CMIP-Forge Pairs 6,581 CMIP6 Papers with Agentic Code Execution and Adversarial Review for Autonomous Climate Workflows
CMIP-Forge demonstrates a retrieval-augmented agentic architecture that autonomously retrieves CMIP6 knowledge, executes constrained code on ESGF archives, and subjects outputs to independent model review. The preprint quantifies both successful workflows and concrete failure modes of the review layer. This establishes a reproducible template for scaling Earth-system research while preserving methodological invariants.
The system ingests unstructured CMIP6 knowledge and executes end-to-end pipelines on atmospheric teleconnections, ocean dynamics, extremes, and projections. Defense-in-depth combines static AST checks, curated scientific primitives, and an adversarial reviewer panel that issues REVISE verdicts. Telemetry logs reveal recurring failure modes including sycophantic regression and unresolved review loops, all traceable to immutable provenance records. Traditional CMIP workflows require sequential human coordination across literature synthesis and data access; CMIP-Forge collapses these steps into a single audited loop while surfacing its own limitations for later human inspection. As CMIP7 planning accelerates, such agentic systems offer a measurable route to scale analysis without relaxing physical constraints, provided the documented review pathologies are systematically mitigated.
CMIP-Forge: At least three CMIP7-endorsed groups will release code repositories containing agent-generated, reviewer-audited analysis scripts by June 2027.
Sources (2)
- [1]Primary Source(https://arxiv.org/abs/2606.17076)
- [2]Supporting Source(https://arxiv.org/abs/2307.03172)