A single-user research monitor where code fetches and versions evidence, Jev screens and scores it, and Qwen drafts cited notes for human review.
For each (fetched source, open research question) pair, Jev answers Noul gates (on topic, method transferable, evidence compatible) and then Score dimensions through Pydantic AI's native `typesafe:` model; deterministic code applies the thresholds and routes each pair to keep / review / drop.
Links source-level judgments to versioned questions and human labels; threshold fitting writes proposals rather than silently changing policy.
Fixed-version documentation and implementation material were inspected on 2026-10-24. No vault writes, Qwen generations or live judgments were performed. The author reports unfinished runtime/threshold goals; this review does not certify them as achieved.
Monitor standing research questions while keeping fetched evidence, judgments and generated notes separate.
Material analysed：
core.py hashes instructions, transmitted schema and Score level names. Changing wording or level descriptions changes what was asked even if the high-level function name stays the same.
Inspect the supporting source ↗The README describes fit as writing a proposal rather than applying it. Human labels are inputs to review, not a license to retroactively claim better production results.
Inspect the supporting source ↗Useful for a single-user research vault with explicit standing questions. Prefer a simpler feed if you cannot review generated claims or maintain question versions.
Read Ask’s question-identity logic in core.py and the README’s fit/export-cases workflow. Check that a rubric change produces a distinct identity and that proposed thresholds require a separate decision.