codex-science

Provide a one-stop, auditable scientific research service from vague intent through question framing, literature, datasets, experiments, compute, artifacts, manuscripts, evaluation, review, iteration, and final handoff. Use for end-to-end research projects, literature reviews, data analyses, reproductions, paper or grant workflows, long-running studies, scientific-agent comparisons, repeated eval-and-repair work, or requests for a Claude Science-style research workbench in Codex.

Build evidence-grounded scientific literature reviews, source and claim ledgers, search logs, paper cards, contradiction maps, gap analyses, and citation audits. Use for papers, PDFs, BibTeX or Zotero exports, literature folders, systematic or scoping reviews, research landscapes, disputed claims, or requests to find and synthesize scientific evidence.

Prepare, record, grade, validate, and compare reproducible scientific-agent benchmark runs across Codex, Claude, or other systems. Use when measuring evidence traceability, uncertainty, protocol discipline, safety gates, reproducibility, loop decisions, regression behavior, or claims of scientific-agent parity on the same tasks and resource constraints.

Search Crossref, PubMed, and OpenAlex through read-only scientific metadata APIs, save auditable normalized snapshots, and import reviewed records into Codex Science evidence ledgers. Use for literature discovery, DOI or PMID metadata, reproducible database searches, source intake, connector verification, or testing scientific database access without treating search results as scientific evidence.

Run bounded, auditable improvement loops for scientific research and engineering. Use when a task needs repeated plan, execution, trace capture, evaluation, repair, and re-evaluation; when coordinating external skills, plugins, MCP tools, or compute backends; or when success must be proven by explicit gates instead of declared from prose.

Create and refine reproducible scientific artifacts including figures, tables, notebooks, reports, manuscripts, slide decks, spreadsheets, interactive HTML, molecular or structural views, and data packages with code, input lineage, environment identity, checksums, and visual verification. Use whenever scientific results must become publication-ready or reviewable artifacts.

Design and manage falsifiable, reproducible scientific experiments and computational studies with preregistered plans, dataset lineage, environment snapshots, controlled local or remote compute, immutable result events, sensitivity checks, and explicit stop criteria. Use for experiment design, data analysis plans, simulations, replication studies, scientific software validation, HPC or cloud jobs, and requests to reproduce or compare results.

Perform adversarial scientific review of studies, literature syntheses, datasets, experiments, figures, manuscripts, and research packets. Use for independent verification, manuscript hardening, reproducibility audits, citation checks, statistical and causal-claim review, peer-review preparation, or deciding whether evidence supports a claimed conclusion.