Introduction

DeepSeek Harness (DSH) adopts a plugin-based architecture to address common hallucination issues in biomedical report generation by large models. Unlike resource-routing plugins, dsh-bioinf-verify focuses on the single responsibility of “evidence verification.”

The most subtle failures in biomedical/scientific reports usually fall into three categories: nonexistent citations (hallucinated DOI/PMID), broken links, and misattribution. This plugin engineeringizes this discipline through a deterministic registry and temperature-0 semantic judgment.

Core Features

The plugin provides standalone verification capabilities, covering common evidence types in biomedical research.

  1. Existence Verification
    - Check whether citations exist (hallucinated DOI/PMID/Accession).
    - Verify that links truly exist/are accessible (redirect following, distinction between 404/410 and 403/401, soft-404 heuristics).
    - Verify clinical trial status (existence/status/phase/result posted).
    - Verify GEO/SRA dataset/sequencing archive accession.
    - Verify UniProt protein accession + gene correspondence.

  2. Metadata Verification
    - Verify DOI existence + metadata + title comparison.
    - Verify PMID + retraction/Expression-of-Concern scan.
    - Verify titles (papers/patents/datasets/news web articles), match ≥75% or containment / close 40–75% / mismatch.

  3. Semantic Consistency
    - Consistency between a single report claim and the original meaning (registry source text + temperature-0 LLM entailment judgment).
    - Batch citation audit matrix (final-draft gate).

  4. Report-level Workflow
    - Decompose report → multi-aspect checks → persistent ledger → generate inline annotations and a complete verification report.

Component Tools

The plugin provides report-level workflow tools and individual check tools.

Report-level Workflow

Tool Purpose
report_verify_start Decompose the report into verification points, classify them as literature/patent/clinical_trial/dataset/protein/webpage, and record source anchors
report_verify_step Execute single-aspect checks and immediately persist the ledger, supporting crash recovery
report_verify_status Read structured progress from the ledger
report_verify_finish Generate inline annotations in the original report and a complete verification report at the end
report_verify_list List historical verification jobs

Individual Check Tools

Tool Validation Content Data Source
url_verify Link truly exists/is accessible Direct HTTP
title_verify Title verification (paper/patent/dataset/news) Crossref, PubMed, PatentsView, GEO, page og:title
doi_verify DOI existence + metadata + title comparison Crossref
pmid_verify PMID verification + retraction scan PubMed
clinical_trial_status Trial existence/status/phase/result posted ClinicalTrials.gov v2
geo_accession_verify / sra_accession_verify Dataset/sequencing archive accession NCBI gds / sra
uniprot_verify Protein accession + gene correspondence UniProtKB
claim_audit Batch citation audit matrix Aggregation of the above
claim_semantic_check Single “report claim vs. source meaning” consistency Registry source text + temperature-0 LLM

Ledger Data Structure

Each verification job generates a persistent JSON (<workDir>/<jobId>.json), containing status, the full original report, decomposed claims and their aspects, and an event log. The ledger supports crash recovery.

Installation and Configuration

Installation

Install via the DSH command line (recommended):

dsh plugin --profile <profile> add git+https://github.com/gsh150801/dsh-bioinf-verify.git

After installation, the package name is automatically appended to the profile’s dsh.profile.bundles. Additional configuration must be added in the profile (see below) and dsh web must be restarted.

Configuration

All secrets are read only from profile configuration or environment variables. Only PatentsView requires a free API key; if not configured, related components return application guidance.

- insert:
    - id: bioinf-verify
      name: dsh-bioinf-verify
      config:
        workDir: ''                # 账本/输出根目录；默认 ~/.dsh/science-verification
        contactEmail: ''           # 或环境变量 CONTACT_EMAIL
        pubmedApiKey: ''           # 或环境变量 PUBMED_API_KEY
        patentsviewApiKey: ''      # 或环境变量 PATENTSVIEW_API_KEY
        guidance: true             # 注入工作流使用规范到系统提示
        llmRouter:                 # 分解 + 语义蕴含用（温度 0）
          enabled: true
          baseURL: 'http://127.0.0.1:8012/v1'   # OpenAI 兼容端点
          model: 'deepseek-v4-flash'
          apiKey: ''

Applicable Scenarios

  • Quality gate after biomedical AI report generation.
  • Cross-verification of citations and data during scientific paper/patent review.
  • Long-cycle verification tasks requiring crash recovery and traceability.

The plugin runs with the permissions of the current DSH process. It is recommended to review the source code and license (MIT) before installation.