An AI-driven research workflow for generating ideas, running experiments, and writing scientific papers.
Agent Claude Code
Resource gatekeeper. Other agents must call the budgeter before launching long-running training, expensive LLM calls, or held-out queries. The budgeter consults resbudgetledger and either approves the request, suggests a smaller request, or halts.
Agent Claude Code
Implementation and experimentation. Can write code, run scripts, and record findings to memory.
Agent Claude Code
Discover and structurally ingest papers from arxiv / openalex. Populates the literature index.
Agent Claude Code
Attempt formal proofs in Lean 4 for stated lemmas. Scope: small statistical identities (sample mean unbiasedness, Chebyshev/Cauchy-Schwarz/Markov/Bonferroni inequalities, simple CLT/MLE statements). Triage gate: only spawn when triageforformalization returns eligible=True.
Agent Claude Code
Read-only literature review, idea generation, and hypothesis proposal. Cannot modify code or files.
Agent Claude Code
Adversarial reviewer of finished manuscripts. Refuses to sign off on central result claims until publication-critical metrics trace back to pinned evidence (empirical checklist) AND theorem claims trace back to an empty diagnostic manifest plus either a Lean verification or an explicit unverified flag (proof…
Agent Claude Code
Independent verification of claims. Read-only access to code; can run verification tools but cannot edit.