Feedback erases evidence of software bugs: Virus research proves correction conflicts with verification

The idea of feeding software models with errors for correction may seem like a straightforward step in code engineering, yet this feedback operates covertly in a counter-productive way that threatens the reliability of safety-critical software. A new study by independent researcher Adarsh Agarwal, accepted at the Workshop on Verifiable Programming AI within the NeurIPS 2026 conference, found that feedback loops aimed at correction actually erase the evidence and program traces that specification-completeness audit tools were designed to detect, resulting in programs that satisfy the written contract but violate the intended operational behavior.

The experiment relied on twelve programming tasks built on the Virus environment for formal verification of Rust code, and they were tested across four different model families under precisely matched invocation budgets. The data showed that independent discovery of evidence without any feedback achieved more than twice the amount of failure evidence compared with the corrective feedback loop in the Intentshield system, recording 59.0 % versus 27.1 %. Repeated control experiments with identical iterative settings confirmed that the bulk of this gap stems from the feedback content itself rather than from independent re-runs, with this advantage persisting across all model families and extending beyond the stage of fixing linguistic analyzers.

Completeness auditing and correctness fixing carry conflicting objectives at the moment of direct execution.When a model is asked to fix a bug immediately to pass the check, it promptly modifies the code to satisfy the incomplete contract, sweeping away edge cases that expose contract gaps. By contrast, a staged-execution approach demonstrates the ability to retrieve the same total evidence yield without reusing prior responses, with most of the audited evidence failing in hidden test cases. Retaining these evidences led to a substantial tightening of the software specifications, achieving an execution success rate of 81.3 % for contracts adjusted via staged execution, compared with roughly 62.5 % for contracts modified using the traditional Intentshield approach.

The study set clear experimental limits for this effect, as the behavior did not recur in evaluations using the Dafny and Lean environments, confining the result to the workflow specific to the Virus environment with Rust. This discrepancy suggests that Rust’s strict memory-safety guarantees and mathematical verification demand special care when generative models intervene in contract formulation.

The findings have a direct practical implication for engineers and tech teams in the Gulf, Egypt, and the Levant, where banks, fintech firms, and sovereign cloud platforms are accelerating the adoption of Rust to secure critical services, alongside experiments with AI agents drafting contracts and verifying program logic. This research calls for a reassessment of reliance on rapid-iteration styles that pressure the model to bypass error messages, because the result may be code that superficially passes automated verification tools while hiding behavioral failures in the actual production environment. The practical step here becomesSeparate the evidence-extraction and specification-completeness checking phase entirely from the code-repair phaseand document failure cases and adopt them as test assets before allowing any model to modify the program formulation.

The research papers conclude with a decisive operational rule for anyone building high-reliability systems: preserve the documented failure and record it well before attempting to fix it.