dismech: Skill for Claude Code

Review Hypothesis Exploration

Treat a provider report as a research lead, not as a source of truth. Assess the report's reasoning and claims without automatically promoting its claims into a disorder YAML.

Assess one report

1. Locate the raw report at kb/hypotheses/<Disease>/<hypothesis_id>/<provider>.md and identify the hypothesis directory, provider slug, and any provider artifact bundle. Read the report and inventory before judging the run.
2. Inventory every material data source in data_sources, including databases searched with no result. Distinguish CITED_NOT_ACCESSED, ACCESSED, SEARCHED_NO_RESULT, and UNVERIFIABLE; a proposed future dataset is not an accessed input. Preserve a committed query/response or search-log artifact for ACCESSED and SEARCHED_NO_RESULT. Verify each accession with just verify-datasets --accession <CURIE> where supported, then separately check disease, tissue, cohort, organism, and assay relevance. Resolution does not establish relevance.
3. Inventory every claimed computation in analyses. Trace input data-source IDs through method, software/version, parameters, code/environment, and outputs to the assessment claims the report attributes to that analysis; status and auditability determine whether execution actually supports them. Use SUCCEEDED only when inspectable artifacts substantiate execution; otherwise use PARTIAL, REPORTED_ONLY, FAILED, or SKIPPED as appropriate.
4. Treat tool availability and fallback as evidence-quality facts. If retrieval, a data lake, or a scientific tool failed, record the failure and any fallback; do not silently treat a literature-only or model-knowledge fallback as a provider analysis or independent result. Biomni is disabled at repository entry points unless DISMECH_ENABLE_BIOMNI=1; do not bypass that opt-in. Without it, Biomni must also remain unavailable to automatic provider fallback. The hypothesis runner's dry-run command inspection is safe without the opt-in.
5. Check the report's consequential claims against the cited primary literature. Distinguish direct support, external plausibility, qualification, and contradiction. Check taxonomic level, disease/model context, directionality, and causal versus correlational language.
6. Create one authoritative YAML sidecar per provider and assessor at assessments/<provider>-assessment-by-<assessor>.yaml. Use a lowercase, hyphenated assessor slug (for example, openai-5-pro).
7. Include the relative source_report, hypothesis_id, an overall verdict, and claim-level dispositions. For every assessment claim, use a short report_quote copied verbatim from the provider report and explain the assessment in rationale; attach analysis_ids only to claims the report attributes to those analyses, and use status/auditability to state whether execution supports them.
8. Add an optional Markdown narrative and/or PDF using the same stem when it helps human review. The YAML remains authoritative; the rendered artifacts must be listed in artifacts.
9. Apply the hypothesis artifact policy in docs/hypothesis-report-assessments.md: commit manifests, code, environment specifications, and small derived outputs; keep large/raw, controlled, or credential-bearing data outside Git and record whether each artifact is committed, external, local-only, missing, or not produced. Structured paths must be non-empty files beneath artifact_root: ../<provider>_artifacts. For a computational bundle, require canonical MANIFEST.yaml, run just validate-hypothesis-analysis-run <report> <artifact_dir>, and replay the saved code separately; the validator does not execute generated code. Confirm that report frontmatter has an artifact_manifest_sha256 binding to the exact current manifest bytes; a manifest edit makes an older report stale. If you correct provider code or results after its response, record the exact correction and before/after hashes in the manifest, replay it independently, and keep the provider analysis at most PARTIAL until the provider reruns or attests the corrected bytes. Never stamp a stale provider report onto an assessor-corrected bundle merely to make the gate pass.
10. Keep a report-review citation as context only. Promote a paper-derived claim to the disease YAML only after normal reference-cache and evidence validation. Never hand-edit references_cache/*.md; use just fetch-reference <ID>.
11. Validate the sidecar: