One whole architecture, composed and re-decided nowhere

Composition, not a new opinion

This engine owns exactly one judgement, and it is the one it labels loudest: how to put two model stages on one accelerator. Everything else belongs to an engine that already answers it. Hardware Match decides the platform, the node count and the per-node camera split, and that answer is copied verbatim — there is no second ranking pass here, no re-pick on confidence, and the alternatives are its ranked list, in its order, carrying its own role labels. Model Match decides the model for each stage on that platform; a stage whose model you pinned goes to Compute Fit instead. The node-level constraints come from Camera Stream Capacity, the Thermal Feasibility Checker, Carrier Finder, the Jetson Configuration Checker, Jetson Power & PSU, Frigate Fit and Drive Endurance, each stamped with the engine that asserted it, so any number on the page can be traced back to the engine responsible for it and reproduced by running that engine directly.

The pipeline sum is class E, and it caps the verdict

When two stages share one accelerator, their utilizations are summed and their memory is added on the shared memory budget. Nothing measures that. The multi-stream scaling curve in this repository models N streams of one model, and the only multi-model source on record recommends separate containers with CUDA stream priority rather than MPS and budgets about three per cent container overhead — a recommendation, not a throughput measurement, so it is cited and never turned into a number. A class E sum can prove a miss but it can never confirm a fit, which is why every multi-stage architecture is capped at NEEDS VALIDATION and each stage is also reported on its own. The measurement protocol is what closes that gap.

Where the whole-system rollups come from

Ingress bitrate is the winner’s Camera Stream Capacity network constraint per node times the node count (class on the tile). Storage is Drive Endurance on the heaviest camera split. With node-local recording every node carries the same whole drives and storage_tb is what they install; with shared recording retention lives on a central pool (storage_tb is its usable capacity) and each node keeps a buffer drive. Pool and buffer drives are separate bill-of-materials lines, and drives_total counts both. Power is module plus drive draw per node × nodes plus camera PoE from Jetson Power & PSU and the PoE Planner; the totals note names what that sum excludes. Hardware and operational cost come from Deployment Cost & TCO: registry-priced lines with per-line evidence classes (camera resolution classes stay class E planning allowances), plus an annual energy figure on a class E $/kWh rate. Hardware cost and hardware power on the architecture card are Hardware Match’s and carry its class.

Full rules, the class E sharing assumption in detail and what invalidates a result: methodology.