Architectural deep-dives, field reports from production deployments, and the things we're learning that we think other engineers and operators should know. No marketing. No thought leadership. Just notes from the bench.
The case for a substrate of constructed specialists, against the case for a single learned generalist. A walk through the architectural decisions behind v2.2 — what we kept abstract, what we exposed, and the trade-offs we'd defend in front of a regulator.
Read the essay→A treasury team had three Excel models that gave three different answers. We routed the question through Titan mode and watched the dissent.
Conventional scorecards rate suppliers in isolation. The Coupling-force aggregator does not. Here's what changed when one auto manufacturer let us look.
Cat models don't talk to macroeconomic feedback. We made them. The implications for life insurance reserves are larger than expected.
The fifth layer is the only one whose job is to attack the work of the other eight. A technical note on construction, calibration, and a few surprises.
Notes from three conversations with regulators in three jurisdictions. The DAG ended every one of them faster than expected.
Focused ⊂ Strategic ⊂ GOD ⊂ Titan ⊂ Omega is a design constraint, not an accident. Here's the architectural argument for why it has to be.
The energy domain pack runs in Focused mode at substation granularity. A backtest, the failure modes we surfaced, and the curtailment we eliminated.
The L7 memory architecture is what makes this work. A note on how episodic memory composes against semantic memory in the underwriting pass.
Five reasons, with code. The honest answer is the fifth one.