Claude Code Academic Research Skill Set v3.8 Adds Claim Audit
Decision Brief
ARS v3.8 closes the second half of the gap left by v3.7.3. v3.7.3 already made each citation carry a locator anchor; v3.8 adds an optional audit (ARS_CLAIM_AUDIT=1) that fetches cited sources by anchor and judges whether the claim is truly supported by that source. Five new high-warning classes are introduced—unsupported claim, negative-constraint violation, fabricated citation, missing anchor, and uncited-constraint violation—rejected outright by the formatting terminal's hard gate. It also ships a calibration runner and a 25-item synthetic gold set, with acceptance thresholds of FNR<0.15 and FPR<0.10. For researchers and developers using this skill set, this means citation authenticity checking moves from "anchor tagging only" to "optional automatic comparison," which users can enable or not after installing. Note that the repo drives tests with a stub reviewer that returns gold labels, validating only the toolchain itself; calibration results from a real review model are not yet recorded, and the official team says its rollout plan is waiting on that result.
Sources
- Skill Radar(GitHub 趋势)
Trending hands-on MCP servers, agent skills, and AI-coding tools discovered from GitHub search momentum.
- GitHub:Imbad0202/academic-research-skills
留言
登入后即可留言,和其他 builder 交换实测心得。
还没有留言,抢头香。