Why the second QANTIS paper follows the next decision
A method can work once and still struggle when used again. The second QANTIS preprint follows that question through a sequence of decisions.
A machine rarely gets the whole picture at once. It makes an estimate from what it has seen, receives another observation and changes its estimate before deciding what to do. In a sequence of decisions, that updated view becomes the starting point for the next one. An error can travel along with it.
This is the question that stayed after the first QANTIS study. We had investigated probability updates on quantum hardware, including an initial closed-loop Tiger example. I wanted to follow the repeated use more closely: would the information remain helpful when the planner had to live with the previous answer, again and again?
The second preprint, arXiv:2607.06760, was submitted on 7 July 2026. It follows that narrower question through sequences of updates inside a classical planner. The important part of the story is what the planner did with the answer, not only how closely two sets of probabilities matched.
What the first paper gave us
The foundational study, arXiv:2603.00785, ran 45 experiments across three IBM Heron backends. In a Tiger belief-oracle case, a rare-observation probability rose from 0.179 to 0.907, while the posterior stayed at a Hellinger distance of 0.0015 from exact Bayes. Those are results for the stated experiment conditions.
The first paper also includes an initial closed-loop Tiger example. The follow-on examines repeated service use in a more focused comparison inside a classical planner. Its main campaigns follow 8-step and 12-step trajectories; controls extend to 20 and 32 steps to explore the operating range.
A changed probability does not always change the action
At every checked decision point, the hardware-derived and exact-Bayes posteriors selected the same immediate action. This is the result I find most useful to discuss. It describes the effect on a decision, rather than only the distance between two distributions.
The two measures answer different questions. A distribution can shift without changing which action looks best. A small shift near a decision boundary can do the opposite. Following both helps us see what the system actually depends on.
What we can take from it
The planner and permission to act remain classical. The quantum processor handles a defined subtask, with a classical path available if it cannot contribute. This makes it possible to study the component without handing the whole decision process over to it.
Both papers are first-author preprints and have not been peer-reviewed. Neither claims a wall-clock advantage over strong classical methods, general scaling or production readiness. The public repository is an explanatory edition, not the full experimental tooling. A reader can use the work to judge whether the tested conditions are relevant to their own problem.
- Update the probabilities
- Let the classical planner choose
- Use the new observation in the next step