v0.2.0 — Built with Claude: Life Sciences hackathon submission

Added
MCP server — the contradiction engine is now a Model Context Protocol server (src/mcp.mjs), so Claude or any agent can call it as a tool: compare_studies, compare_example, list_examples. Published to npm as studydiff-mcp and to the official MCP Registry as io.github.nickjlamb/studydiff. compare_example runs fully offline — no API key, no network.
Shared Markdown report (src/report.mjs) — one auditable report format across the web app, the CLI and the MCP surface: verdict, ranked drivers, verification counts, and every value with the verbatim sentence that supports it.
Third worked example — Treg lineage stability (Zhou 2009 vs Rubtsov 2010): a landmark T-cell-immunology controversy where the disagreement (stable vs. unstable Treg lineage) traces to the fate-mapping method.
Unified dropzone — one drag-and-drop target that splits into Paper A / Paper B on first interaction, so two PDFs can be dropped in a single action while still allowing mixed inputs (PDF · PMID/DOI · paste).
Changed
Run time roughly halved. The two Claude extractions now run concurrently, extraction input is capped, and quotes are constrained to the shortest exact span. Output dropped from ~2.5k to ~1.2k tokens per card, taking a full-text pair from ~19–21s to ~11s. Grounding still runs in 0–15ms.
Results rebuilt around the answer — a scannable "why these studies differ" card (verdict → most likely reason → each study's conclusion), a visually dominant primary driver, in-body accordions for the detail panels, and a unified right rail with verification metric tiles (claims verified · fields not reported · 0 invented).
Colour now carries meaning only: teal = Study A, purple = Study B, green = verified.
En dashes throughout the app.
Fixed
Full-text PDFs no longer fail with "Request too large" (compare-body limit raised to 8 MB).
The results reveal no longer jumps: layout settles before scrolling, and the embedded iframe no longer fights its own resize.
Added HSTS and upgrade-insecure-requests, so the app can't be reached over plain HTTP.

Full Changelog: v0.1.1...v0.2.0

StudyDiff v0.1.1 — hackathon submission

Verbatim supporting sentences now shown on screen rather than in a hover tooltip. Determinism claim scoped correctly: grounding, comparison and driver ranking are deterministic; extraction is an LLM call. New hero subhead.

StudyDiff v0.1.0

First public release. See CHANGELOG.md.