Skip to content

Literature Evidence MCP v0.5: scientific scope and residual limitations

This release adds auditable human-review workflows and bounded numerical or provenance operations. It is NOT certified as a complete systematic review or as an automated Cochrane RoB 2 / ROBINS-I V2 / GRADE rating engine.

Risk of bias and GRADE

  • record_risk_of_bias: two independent identified human reviewers, all domains, attributed evidence identifiers and rationale; conflicting domains require adjudication. ROBINS-I V2 uses the six-domain November 2025 draft (NOT the historical seven-domain 2016 version), and explicitly requires the assignment/per-protocol effect of interest. No official signaling-question algorithm or independently calibrated assessment is implemented. Official V2 reference: https://www.riskofbias.info/welcome/robins-i-v2
  • record_grade: human-entered downgrade and upgrade domain considerations, an identified reviewer and rationale. No automatic final certainty category; threshold selection and the ROBINS-I initial-certainty conventions need experts.

References: https://www.cochrane.org/events/version-2-robins-i-tool and https://book.gradepro.org/guideline/overview-of-the-grade-approach

PRISMA 2020

  • prisma_dual_screen: two reviewer decisions and explicit exclusion reasons, never force reconciliation of disagreements.
  • prisma_review_flow: verifies record counts after identification/duplicate removal and tracks the actual 42 reporting rows / 27 main items of PRISMA 2020, including 10a-b, 13a-f, 16a-b, 20a-d, 23a-d and 24a-c. Full-text eligibility, registration/protocol, study-level inclusion and adjudication remain explicitly unknown until documented by reviewers.

Reference: https://www.prisma-statement.org/prisma-2020-checklist

Full-text PDF structural provenance

pdf_page_provenance takes only bounded base64 PDF bytes, never untrusted URLs or filesystem paths. Returns PDF and page-text SHA-256 fingerprints, exact page number, textual quote offsets, text bounding boxes, table rectangles and image object bounds. It DOES NOT validate figure semantics, transcribe images, certify table contents, authenticate publishers or bypass copyright. Hostile documents should be parsed in sandboxed workers in production.

Publication update alert graph

  • propagation_alerts traverses user-supplied source-attributed citation edges.
  • crossref_update_alerts queries Crossref update metadata, reporting indeterminate status when unavailable. Both generate alert chains for correction, retraction and expression-of-concern categories without relabeling downstream citing papers as retracted. Graph edges and notices still need independent verification.

Quantitative analysis

cpp_random_effects implements DerSimonian-Laird inverse-variance random-effects pooling, Q, I-squared, tau-squared, approximate normal CI, approximate prediction interval, and optional leave-one-out refits. Known-value independent Python comparisons and invalid numeric inputs are tested. It does not cover REML, Hartung-Knapp, correlated estimates, meta-regression, risk of bias, effect extraction or determination of study eligibility.

Go retrieval and operations

go_literature_records fetches bounded real provider metadata (not just HTTP status) concurrently for Crossref, OpenAlex and Europe PMC, using a compiled allowlist, no redirects, fixed timeouts, result/body limits and JSON parsing. Python adds a local SHA-keyed persistent cache with a configurable TTL, plus cross-process one-batch-per-five-second request gating. Cache contents are UNVERIFIED local data, not evidence, and authenticated multi-user hosting and distributed quotas are not implemented.

Cross-platform and performance

Rust, Go and C++ run Linux CI with independent native code tests. GitHub also compiles native projects on Windows and macOS for portability. The benchmark script scripts/benchmark_native.py measures only warmed machine-specific Python/Rust literal search times and C++ DL operation times, not identical-language speed comparisons for all features or scientific retrieval accuracy. No claim of native acceleration superiority is made.

What remains scientifically open

Independent expert-labeled RoB/ROBINS-I/GRADE accuracy validation; full two-stage PRISMA retrieval, study-level inclusion/adjudication and reports; verifiable figure/table contents; independently verified citation edges; Hartung-Knapp/REML and publication-bias sensitivity; and tested production authentication, multi-host resource controls and licensed PDF redistribution. The project must not claim those are complete.

No npm Registry publication was requested or executed in v0.5.