Closure Instead of Training: Agentic Domain Modelling for Life Cycle Assessment

Can an agent match trained domain models without any training? An explicit domain model delivered through context and closed-world tools raises accuracy on a 2,335-item life cycle assessment benchmark while cutting the cost per execution.

Abstract

Specialist domains such as law, medicine and climate are served by models trained for them, through continued pre-training and fine-tuning, because their answers rest on conventions, tables and regulations rather than on verifiable rewards. We ask whether an agent can match trained domain models without any training. We treat the pre-trained model as a population model and a domain as a sub-population whose scenarios are samples of its statistics, and we propose agentic domain modelling: an explicit domain model of typed elements with coordinates in a factored scenario space, delivered to an agent through its context and through closed-world tools. Its measurable object is closure, the probability that a task’s determining elements lie in the model, the complement of open-set risk. We show that accuracy decomposes affinely in the closure degree, that coverage under a context budget is submodular, and that the completeness of a benchmark drawn from the domain is estimable with bias at most 1/N . In life cycle assessment we build LCA-Closure, 2,335 items in twenty tracks across all assessment phases, constructed from a corpus of 20,451 corporate reports and from official tables by an LLM-as-judge protocol, with measured completeness, and an agent system on a distributed harness with a typed locator, phase specialists, a closed-world tool and an open-set gate. The same agent rises from 76.6% open to 86.1% with full closure, almost entirely on tracks whose knowledge lives in tables (51.2% to 73.1%), while the cost per execution falls by 41%; growing the domain model at random moves accuracy along the affine law with stationary conditional accuracies; without training, the agent beats an encoder fine-tuned on ExioNAICS inside its own distribution (43.0% vs. 25.5% six-digit accuracy) but trails fine-tuned ClimateBERT classifiers on their label conventions (79.7% vs. 83.9%); and a gate built on the closure answers every benchmark task and no far out-of-domain probe. The study cost 56.0 CNY.

This is a working paper. The full text is in preparation for publication; the abstract above reports its current results.

@techreport{guo2026agentic,
  title       = {Closure Instead of Training: Agentic Domain Modelling for Life Cycle Assessment},
  author      = {Guo, Yanming},
  institution = {AIDC Research},
  year        = {2026},
  type        = {Working paper},
  url         = {https://www.ai-dc.ai/research/agentic-domain-modelling}
}