If an AI can predict what running code would do, can it skip actually running it when you ask?

This explores whether a simulator that predicts what running code or a tool would do, once corrected to be reliable, can replace actually running it when a model is answering a live request and not only when it is being trained. The corpus has no note that tests this directly, so what follows is the nearest evidence and what it implies.

The closest result is on the training side. Structured reasoning templates let a model judge whether two code patches behave the same without running either, and it gets this right 93% of the time on real agent code. That is described as crossing the reliability bar for an RL reward signal, and it holds for specific task types like fault localization and code reasoning (Can structured reasoning replace code execution for RL rewards?). Note the scope. This is a verifier used to score training samples, not a simulator run while serving a user. The note also doesn't claim it works for every kind of task.

My reading is that the two settings tolerate errors differently. A reward signal averages over thousands of training samples, so a 7% error rate adds noise that training can usually absorb. At inference, one wrong simulated outcome goes straight into the answer or action the user sees. The corpus doesn't test this, so treat it as a reasoned guess. It would mean the threshold is stricter at inference than it is for rewards.

The inference-time notes point the other way, toward more real execution. In one case a stronger model built a harness that nearly doubled a weaker model's Theory-of-Mind performance mostly by moving unstable reasoning into deterministic code (Can a stronger model lift a weaker one at test time without retraining?). That swaps model reasoning for real computation, which is the opposite of swapping execution for simulation. Similarly, StateM lifts several frozen models on Terminal-Bench 2.1 (95.3% on GPT-5.6, +5.4 points on DeepSeek-V4 Flash) purely by improving the execution system around them (Can execution harnesses lift model performance without retuning weights?). And where a claim has to be trustworthy, BenchShield grounds it in recorded infrastructure evidence of what actually ran, not in a score (Can infrastructure evidence replace terminal scores in benchmark validation?). A simulated run can't produce that kind of evidence.

The plausible middle path is a hybrid, though this is my extrapolation and no note tests it. Adaptive compute allocation shows that spending effort unevenly by prompt difficulty beats spending it uniformly (Can we allocate inference compute based on prompt difficulty?). The same logic could apply here, with a cheap corrected simulator handling easy or low-stakes cases and real execution kept for hard, irreversible, or auditable ones. So the current answer is not yet for full replacement at inference time. A simulator that is good enough to train on has not been shown to be good enough to trust at serving time.