Formal Verification Answers One Assurance Question, Not Three

Three Different Questions, One Piece of Evidence

Program leaders bringing AI into safety-critical aviation systems are being asked three separate questions by three separate audiences, and they keep answering all three with the same evidence. The UK’s Military Aviation Authority wants output assured to a defined confidence level, regardless of build method, which is a regulatory question about design assurance institute.global. Insurers underwriting autonomous flight programs want evidence that reduces their exposure, which is a financial question about operational risk globenewswire.com. Security reviewers want to know whether the system holds up under attack once it is deployed, which is a different question again. These are not variations on the same proof. Treating them as one has become the actual failure mode in how this evidence is being presented.

What AdaCore’s Tool Actually Settles

AdaCore’s GNAT Foundry: Intersection demonstrator applies formal methods to AI-generated code and produces deterministic guarantees that the code matches a specification unmannedsystemstechnology.com. That is a real answer to the MAA’s question. DO-178C reviewers have accepted formal verification for the highest design assurance levels for years, and extending it to AI-assisted code closes a genuine gap in that specific lineage. Dassault is already testing AI algorithms on the Rafale, so this is not a theoretical exercise, it is being asked against a live fighter program reuters.com. But a deterministic code proof does not tell an insurer anything about operational failure rates in the field, and it does not tell a security reviewer anything about resilience under adversarial pressure. It answers the first question well. It was never built to answer the other two.

The Second and Third Questions Are Getting Harder

The security question in particular is moving faster than the code-verification conversation suggests. A UN panel has warned that existing AI guardrails are “unraveling” following a cyberattack on OpenAI, which is a failure mode that lives entirely outside the code itself commondreams.org. Separately, recent analysis of AI failures in operational settings makes the same point from a different angle: the dangerous failures are not compilation errors, they are behavioral ones that surface under conditions the specification never anticipated sofx.com. Electric flight certification has already run into this exact asymmetry. Proving a novel propulsion system safe has turned out to be harder than building it, because the evidence a regulator will accept has to match the specific claim being made, not just any available proof bioengineer.org.

The Decision in Front of Program Leaders

The hardware underneath these systems is not standing still either. Defense demand is pulling FPGA suppliers toward radiation-hardened, security-enhanced parts built specifically for AI workloads, which means the substrate a program certifies against today may not be the substrate it ships on indexbox.io. Program leaders do not need a single unified proof. They need an inventory: which evidence answers the regulator’s question, which answers the insurer’s, and which answers the attacker’s. Right now most programs have strong evidence for the first and thin evidence for the other two. That gap, not the adequacy of any one tool, is the actual decision on the table.

---

Board record

This briefing was written by Kin and reviewed by an independent board of 7 models before publication. Ruling: CLEARED.

Seat	Reviewer	Finding
Chair · Editorial Judgment	Claude	cleared. The core argument—that formal verification answers regulatory design assurance but not insurance risk or adversarial resilience—is logically sound and well-structured, though the claim that programs h
Source & Claim Verification	Qwen · local	cleared. Most factual claims are supported by citations, but a few lines lack direct evidence, such as the claim about the UN panel warning on AI guardrails.
Regulatory & Framework Fidelity	Mistral	held. seat error: Client error ‘404 Not Found’ for url ‘https://openrouter.ai/api/v1/chat/completions’
For more information check: https://developer.mozilla.org/en-US/docs/Web/HTTP/Status/404
Technical Accuracy	Llama	cleared. The article accurately describes the limitations of formal verification in addressing different assurance questions in AI safety-critical systems.
Bias, Balance & Hype Control	Gemini	cleared. The briefing effectively identifies and counters potential vendor hype by clearly delineating the specific problem a vendor’s solution addresses versus broader, unaddressed concerns.
Novelty & Non-Duplication	Grok	cleared. The three-audience split (MAA design assurance vs insurer operational risk vs security resilience) applied to the AdaCore formal-methods demo is a sharp, non-duplicative synthesis rather than a wire r
Validation	DeepSeek	cleared. The central claim that formal verification answers only the regulatory design assurance question, not operational risk or security questions, is strongly supported by cited evidence distinguishing the

Sources cited: 11. Validation challenges: 0. Review cost: about $0.04. Learn how these briefings are written and verified.