Governance Records as Supervision: Verifier-Selected Self-Training for Structured Workflow Repair

In the world of artificial intelligence, a model is often judged by how well it can hold a conversation or write a story. But in many real-world tasks, the goal is not to sound good, but to be correct according to a strict set of rules. Imagine a robot that must move blocks on a table to build a tower. It might say the right words, but if it tries to lift a block that is already holding up another one, the plan fails. The failure is not a matter of style; it is a mechanical error that a computer can check instantly. For years, researchers have used these checks only to reject bad plans and try again, wasting time and energy. A new line of inquiry asks a different question: if a computer can spot a mistake and find a correct solution, can it use that record of success to teach the model to get it right the first time, without needing to think so hard next time?

This question lies at the heart of recent work by independent researcher Jesus Salas, who explored whether the digital "paper trail" of a successful task could become a teacher. The study focuses on a specific type of problem where the rules are clear and the outcome can be verified by a machine. The researcher used a well-known test environment involving moving blocks, where a computer program acts as a referee to decide if a plan works. The goal was to see if a large, expensive AI model that sometimes solves these puzzles could generate a few correct answers, and if those specific answers could be used to train a smaller, faster version of the same model to solve them reliably on its own.

The experiment began with a large AI model known for its ability to "think" before it answers. This model was given a series of block-moving puzzles. It did not get the right answer every time, but in a few dozen attempts, it managed to produce plans that the referee program accepted as valid. The researcher took these rare, successful plans and used them as a training set. The goal was to teach the same AI model, but this time in a mode where it does not spend time thinking or reasoning, to produce those same correct plans immediately. The model was not given the answers in advance; it only learned from the few times it had stumbled upon the right solution on its own.

When this trained model was tested on a fresh set of eighty new puzzles it had never seen before, the results were striking. The untrained model, when asked to solve the puzzles without thinking, succeeded only once. The model that had been taught to "think" before answering succeeded thirty times. But the model that had been trained on the few successful plans from the thinking version succeeded fifty-seven times. It did not just get better; it became substantially more reliable on this specific test. Furthermore, this trained model solved the problems much faster and used far fewer computer resources than the version that spent time reasoning. It was able to produce a valid plan in 57 out of 80 cases, whereas the original thinking version did so in 30 out of 80 cases. While the trained model was highly efficient, the study noted that the specific goal of proving it fixed interface errors better than the thinking model was not met in the preregistered test, even though the trained model remained valid in its output format for all cases.

To ensure that this improvement came from the quality of the answers chosen for training, and not just from having more examples, the researcher ran a second test. They took a pool of valid plans generated by the model and split them into three groups. One group was chosen by the referee program, one by the model itself picking blindly, and one by a simple rule that picked the first available option. The model trained on the plans chosen by the referee program performed significantly better than the one trained on the plans the model picked for itself. This proved that the referee's judgment was the key factor. The model was not just learning from any success; it was learning from the specific kind of success that an independent checker had confirmed was truly correct.

The study also explored what happens when a much smarter AI model acts as the teacher. In this scenario, a powerful reasoning model generated the correct plans, which were then used to train a smaller, less capable model. The smaller model improved dramatically, moving from solving only two puzzles to solving fifty-one out of eighty. However, the researcher was careful to distinguish this from the first experiment. In the first case, the model taught itself using its own rare successes. In the second, it learned from a superior teacher. Both methods worked, but the first one showed that a model could improve itself without needing a smarter friend, as long as it had a way to verify its own work.

The research also looked at the limits of this approach. The improvements were real and measurable, but they were not infinite. When the researchers tried to train the model over and over again on its own new successes, the gains eventually stopped. The model reached a point where it could not learn any more from the available data, suggesting that there is a ceiling to how much a model can improve on its own without new information. The study also noted that while the trained model was excellent at following the specific rules of the block puzzles, it did not become a general expert at all types of planning. It became a specialist for the specific rules it was taught.

This work suggests a new way to build AI systems that are both smart and efficient. Instead of relying on a massive, slow model to think through every problem, we can use a large model to find a few correct solutions, verify them, and then train a smaller, faster model to replicate that success. The smaller model becomes a specialist that can handle routine tasks quickly and accurately, while the larger model and the referee program stand by for the difficult cases that require deeper thought. The research shows that the record of a successful attempt is more than just a log; it is a valuable resource that can be turned into a teacher, allowing machines to learn from their own occasional brilliance and make it a consistent habit.