

• Create scripts/eval_regression.py
• Load a fixed set of 5 golden topics from data/eval_topics.json
• Generate drafts, compare metrics (word count, verified citations, structure quality)
• Fail CI if any metric degrades >10%

Skills needed: Python, basic understanding of the agent pipeline

Related: EVALUATION.md

[sap6011]

Want me to work on this
??

[federicodeponte]

Hi @sap6011 — yes, happy to have you take this!

A few pointers before you start:

• The issue references EVALUATION.md, but that file doesn't exist in the repo right now. Use engine/utils/quality_gate.py as the scoring reference — it already breaks quality into word count, citations, completeness, and structure.
• Use engine/draft_generator.py::generate_draft() to generate each draft. It returns (pdf_path, docx_path) and writes the final markdown to exports/{base_filename}.md, which your script can parse for metrics.
• Keep the topics cheap/fast: 5 topics, one academic level (e.g. master), one citation style (apa). The script will be slow and cost API credits, so it should run in .github/workflows/live-validation.yml, not on every PR.
• Store golden baselines (the metrics from a known-good run) in a JSON file and fail if any metric degrades >10%.

If you want, sketch your plan in a comment here before opening the PR and I can confirm the approach.

[sap6011]

Yes sure, on it!!

[sap6011]

scripts/eval_regression.py calls generate_draft() per topic into a temp dir, extracts word count, citation count (flush-left lines in the References section), and header count from the final .md in exports/. Compares against a stored data/eval_baseline.json and exits 1 if any metric drops >10%. The --update flag writes a new baseline on a known-good build.

Does this sound good?