Platform class is the single largest determinant of allowable model footprint, thermal envelope, and update cadence — an Orin NX 16GB wearable and an AGX Orin vehicle compute have an order of magnitude difference in inference capacity. Replicator Initiative Tranche 1 and 2 selections span all-domain platforms, and a model architected for one class rarely ports cleanly to another without quantization and distillation work. Platform mix also determines whether you need a single unified model or a family of size-tiered variants sharing a common training pipeline.

Mission function determines which DoD policy regime applies — a perception model for navigation sits under LOAC and DoDD 3000.09 differently than a targeting-support model, and a sensor-to-shooter function triggers the autonomy-in-weapons review path under the January 2023 update to DoDD 3000.09. JP 3-0 Joint Operations and joint targeting doctrine require a clear human-machine role definition long before the system is fielded. Mixing mission functions in a single model almost always forces the program to the most restrictive review path, so scoping is also risk management.

Latent AI's LEIP platform under Project Linchpin measured 3× faster inference on Jetson AGX Orin vs. an unoptimized baseline, and the Army reported a 70% improvement in effective decision speed when sensor data is processed at the edge rather than cloud. A cloud round-trip in a degraded SATCOM environment is not just slow — it fails entirely, which makes the cloud latency number a survivability metric, not a performance metric. The latency budget has to be allocated across feature extraction, inference, fusion, and post-processing before any single stage is over-engineered.

SWaP is the non-negotiable ceiling that determines whether a model class is feasible at all — a dismounted warfighter carrying Orin NX at roughly 1kg inclusive of battery cannot absorb a 200W AGX-class workload regardless of software optimization. The 15–75W envelope typical of rugged embedded GPU deployments (Jetson Thor/Orin family) forces quantization, pruning, and distillation into the model development plan from day one. Programs that defer SWaP analysis until integration almost always discover a 2–3× over-budget draw that forces a model redesign at the worst possible time.

The CJADC2 architecture explicitly assumes disconnected, intermittent, limited (DIL) operating conditions, and any edge AI that depends on connectivity for inference becomes a single point of failure in exactly the conflict it was fielded for. Near-peer EW capabilities deny SATCOM and terrestrial links with high confidence — Replicator Initiative platforms are being selected on the basis of autonomous operation under these conditions. Assuming a connectivity floor that won't exist in combat is the fastest way to field a system that fails its first operational test.

DoD Cloud Computing SRG Impact Levels (IL2 / IL4 / IL5 / IL6) constrain where workloads may physically run and who may touch them, and a mismatch between the model's training environment and its deployment environment is an authorization blocker that surfaces at ATO time, not at architecture time. ITAR (22 CFR 120-130) and EAR (15 CFR 730-774) add export-control obligations that can invalidate otherwise-clean cloud deployments if foreign-persons access is possible. Sovereignty is a design input, not a documentation exercise.

Training data classification propagates to the weights — a model trained on SECRET imagery inherits SECRET classification, and the weights themselves become classified material that must be handled accordingly. DFARS 252.204-7012 requires covered contractors to protect CUI at a minimum of NIST SP 800-171 controls, and CMMC 2.0 Level 2 aligns to that control set with third-party assessment for prioritized programs. Programs that discover weight classification after training has already happened in a mismatched enclave have restarted from scratch — this decision must be explicit, not inferred.

The DoD AI Ethics Principles (adopted February 2020) require AI systems to be governable — humans must be able to disengage or deactivate, which is a design constraint, not a policy wrapper. DoDD 3000.09 (January 2023 update) governs autonomy in weapon systems and sets a senior-review requirement for systems that select and engage targets without human input. The Political Declaration on Responsible Military Use of AI and Autonomy (February 2023) reinforces the human-oversight expectation internationally. Picking the wrong posture triggers a fundamentally different review path.