A research framework called PHACT strips language models of the power to certify their own physics outputs — and, in 80 adversarial trials, produced zero false certifications.
Physics-Anchored Certification works as a loop: the language model proposes a value, and a separate deterministic engine decides whether that value is certified, impossible, or unknown. The key design choice is that the checker derives the certified quantity from fixed inputs rather than accepting a model-supplied value at face value. That distinction matters because a checker that simply trusts the model's answer can be fooled; one that recomputes from scratch cannot. The researchers tested the system across five scientific domains, two models, two decoding temperatures, and a deliberately broken engine.
The result points at a structural fix for a well-documented problem: language models confidently produce wrong answers, especially in technical domains where errors are hard to spot. Outsourcing the authority to certify — rather than trying to make the model less wrong — sidesteps the reliability problem at the architectural level. That approach is more auditable and, in principle, more trustworthy than fine-tuning or prompt-level guardrails.
The zero-false-certification record across adversarial conditions is a strong early signal, though 80 trials across five domains is a narrow sample. The harder question is whether deterministic checkers can be built at useful scale across the full range of physical design problems — or whether PHACT works best as a narrow safety layer for well-defined, computable quantities.