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 isn't 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's 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 doesn't 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 doesn't 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 isn't, with the honest finding of what the available rewrites do and don't restore.

Formulating an Initiative Trajectory Model, and why training waits

No trained model here. A formulation: why the right target is coordination, not people, why a Joint-Embedding Predictive Architecture is stage three of four rather than stage one, and why four independent adversarial reviews are the reason training is postponed rather than reported.

Individual and group failure modes an ever-learning organization is built to close

My previous report stayed above the level of individual people on purpose. This one takes up what that one declined to touch: prediction error is generated by people, as individuals whose judgment either engages or quietly surrenders, and in groups that reliably fail to say what they collectively already know. I name both failures precisely enough to detect without diagnosing a person, and argue a well-built ever-learning organization has no room for either to persist.