What we protect
A filed foundation patent grounds the program; a companion family extends it; one mechanism is held back. Every principle below is implemented and running in DiaCroma™ today.
The foundation — Admissibility-First Decision Control
The backbone, with priority secured across 35 claims: safety enforced as a wall before any scoring, the gates that define what’s allowed, typed answers, and a core that learning can’t move.
Forbidden options are removed before any scoring runs — because a finite penalty can never hold a hard limit against a big enough reward, so “allowed” has to be a wall, not a price.
Request the paper →Feasibility, safety, law, and legitimacy as four mathematically distinct yes/no tests — not one check repeated four times; only their conjunction can hold all four at once.
Request the paper →Follow-through is capacity × willingness × coordination — multiplicative, so the weakest factor caps the whole result, and you repair the one that’s limiting instead of asking for more effort.
Request the paper →When there is no single right answer, the response is typed — act, ask, a frontier of options, or an honest refusal — and every one, refusals included, is governed and reproducible.
Request the paper →What may learn from experience is separated from what must stay fixed — the line between a system that gets better and one that quietly gets bolder past a hard limit.
Request the paper →The refinements that extend it
Four results that sharpen the foundation — how belief, habit, honesty, and human limits are handled — each already running in the product, each the subject of a paper you can request.
How to read uncertain, stale, or low-quality evidence so the system’s confidence never outruns what the data actually support — a prerequisite for every gate that acts on belief.
Request the paper →Why steady repetition, not intensity, is what makes a behavior durable — and what a system must understand about a person’s state to support that rather than undermine it.
Request the paper →What it means, formally, to be honest about a recommendation: the frontier of options when the trade-off is genuinely unsettled, and confidence that matches how often the event actually occurs.
Request the paper →Constraints that must always hold, written into the mathematics itself — a fixed projection the optimizer and the learner can’t reach, reprice, or argue with. What most systems treat as policy, this makes a theorem.
Request the paper →The time-and-causation layer
How the system reasons about when a plan runs out and what actually caused a change — the layer that keeps feasibility and credit honest over time.
Feasibility runs out when the first budget — time, money, energy, attention, others’ cooperation — is exhausted, not the average; and credit for an action subtracts the improvement that would have happened anyway.
Request the paper →The held mechanism
One mechanism is deliberately not described. It is what keeps an agent on-mission across a whole trajectory — catching the slow, sustained drift that per-step checks miss. It opens to request once filed.
Governs the whole trajectory against the original mission, so an agent can’t quietly drift off the job — the failure that sinks ordinary AI agents. Held under a dedicated filing; the mechanism is deliberately not described.
Held · not described