v0.1.5

0.1.5 — 2026-11-12

• Claude backend: a malformed answer no longer fails the paper outright. The parser now tolerates prose around the JSON object and probabilities written as strings ("0.85"); a still-malformed answer gets one repair attempt, with the model's own reply echoed back and what was wrong with it. Verified against the live API: a bad reply {"ai-agents": "high"} was repaired to {"ai-agents": 0.95} in one turn.
• Quarantine: a paper that fails three days in a row is skipped for 30 days (data/quarantine.jsonl) instead of being retried — and billed — every morning; it gets one fresh attempt when the window expires, and run --rejudge gives it a fresh chance immediately. This stops the 13 papers in the Claude example repo that failed every day on Claude response-format errors from burning tokens on retries that never resolve.

v0.1.4

0.1.4 — 2026-11-12

• Fix __version__ lagging behind pyproject.toml: it still read 0.1.0 on the 0.1.3 install, so paper-radar --version printed 0.1.0 while pip reported 0.1.3. New regression test asserts __version__ matches pyproject.toml, so a future bump cannot silently drift again.

v0.1.3

What's in 0.1.3

• Discoverability (PR #19): Claude is now named on the first screen of README and README.zh-TW.
• The PyPI description now mentions Jev or Claude, and the PyPI keywords add claude and anthropic.
• This release also ships the Claude calibration doc from PR #18: docs/claude-calibration.md describes the Claude run calibration.

v0.1.2

What's in 0.1.2

• Honest calibration labelling (A6): paper-radar check now warns when the backend is Claude that its probabilities are self-reported, not calibrated, and paper-radar calibrate says the same when fitting thresholds to Claude-judged labels. The Cochrane screening numbers (96.9% recall, 78% work saved) were measured with Jev and do not apply to Claude.
• GitHub Action (A7): new action.yml at the repo root — any repo with a radar.toml can run Paper Radar with uses: Eliot5566/JEV-Paper-Radar@v0.1.2. This is the first tag containing action.yml (@v0.1.1 predates this file). Keys go in as secrets and are never printed.
• Screening story (A8): new docs/screening.md for systematic-review readers — how screening mode works, how to run and measure it, and the CLEF TAR 2019 benchmark against Cochrane reviewers' decisions.

v0.1.1

Fix: Claude backend no longer sends temperature — claude-haiku-5-5 rejects it with 400. Verified live against the Anthropic API.

v0.1.0

First packaged release: pip install paper-radar. Adds Claude backend (claude-haiku-5-5) alongside Jev, plus CHANGELOG.