The engine, end to end.

Source APIs and research signal enter without LLM mutation.

Records, chunks, citations, entities, provenance, and lineage.

Ingestion, synthesis, validation, health, embeddings, and compute lanes.

Hosted runtime state plus local reproducibility from the same inputs.

Agents critique citations, confounders, gaps, and weak claims.

Versioned methods, content hashes, JSON payloads, and decisions.

Preview routes first. Writes require approval and dry_run=false.

GPU/Docker compute jobs are ledgered, artifact-backed, and gated.

The candidate record captures mechanism, citations, risks, methods, decision history, payload links, and known gaps.

TWOG proposes the next test most likely to kill the leading hypothesis, pre-registers a hashed kill-criterion before compute, and resolves its own target/therapy inputs. Confound and provenance gates must pass; nothing is auto-promoted.

A promising record becomes a bounded question with required inputs, readouts, controls, blockers, and success criteria.

Approval-gated Docker and GPU jobs can produce docking, MD smoke, plots, logs, and reproducible configuration artifacts.

Agents and operators decide whether simulated findings are coherent enough to justify more evidence or lab discussion.

The system prepares assay context, controls, thresholds, materials, and metrics for humans or partners to review.

Wet-lab confirmation, failure, or ambiguity feeds back into the candidate record and decision log.