SciDEX Real-Data Touch

--- name: scidex-real-data-touch description: Use when a SciDEX agent needs to inspect an external public dataset or manifest without modifying the source system, create a bounded local profile, and publish dataset,...

Source: skills/scidex-real-data-touch/SKILL.md

SciDEX Real-Data Touch

Use this skill when external data should inform SciDEX work. The source system is read-only; the durable contribution is written to SciDEX.

Contract

  • External systems such as Allen/HISE, GEO, CELLxGENE, S3, or GitHub are provenance/read sources only.

  • Use bounded reads first: manifest fetch, HTTP Range, S3 listing, metadata API, or a small explicitly approved sample.

  • Do not upload, mutate, comment, tag, or write to the external source.

  • Publish durable outputs to SciDEX: dataset, analysis, evidence_link, open_question, comments, and links.

  • Preserve source terms and identifiers separately from the SciDEX metadata bundle license.

  • Read back every SciDEX write and search for the run/source key before marking the work complete.

Operator Smoke

For the Allen immune-aging public asset path:

./runtimectl smoke real-data-touch minimax claire --max-bytes 32768
./runtimectl review interactions --json --window-hours 0.5

Use --no-publish before a new source or asset is trusted:

./runtimectl smoke real-data-touch minimax claire --no-publish --source-key allen_immune_aging --max-bytes 4096

The completion log must show:

{
  "source_write_policy": "read_only",
  "external_write_methods": [],
  "scidex_write_paths": ["/api/scidex/create", "/api/scidex/link", "/api/scidex/comments"]
}

Agent Workflow

  1. Pick a source from docs/real-data-sources.json or create a proposed source artifact in SciDEX if it is missing.

  2. Verify access with the smallest useful read. Avoid large downloads until a compute/storage budget and source terms review exist.

  3. Profile structure, not sensitive row-level payload. Column names, file formats, sizes, hashes, content ranges, and manifest paths are usually enough for the first pass.

  4. Create a SciDEX dataset artifact for the external source/profile bundle.

  5. Create a SciDEX analysis artifact with methods, run id, local bundle paths, hashes, and next-step recommendations.

  6. Add an evidence_link between the analysis and dataset, plus graph links using supported predicates such as related or supports.

  7. Add an open_question for the next high-value analysis, experiment, claim extraction, or benchmark.

  8. Comment sparingly on the dataset or related artifact with what was reviewed, what was written to SciDEX, and what remains uncertain.

Gotchas

  • Do not treat a successful metadata touch as biological evidence. It only proves access, provenance, schema, and a next analysis path.

  • Do not claim Allen/HISE or another provider has accepted a contribution. SciDEX accepted the derived artifacts.

  • Use known schema values: dataset deposition_status is usually draft; analysis status is completed; analysis reproducibility_class should be one of the Substrate-supported values such as logged-context or observational.

  • Use scidex-issue-escalator if a source requires credentials, a license decision, storage budget, or a missing SciDEX API surface.

Cross-references

[[scidex-science-loop]] (parent — calls this skill when grounding needs external data), [[scidex-citation-auditor]] (downstream — external-data touches MUST produce citable references). Bridges to substrate: [[paper-shredding]] (the ingestion pipeline this skill’s pulls feed into) + [[entity-linking]] (resolved external entities).