让科研 Agent 真正进入研究流程：SciForge 的设计与实践

TL;DR - SciForge is an open-source workbench that wraps mature coding agents (Codex, Claude Code) with persistent research-project state, structured scientific-object references, and evidence graphs so multi-day scientific work stays auditable, reproducible, and handoff-ready. It matters because it targets the gap between one-shot agent answers and real iterative research workflows.

• Separates research goals (questions, object scope, metrics, stopping conditions, release criteria) from any single chat session; each agent run persists inputs, artifacts, and judgments for later researchers.
• Structured scientific object references carry file path, content hash, version, viewer selection, and model/chain/residue locators; a Scientific Model Router dispatches protein sequence/structure and small-molecule inputs to domain models, with single-cell data handled by a Cell2Sentence worker.
• Two-tier provenance: a session-level Evidence DAG (demo: 32 nodes, 21 edges — flagged an overstated/hallucinated PDB citation) and a Project DAG aggregating 4 sessions, 25 evidence records, 16 claims into 45 nodes / 48 relations; background auditing flags unsupported or conflicting claims without gating routine steps.
• Eight end-to-end demos with configurable autonomy and human checkpoints, including ESMC-6B ContactProbe hyperparameter search (24 runs, 7-min budget), MCFST spatial-transcriptomics reproduction (ARI 0.7007 vs. reported 0.693, with logged discrepancies), EGFR molecule optimization (+1.7 kcal/mol, below its preset 2.0 threshold), and genome-to-BGC prioritization (430 regions → 23 candidates); code, releases, and paper are public on GitHub.