All notable changes to Evidentia are documented here. This project adheres to Semantic Versioning and the Keep a Changelog format.
1.1.0 — 2026-06-18
- arXiv verification: arXiv IDs now resolve against the arXiv API instead of always falling back to manual review.
- Agent-readable resolver traces: JSON results now include
lookupVerifiedandresolverOutcomes, exposing matched / unmatched / unreachable / skipped registry outcomes without changing the human-facing 4-tier verdict. - Optional local registry cache:
--cache <file>for the CLI,cachePathfor MCP, andVerifyOptions.cachePathfor library callers. - Codex Desktop plugin metadata via
.codex-plugin/plugin.jsonand.agents/plugins/marketplace.json. - Benchmark expansion from 19 to 22 cases, adding arXiv and ISBN manual-review coverage.
- Hardened TypeScript type resolution with
types: ["node"]so stray copiednode_modules/@types/*folders do not break local builds. - Updated README, Japanese README, plugin metadata, and package metadata for arXiv, resolver traces, local cache, and Codex plugin installation.
1.0.0 — 2026-06-12
First public release. Evidentia becomes a tool, not just a prompt.
- Deterministic citation-verification engine (
evidentianpm package): extracts DOIs/PMIDs/arXiv IDs from any text and resolves them against CrossRef, PubMed, and OpenAlex — no API key required. - 4-tier classification: Verified / Bibliographic mismatch / Hallucination / Content-review-needed, distinguishing "real paper, wrong DOI" from "this paper does not exist" via title fallback search.
- CLI:
evidentia check <file|url|->with--format md|text|json,--out,--mailto,--offline, and--fail-on-fabrication(CI exit code). - MCP server (
evidentia-mcp): exposesverify_citationsto any MCP-capable agent; install withclaude mcp add evidentia -- npx -y evidentia-mcp. - Claude Code plugin packaging:
.claude-plugin/plugin.jsonandmarketplace.json, installable via/plugin marketplace add kgraph57/evidentia. - Skill integration: the 15-criteria appraisal skill now calls the engine for Step 4 citation verification, using its output as ground truth for citation existence.
- Clinical-trial verification: recognizes ClinicalTrials.gov NCT IDs and resolves them against the registry (real → Verified, unregistered → Hallucination).
- Book & grey-literature awareness: an ISBN, a clinical guideline, or any source not indexed in these registries is marked "verify manually" (Tier 2) — never a false hallucination. Identifier-less entries in a reference list are surfaced, not skipped.
- Batch input:
evidentia check a.md b.md …verifies multiple files and prints an aggregate report;--fail-on-fabricationconsiders all of them (used by the CI example). - Worked examples with real registry output, a GitHub Actions example for medical content repos, issue templates, a 17-case live benchmark, and a test suite (38 unit + 5 live API tests).
- Restructured the repository into a plugin layout (
skills/medical-fact-check/...). - Reference-file paths in
SKILL.mdare now relative and plugin-portable. - README rebuilt around a one-line install, a real demo, and the citation layer.