AI Designs Real Physics—And Now It Can’t Lie About Whether It Works

A language model that hallucinates a bridge load rating or a chip thermal spec isn’t a productivity tool—it’s a liability bomb. Two researchers just published a framework claiming zero false certifications across eighty adversarial trials. That number deserves both your attention and your suspicion.

What happened

Vyas and Stoev introduce PHACT (Physics-Anchored Certification), a “propose-certify” loop in which a language model never gets to be the final word on whether a physical design is valid. The model proposes; a deterministic engine independently derives the certified quantity from fixed inputs and returns one of three verdicts: certified, impossible, or unknown. The key architectural insight is anti-hallucination by construction: because the certifier derives values from scratch rather than checking the model’s supplied value, a model that lies about its answer cannot forge a passing certificate. The system was tested across five scientific domains, two models, two temperature (sampling) settings, and—critically—a deliberately faulted engine. Across all eighty adversarial trials, the framework produced zero false certifications. The “unknown” return code is load-bearing here: the system admits ignorance rather than confabulating confidence.

Cold read

Eighty trials across five domains sounds like a broad test; it is actually about sixteen trials per domain, which is thin coverage for any claim approaching “reliable.” The paper tests adversarial inputs, not adversarial domain breadth—we don’t know what happens when the underlying physics is outside the deterministic engine’s model, a situation that is common in real engineering where approximation regimes collide. “Zero false certifications” is a strong result, but it’s trivially achievable if the system routes aggressively to “unknown”—the abstract doesn’t break down how often that third verdict fires, which is the number that actually determines whether this is useful or just cautious. The faithfulness vs. groundedness problem isn’t solved either: PHACT can certify that a design satisfies the equations the engine knows about, not that those equations are the right ones for the real-world artifact. Building the deterministic certifier for a new domain is itself a hard, expensive engineering problem that the paper does not address.

What it means for you

• Signal maturity: 2/5 — proof-of-concept with clean lab conditions; no production deployment data
• Who gets hurt: AI-native CAD and physical simulation startups pitching LLM-generated designs directly to regulated industries (aerospace, medical devices, civil infra)—their “the AI checked it” assurance just became obviously inadequate
• What breaks if this is true: The current market assumption that human review is the only viable safety layer for LLM-assisted engineering gets replaced by a certifier-as-moat architecture, shifting defensibility to whoever owns the deterministic engine
• Why it might not land: Every new domain requires a complete, trustworthy deterministic certifier—that’s the hard part of physical design software and it doesn’t get easier because you add an LLM in front of it
• Watch for: A major EDA vendor (Cadence, Synopsys, Ansys) announcing a formal “propose-certify” integration; that would be the signal that this architecture is crossing from paper to product

Forecast as of 2026-06-30

By Q3 2027, at least one venture-backed engineering-AI startup will publicly adopt “propose-certify” framing in its safety documentation—but fewer than half of those implementations will feature a genuinely independent certifier, defaulting instead to LLM-as-judge loops that violate PHACT’s core anti-forgery property.

Source: Structural Certification for Reliable Physical Design with Language Models — Nakul Vyas, Iliya D. Stoev. https://arxiv.org/abs/2606.30107v1