paper-cluster-by-theme

Take a list of relevance-filtered papers and produce thematic clusters that the synthesis layer can write about rather than list.

A "cluster" here is not a topic label — it's an argument. "Diffusion models for protein design" is a topic. "Diffusion models are converging on AlphaFold-comparable accuracy without MSAs" is a theme. The synthesis is much stronger when clusters describe what the papers collectively argue.

Workflow

- [ ] Step 1: Read all kept papers' titles, abstracts, and matched keywords
- [ ] Step 2: Identify 2-5 candidate themes from the abstract content
- [ ] Step 3: Assign each paper to its best-fit theme (or to outliers)
- [ ] Step 4: Re-name each theme as an argument-shaped phrase
- [ ] Step 5: Validate clusters (no empty cluster, no over-stuffed cluster, no paper in two)
- [ ] Step 6: Return clusters + outliers + assignment rationale

Step 1 — Read inputs