Research Toolkit

Setup guidance makes MCP optional and describes read-only installation comparison, backups, and preserving local changes. Review backends and record handling are also documented, supporting reasonable limits on access, user choice, and external effects; however, the supplied evidence does not show per-action confirmation for all state writes or external backend calls. The docs explain sources, review inputs, and version bindings, but do not specify handling and cleanup for sensitive material or credentials. Dependency security evidence is also thin because the dependency file contents and vulnerability response are not supplied. Citation and source-record requirements are thorough. Publisher identity is unverified and is not treated as a deduction.

Guidance for Lite and Full, review tiers, task records, and delivery is mutually consistent. The CI workflow lists many static contract checks across platforms, supporting a strong self-consistency score. A standard-library CLI and MCP alternative reduce dependency barriers, though MCP still requires installed dependencies and a compatible client. The installation checker documents exit-code meanings, but failure diagnosis and repair guidance elsewhere is limited in the supplied material.

The materials address industry, product, company, and technical research, with examples specifying audiences, scope, materials, and length. Lite and Full use cases and review boundaries are clear, including the rule not to present self-review as independent review. The Skill entry relies on users and agents to follow its task-based reading and selection instructions, so trigger precision is not fully established. Direct reading, Skill/plugin packaging, and local MCP/CLI paths support varied environments.

The bilingual entry points, research guide, quickstarts, setup, development, and evaluation docs have a clear structure. Installation, version comparison, update safeguards, and Python/MCP requirements are specific. Renames and older installations are addressed, but old and new names can coexist, so naming stability loses a point. Examples are extensive, though the supplied material does not include a complete FAQ. Limitations include the lack of proof for general quality advantage and the fact that installation checks do not prove runtime loading. MIT licensing is explicit; VERSION, compatibility policy, bilingual changelogs, and release steps are documented. Contribution and synchronization processes are described, but no accountable maintainer or support commitment is evident.

The workflow produces a complete report with usable citations, source and claim records, review findings, and revision records, making outputs usable and reviewable. A case shows a concrete recommendation changing after review, while the project explicitly says it has not established a general quality or efficiency advantage over a native agent. Thus there is an example of added value but limited overall proof. Lite/Full and optional CLI/MCP paths balance process overhead, but no systematic cost quantification or broad comparison is provided.

The methods require registering sources and important claims and retaining version-bound review records. The case links to a report, initial draft, sources, and revision archive, supporting strong traceability. CI names source-integrity, regression-fixture, and evaluation checks, but their implementation and full results are not among the supplied files, limiting assessment of actual cross-source corroboration. Examples call for information cutoffs, treating unspecified conditions as assumptions to verify, and stating evaluation limits, so fact and inference are separated well.