Relational Viability Under Incomplete and Changing Knowledge
The programme written up: the formulation, the propositions, the counterexamples, and an audit of what the earlier experiments actually showed.
A preprint, deposited openly and citable by DOI. It is a document of a research programme and a formal synthesis of it. It is not a validated new physical law, and it does not establish a new theorem or a performance advantage over good robust control — the paper says so itself, in those terms.
- The augmented state and the parameter table — every symbol, and how each one has to be established before it can be used
- The functional, and what its sign does and does not mean
- The conservative one-step drift certificate, with latency handled explicitly
- Four propositions with elementary proofs, each labelled as a restatement where it is one
- Six counterexamples to the strong hypothesis
- An audit of the earlier experiments, including the ones that found nothing
- Five falsifiable tests that would have to pass before any stronger claim is defensible
Of the earlier studies: one matched a reference solver on eighty small decisions in a particular binary model. One found the adaptive heuristic slower than ordinary interval splitting. One found no advantage over optimal robust planning at all. Two reported speedups whose value depends entirely on amortising setup costs that may never be amortised.
Those are reported as what they are. They are historical exploratory runs, not one pooled independent experiment, and I have not re-run every one of them.
The hypothesis, the name and the research direction are mine. The draft, the mathematical restatement, the code checks and the presentation were prepared with AI assistance. No peer review and no independent hardware test is claimed.