Phase 1 of 6
Scoping & EMS Mission
Define the electromagnetic spectrum mission, latency envelope, SWaP constraints, and classification posture that govern every downstream architectural decision.
0/8
Phase Progress
Required Recommended Optional Open-Source Proprietary Trinidy
EW Mission & EMS Domain
Identify EW mission sets in scope
Why This Matters
JP 3-13.1 Electronic Warfare and JP 6-01 Joint EMS Management define ES, EA, and EP as distinct mission sets with different latency envelopes, different classification caveats, and different approval authorities. A model that fuses all three into one monolithic inference path invariably loses to a specialist against a thinking adversary, and the approval chains for EA effects are very different from ES collection. Treating EMBM as an afterthought is the most common mistake — the 2020 DoD EMS Superiority Implementation Plan makes it the coordinating function.
Note prompts — click to add
+ Which mission set is the primary use case, and which are adjacent capabilities we plan to reuse features for?+ Do we have the correct approval authority chain for EA effects versus ES-only deployments?+ How will this capability plug into EMBM so it contributes to, rather than fights, joint EMS C2?Confirm which of the three JP 3-13.1 EW divisions your model must support.
Select all that apply
Define response-loop latency budget
Why This Matters
Modern frequency-hopping radar and LPI/LPD waveforms change parameters inside a millisecond — any architecture that includes a network hop to a datacenter has already lost before the inference even runs. DARPA SC2 demonstrated that AI-driven cognitive radio dramatically outperforms rule-based spectrum management, but only because the whole loop sits at the RF front end. A stated latency budget is what forces the hardware conversation early, before an infeasible architecture is committed.
Note prompts — click to add
+ What is the tightest loop on the platform today, and is it bounded by physics (TOA) or by software?+ Have we budgeted RF front-end digitization, feature extraction, inference, and PA keying separately?+ Is our worst-case jitter (not average latency) inside the pulse repetition interval we must react within?Select the end-to-end EMS reaction budget the inference path must hold, emitter-edge-to-effect.
Single choice
Trinidy — Any network hop introduces milliseconds of jitter — incompatible with sub-ms cognitive EW. Trinidy runs classification on airborne/on-platform FPGA fabric with deterministic pipeline latency, and decision models on co-located GPU, eliminating the network from the reaction loop entirely.
Define SWaP (size, weight, power) envelope
Why This Matters
SWaP is not a secondary constraint on airborne and dismounted EW — it is the primary constraint. A model that needs a data-center GPU to hit sub-ms latency will never fly on a fighter or a Group 3 UAS, and retrofitting a smaller model later almost always sacrifices the accuracy that justified the program. The L3Harris Mid Band Array and the BAE Systems Compass Call EC-37B are examples of fielded envelopes — target one of them or a smaller pod class from day one.
Note prompts — click to add
+ Is our target platform a retrofit on existing aperture/power or does it justify a new MIL-STD-704 power envelope?+ Have we validated thermal dissipation at sustained classification load, not peak inference?+ What is the weight budget for the inference card, and does it include cabling, shielding, and TEMPEST?Select the platform SWaP class the inference substrate must fit inside.
Single choice
Establish classification and compartment posture
Why This Matters
Signal collection and EW capabilities are among the most highly classified military capabilities, and a model trained on SIGINT-derived signal libraries inherits the classification of the training data. Models trained at TS/SCI with SI/TK cannot be fielded on a collateral platform without a sanitization that usually destroys the model's value. DFARS 252.204-7012 and NIST 800-171 are the floor for the unclassified CUI side of the pipeline; the classified side is governed by the owning ICD/SCI authority.
Note prompts — click to add
+ Where does our training data's highest caveat originate, and does our intended platform hold that caveat?+ Is there a sanitized/releasable variant of the model on the roadmap, and is it planned from day one or bolted on?+ Who is the accreditation authority (AO) for the deployment enclave and have we engaged them pre-architecture?Map the enclave classification and compartment caveats the inference and training environments must hold.
Select all that apply
Select target deployment platform family
Specify the platform type that will host the inference substrate.
Select all that apply
Trinidy — Trinidy supports airborne FPGA fabric (sub-1ms deterministic inference), on-platform GPU for decision models, and ground-station racks on the same deployment toolchain — so a capability prototyped at the ground station can be promoted to the airborne pod without reimplementation.
Define effects-authorization & ROE posture
Why This Matters
DoD AI ethical principles and EW-specific ROE both require a documented authorization posture — it is not an implementation detail. EP / self-defense is the one area where most operational concepts permit autonomous reaction because the adversary timeline (sub-ms) precludes a human; EA against other emitters almost always requires authorization because of fratricide and escalation risk. Specifying this up front drives whether the model is a classifier-plus-recommender or a full policy network.
Note prompts — click to add
+ Which effects are pre-authorized today and which require a human vote at the moment of action?+ Is our reaction model a policy that emits actions, or a classifier that feeds a deterministic effects library?+ How is the authorization state logged so that post-mission review can reconstruct every decision?Confirm the authority chain for EA effects and the degree of model autonomy permitted.
Single choice
Map fratricide and deconfliction constraints
Why This Matters
Fratricide in the EMS — jamming a friendly radar, datalink, GNSS receiver, or allied emitter — is the single most common and most expensive EW failure mode, and it is almost always caused by a model operating on a stale or incomplete friendly-force picture. DoDI 8320.05 specifically addresses spectrum data sharing so EW systems do not have to rediscover friendly emitters in the field. The model must treat the JRFL and ACO spectrum annex as hard constraints, not soft features.
Note prompts — click to add
+ How fresh is our JRFL/ACO feed at the inference node, and is staleness a hard block on EA effects?+ Have we red-teamed a fratricide scenario where the deconfliction feed is degraded or spoofed?+ What is the fallback posture when deconfliction data is unavailable — deny EA, or proceed on last-known list?Define friendly-force spectrum deconfliction inputs the model must respect at decision time.
Select all that apply
Define mission success and measures of effectiveness
Select the primary measures of effectiveness the capability will be evaluated against.
Select all that apply