v1.4.0 — Tightened reference verification

What's New

Reference integrity hardening (driven by real test failure)

In testing, the model produced a DOI differing by just 3 characters from the real one (blood.2018894368 vs blood.2018894618) and attributed the paper to the wrong first author — because it wrote from memory instead of copying from PubMed. This release makes that failure mode much harder.

Changes

• Immediate write-to-file — metadata from get_article_metadata must now be written to _working_ledger.yaml immediately after each call, in the same turn — not batched or held in memory
• Step 4: verify ALL references — previously "spot-check at least 5"; now re-fetches metadata for every PubMed reference and compares DOI, title, and first author character-by-character
• Step 5: re-read ledger from disk — before writing the reference list, the model must re-read _working_ledger.yaml from file, not from context memory
• Expanded author fields — authors replaced with first_author + authors_full in YAML schema, both copied verbatim from PubMed
• model: opus, effort: max — added to frontmatter for best results
• Repo link fix — disclaimer template prevents URL substitution

Installation

Download literature-search.skill below and install in Claude Desktop via Settings → Skills → Add Skill.

Requirements

• Claude Desktop with PubMed and Scholar Gateway MCP connectors enabled
• pandoc (brew install pandoc)
• Optional: bioRxiv and Clinical Trials MCP connectors for enhanced coverage

See README for full setup instructions.