The research library.

Deterministic convergence as a substrate for organizational decisions

Agents may propose anything; only the engine may decide. A context lattice, a promotion gate, and nine invariants that make determinism a theorem rather than a hope — and an argument that this was never only about language models.

How a plan earns the right to be run

A plan is not sound because a model produced it. Six mandatory stages between intent and execution, a veto instead of a vote for adversarial review, and a discursive dilemma that shows why majority rule fails a panel of skeptics.

Provable optimization inside a probabilistic loop

Language in, certainty out. CP-SAT and HiGHS run as Suggestors alongside language models, and confidence is a certificate class — proven, feasible, heuristic, none — never a probability to be averaged with one.

Governed memory for multi-agent organizations

Retrieval-augmented generation solved the wrong half of the memory problem. Treating a recalled item as a governed proposal — not a concatenated prompt fragment — contains indirect prompt injection as an architectural consequence, not a patch.

Bounded symbolic assurance for governed decisions

Between "the model argued it is safe" and a machine-checked proof sits a wide, useful middle. An SMT solver searches for counterexamples to organizational invariants — and the discipline is refusing to let a timeout ever be recorded as a proof.

Closed-form analytics as governed Suggestors

A mean, a variance, a linear fit and a threshold answer more organizational questions than fashion admits. Closed-form analytics satisfy the substrate’s determinism, idempotency and commutativity for free — a fitted model only relative to a pinned artifact.

Fuzzy inference for organizational vagueness

A customer is "large," an exposure is "material" — vague predicates with no defensible threshold. Membership, materiality, confidence and probability all live in [0,1] and none of them compose; the compiler should say so, and refuse an inverse that does not exist.

Fitted models under governed provenance

What it takes for a fitted model to belong in a governed decision, stated plainly: the artifact must be a fact, with a hash, a dataset, a seed and an evaluation attached — and the honest line between what is built and what is only designed.

Monotone authorization as a precondition for confluence

Authorization does not belong in the kernel. A two-line counterexample shows that confluence requires policy to be monotone in the context — and a large class of real compliance rules is not, with the honest finding of what the available rewrites do and do not restore.