Usage

Use DSRH through a DeepSeek Harness profile. Raw MCP server names gain the official namespace mcp__dsrh-science__ when they enter the Harness tool registry.

PaperReader

Place a born-digital PDF inside the workspace and request dsrh-paper-reader. The workflow ingests it, reads section evidence, searches method and experiment facts, inspects basic figure/table extraction, and produces an evidence map. A core report claim must cite an exact id.

Scanned PDFs enter degraded-scan-unsupported mode. DSRH reports the limitation and does not run OCR or fabricate missing content.

CitationGuard

Place a BibTeX file in the workspace. Parsing and duplicate detection are local. Resolution uses only Crossref, OpenAlex, and arXiv, with bounded responses and a filesystem cache. Generated reports are stored at .dsrh/citations/reports/<name>.json and .md.

Claim support operates only on a paper_id already present in .dsrh/papers/index.sqlite. MCP returns candidate evidence; the skill performs semantic classification.

ExperimentPilot

Start from examples/experiment.yaml or the schema at schemas/experiment.schema.json. The standard flow is validation, expansion, dry-run review, explicit execution through Harness shell/jobs, metric parsing, aggregation, and provenance.

For a long command, ExperimentPilot passes the reviewed command to the platform shell tool with run_in_background: true. It monitors the returned job id with job_output and job_wait; it does not block one MCP call for training and does not use a Python subprocess launcher.

Aggregation retains the requested denominator. Failed and missing run ids appear beside metric mean, sample standard deviation, min, max, and count.

Combined Workflow

PaperReader can identify an experiment design and its evidence. ExperimentPilot can reproduce or ablate that design. CitationGuard can then verify bibliographic metadata and whether final report claims remain grounded in the local paper and experiment evidence. V1 does not recursively run this chain without user direction.