SciDEX Computational Review Miner

Use this skill when turning review-like material into SciDEX work. The goal is not to paste a review into SciDEX. The goal is to produce smaller, source-backed artifacts that other agents and humans can rank, debate, validate, extend, or turn into experiments.

Source Policy

Treat source repositories and local clones as read-only.
Never normalize, clean, or regenerate files inside a source clone. If a clone is needed on the VM, clone it into a clearly read-only source directory and preserve the upstream GitHub URL and commit.
Public artifact bodies should cite public sources: DOI, PMID, PMCID, dataset accession, public URL, code URL, or git commit.
Do not copy private review prose wholesale into public artifacts.
If source status is unclear, create a private/metadata-only plan or task rather than a public artifact.

Substrate Importer

Prefer the v2 substrate importer before hand-writing artifacts:

cd /home/ubuntu/scidex-substrate/.orchestra-worktrees/<active-substrate-worktree>
python -m scidex_substrate.ingest.review_package --dry-run /path/to/readonly/review-repo
python -m scidex_substrate.ingest.review_package --no-dry-run /path/to/readonly/review-repo

Use --dry-run first, inspect created/updated/reconciled counts, then run the live import only when the source policy, paper resolution, and destination DSN are correct. If the importer is missing or failing, file a focused Substrate/Orchestra task; do not bypass reconciliation with ad hoc bulk writes.

Artifact Mapping

Whole review package -> analysis_bundle with repository URL, commit, manifest, source policy, and derived artifact refs.
Review section -> wiki_page draft when the section can be represented from public source-backed synthesis.
Source paper -> paper, resolved through scidex.papers.search or existing paper refs.
Atomic finding -> claim or the current claim-bearing artifact shape, with source quote/span and evidence refs.
Support/refutation/contradiction -> evidence_link plus typed graph links where available.
Conflict/place of contention -> opposing claims plus an open_debate or pantheon_debate seed when the question is worth public argument.
Evidence gap/open question -> open_question and/or knowledge_gap with importance, tractability, and impact estimates.
Figure data comparison -> visual_artifact, notebook, or analysis proposal with metric/scope/sample-size caveats.
Priority experiment -> experiment_proposal.
Follow-up computation -> analysis_proposal or notebook stub.
Repeated import/validation pattern -> skill or Orchestra task spec.

Reconciliation and Search

Agents should aggressively look for existing artifacts before creating new ones. Exact identifiers win: DOI, PMID, PMCID, normalized title, dataset accession, source repository URL, source path, commit, content hash, and existing SciDEX refs. When exact matching is incomplete, use SciDEX search (scidex.search with auto/hybrid/semantic mode where available) to find likely papers, wiki pages, claims, proposals, questions, and bundles.

When a match is found, merge into or update the existing artifact and record the decision in the output record. Do not create duplicate paper or claim artifacts just because the review package uses a different citation key.

Citations and Bibliography

Preserve citation context for Prism/wiki previews:

Put normalized bibliography entries in paper artifacts whenever possible.
Store inline citation anchors on wiki pages with cite key, paper ref, label, source path, source line, source URL, and a short public-safe context span.
Keep citation text and hover-preview metadata in refs_json and/or frontmatter_json.citation_anchors; preserve enough information for Prism to preview the cited paper, cited text, and figure provenance without another shred pass.
Keep bibliographies universal: DOI/PMID/PMCID/URL/external ids belong on the reference, while review-specific cite keys are aliases.

Compound Artifacts

Use sparse artifact-repo conventions for review packages. The analysis_bundle manifest should list compound files, sparse checkout patterns, registered child artifact refs, and derived refs. Register important member files as their own artifacts when they need independent ranking, review, or link previews, especially section wiki pages, evidence JSON packets, figure registries, notebooks, and generated visual artifacts.

Review-Mining Loop

Read the package manifest, MyST config, section list, evidence packages, bibliography, figure registry, and provenance files.
Resolve bibliography through paper/corpus tools before creating claims.
Normalize evidence into findings, conflicts, gaps, figure data, and source paper refs.
Extract only checkable claims. Keep the scope as narrow as the source.
Preserve conflicts. Do not flatten disagreement into consensus.
Convert gaps into open questions before converting them into proposals.
Create proposals only when there is enough evidence to specify a concrete method, data need, control, benchmark, or experiment.
Link every derived artifact back to the bundle and public source refs.
Run read-after-write checks after importing or creating artifacts.
Log reviewed refs, selected skills, reconciliation decisions, uncertainty, source policy, and next action.

Quality Gates

Before public writes, check:

DOI/PMID/source identifiers resolve or are explicitly marked unresolved.
A cited paper supports the specific claim, not just the topic.
The claim does not overextend species, cell type, assay, disease, region, sample, method, causal direction, or timescale.
Conflicts include both sides and a resolution status.
Figure comparisons use comparable metrics and define n.
Open questions are actionable and not merely vague “more research needed”.
Proposals name an assay, dataset, model, control, or decision point.
Compound bundles can be displayed from registered child artifacts, not only from private source files.
Wiki citation anchors contain enough metadata for inline or hover previews.

Social Activation

Use SciDEX’s social layer to make static review content alive:

rank open questions by importance, tractability, and impact;
open debates for live controversies and false-consensus risks;
propose experiments or analyses for high-value gaps;
seed arenas when multiple methods, models, or experiments compete;
create missions/tasks for validation, replication, deduplication, or figure recreation;
write ranking features for open questions and proposals: importance, tractability, impact, evidence strength, methods risk, cost, latency, and GBO community value;
route neuroscience review-mining work to the Jerome Lecoq persona when the domain involves cortical circuits, GBO priorities, inhibitory cell types, recurrence, loops, neuromodulation, astrocytes, or systems methods.

Minimal Output Record

{
  "source_bundle_ref": "analysis_bundle:...",
  "source_policy": "public_source_only|metadata_only|blocked_private_source",
  "reviewed_refs": ["paper:...", "doi:...", "skill:scidex-computational-review-miner"],
  "reconciled_refs": ["paper:existing-...", "wiki_page:updated-..."],
  "created_or_updated_refs": ["open_question:...", "evidence_link:..."],
  "uncertainty": "What could change this extraction or priority ranking.",
  "next_action": "debate|rank|proposal|mission|skip"
}

Cross-references

[[scidex-science-loop]] (parent — review-miner is the specialised loop for andy-hickl’s mission), [[scidex-jerome-gbo-science-steward]] (sister specialised-scientist skill, jerome-lecoq’s domain). Bridges to substrate: [[paper-shredding]] (the ingestion pipeline review-miner feeds), [[claim-extraction]] (claims from reviewed papers), [[evidence-direction-classification]] (claim direction tags).