Which platform, at how many nodes, and why that one?

One feasibility gate

Step 1. Every catalog platform is evaluated at every node count from 1 to the cap through the engines that own each fact. A candidate is eliminated only on a hard FAIL or UNSUPPORTED, or when node count × unit power exceeds a power ceiling you stated — and a check only FAILs on evidence that can prove the miss (elimination authority). Budget is not part of this gate — over-budget candidates stay in the ranked list. There is no × 1.5 stream tolerance, no preview penalty and no interpolated-benchmark multiplier in elimination.

One admission policy

Step 2. Non-eliminated candidates are classified by recommendation evidence: recommendation-eligible or planning-only. Capacity evidence below the published floor, unresolved validation constraints, or a miss that rests on planning-only evidence (a multi-stream or power-mode estimate, a host CPU budget, a class E band) keeps a candidate planning-only even when its score is high.

Step 3. Planning-only candidates are not dropped or rescored. They keep their numerical scores and are listed after every recommendation-eligible candidate: modelled fit first (the tightest modelled check at or below its capacity), then presentation verdict, then score. A planning-only board modelled over its capacity is sized at the fewest nodes that bring the modelled load within capacity, where the node cap allows it.

One score function

Step 4. Every survivor is scored once as Σ wᵢ sᵢ over five terms in [0, 1] — margin, evidence, cost, power and simplicity — using the published weight row for your goal and the fixed cost and power anchors (camera count × $100 and camera count × 2 W for policy version v1).

Step 5. Run a match to show the engine’s published lexicographic ordering on your result.

What the engine does, in detail

Hardware Match answers which complete hardware configuration should deploy for this workload and these constraints, at how many nodes, and why that one. Every complete candidate configuration is evaluated through the domain engines that own its constraints. A candidate is eliminated only by an established hard FAIL. Non-eliminated candidates are then classified by recommendation evidence: recommendation-eligible or planning-only. Recommendation-eligible candidates are ranked once using the published five-term Hardware Match score against fixed published anchors. Planning-only candidates remain visible and are ranked separately — modelled fit, then verdict, then the same score. The highest-scoring recommendation-eligible configuration is the winner. Decode, inference and camera-pipeline working memory are evaluated through Camera Stream Capacity. Model-centric standalone memory checks use Model Memory Fit where applicable. Thermal, configuration, carrier, power and PoE checks are delegated to their owning engines. No hidden ecosystem rating enters the score and no silent re-pick is applied.

One authoritative winner

Step 6. The winner is the highest-scoring recommendation-eligible configuration in that order. Alternative roles (lower cost, lower power, more headroom, stretch) are labels on the same ranked list, not a second selection pass. Every term, weight, input and total on the result is reproducible from scoring_provenance on each ranked row.

Full rules, the class E weight table and what invalidates a result: methodology. Related: Camera Stream Capacity is where the decode and inference numbers come from; Thermal Feasibility Checker owns the enclosure question; Jetson Power & PSU sizes the supply once the platform is chosen; Carrier Finder finds the board it plugs into.