Eight programs.

T.A.A. exists because the honest way to start is small. One capability, one operation, one measurable question agreed in advance — proven against your numbers rather than ours. If it works, the conversation about anything larger happens from a position of evidence. If it does not, you have spent a scoped amount finding out.

ARKHAM is the governance layer over every agent we run. It defines the boundary — draft, propose, prepare on one side; approve, greenlight, release on the other — and enforces it where it cannot be talked around. The mesh is what lets agent count grow without oversight quality falling, because the oversight is structural rather than attentional.

NIGHTWATCH is the set of agents that work while nobody is watching: pulling the day's data, reconciling it against the systems of record, catching the discrepancy that would otherwise be found a fortnight later, and having an answer ready before the first person opens a laptop. Unattended work has a different failure mode to supervised work, so it is built to fail loudly and idempotently.

Most organisations discover the tender that suited them perfectly about a week after it closed. L.O.N.G.B.O.W. runs standing surveillance across public funding, grant and tender streams, matches what it finds against a partner profile, and surfaces the ones worth the effort early enough to actually mount a bid. Long-range targeting, applied to money.

IRONCLAD points independent adversarial review at our own stack and instructs it to break in. It has found real things — a boundary that shared an identifier and therefore trusted too much, a self-approval path that should never have existed. Findings are written down and published, including the ones that were our fault, because a hardening programme that only reports success is a marketing programme.

Dark Sun Protocol is the effort to make a decision against everything known rather than against whichever system happened to be open. Capture, positioning, allocation and agentic reasoning read and write one common model, so the operating picture is singular. The hard part is not the interface; it is the shared data spine that makes a single picture possible without brittle integration between every pair of systems.

An autonomous system that runs for an hour is a demonstration. One that runs for a fortnight is an operation, and the difference is almost entirely fault recovery. Pale Horizon works the long-duration problem: self-supervised navigation, drift correction without a human in the loop, and graceful degradation when a sensor stops telling the truth rather than stopping outright.

Cold Mirror builds the places where autonomy is rehearsed. The value is not photorealism; it is that the failure modes transfer — that a system which handles a doorway in simulation handles a doorway in a warehouse. Closing that simulation-to-reality gap is the entire discipline, and it is measured by how often the real world produces a surprise the mirror did not.