Phase 1 of 6
Scoping & OODA Latency Constraints
Define the C2 echelon, OODA latency budget, classification boundary, and mission authority that will govern every downstream architectural decision.
0/8
Phase Progress
Required Recommended Optional Open-Source Proprietary Trinidy
C2 Echelon & Mission Surface
Identify C2 echelon(s) the decision support must serve
Why This Matters
JP 3-0 Joint Operations distinguishes strategic, operational, and tactical levels of war, and the supporting AI decision surface is not transferable across them. A JTF J-staff consumes a synthesized COP and negotiates phasing decisions in hours; a tactical C2 node needs targeting quality data in seconds. Building a single model for all echelons guarantees it underserves every one of them. Echelon also determines which C2 systems are authoritative — GCCS-J at joint, AFATDS at fires, CPOF at Army tactical — and therefore which feeds the model has to ingest.
Note prompts — click to add
+ Which echelon is the primary user, and which echelons are secondary consumers of the same output?+ Is the model expected to span strategic through tactical, or will there be sibling deployments per echelon?+ Have we confirmed which authoritative C2 system owns each data element at our target echelon?Confirm which echelon the AI decision support pane is supporting — requirements shift by orders of magnitude across strategic, operational, and tactical.
Select all that apply
Define OODA loop latency budget for the decision surface
Why This Matters
Col. John Boyd's OODA construct is not just doctrine — it is the physical budget every decision support feature competes inside. Project Convergence 2020 famously compressed a sensor-to-shooter chain from ~20 minutes to ~20 seconds, and ABMS Onramps since 2019 have validated cross-domain data movement in seconds. An AI feature that adds 400ms of its own latency inside a 1-second kill-chain step is not a feature, it is a regression. The budget must be set before architecture.
Note prompts — click to add
+ What is our current observe-to-orient time without AI, and where do we lose the most seconds?+ Which features in the roadmap would break the budget if added, and how do we stage them?+ Have we wargamed the latency under degraded / denied comms, not only on the fiber test range?Select the end-to-end Observe-Orient latency the AI layer must hold — the budget that the commander is still inside the loop.
Single choice
Trinidy — Sub-second OODA inference requires on-node compute — a cloud round-trip to an IL5/IL6 enclave is already the entire observe-orient budget at tactical echelon. Trinidy runs the decision-support ensemble on the C2 node itself, keeping p99 predictable even when the mission partner network is degraded.
Confirm classification level and enclave
Why This Matters
DoDI 8110.01 governs the Mission Partner Environment and sets the conditions under which coalition enclaves can share decision-support data, and ICD 203 governs analytic standards in the IC. Standing up AI decision support without deciding which enclave owns the model, the weights, and the logs is the single most common reason programs stall at the ATO gate. Classification also determines which commercial foundation models are legally unusable — most of them.
Note prompts — click to add
+ Which enclave is authoritative for the model's training data, weights, and inference logs?+ Are any features planned that would force a cross-domain solution, and do we have a CDS path?+ Have we confirmed our mission partners can consume the model outputs at their classification level?Map the highest classification of data the model will touch and the enclave it must run in.
Select all that apply
Trinidy — Classified enclave RAG and inference must happen inside the enclave boundary — no data path to a commercial foundation model endpoint. Trinidy packages doctrine-trained models, vector stores, and feature stores for standalone operation on SIPR, JWICS, or coalition partner enclaves.
Declare the functions the AI decision surface will perform
Why This Matters
Every function on this list has a different data contract, latency envelope, and accountability chain. Bundling them into one "decision support" program without decomposition is how programs arrive at milestone reviews unable to describe what they built. DoDD 5100.01 functions of the DoD and the JADC2 Implementation Plan (Mar 2022) both push toward decomposed capability that composes at the data fabric — not monolithic C2 applications.
Note prompts — click to add
+ Which function is the anchor capability that the program is accountable for delivering first?+ For each function, which existing C2 system of record owns the authoritative data?+ Are any functions effectively weapons-release recommendations, and therefore in scope for DoDD 3000.09?Select the decision-support functions in scope — each carries its own data, model, and governance envelope.
Select all that apply
Declare autonomy posture under DoDD 3000.09
Why This Matters
DoDD 3000.09 (Autonomy in Weapon Systems, updated January 2023) and the DoD AI Ethics Principles require an explicit declaration of the human role for any AI that could influence kinetic effects. "Decision support" is not a safe-harbor term — if the model's output routinely drives targeting recommendations, the directive applies. Getting the autonomy posture wrong at scoping forces a full redesign at the legal review.
Note prompts — click to add
+ Has the program legal / OGC reviewed whether 3000.09 applies to our decision surface?+ Is the human decision step architecturally explicit (UX gate) or inferred from process?+ How does the system behave when the human is unavailable — does it default to safe, or to last-known?State the human role in every decision the AI surface influences — this is doctrinal, not a design preference.
Single choice
Specify deployment topology for the inference plane
Why This Matters
DoD Cloud SRG IL5 and IL6 authorize classified workloads in the cloud, but the physical layer is still in a fixed data center — a ship at EMCON, an airborne node in a contested RF environment, or a ground unit in a DIL network cannot round-trip to it. The JADC2 Implementation Plan explicitly calls for edge inference at the tactical level, and the US Army TITAN program (Palantir prime contract, March 2024, $178M) is purpose-built around deployable compute closer to the sensor. Topology decisions made after the SLA is set have 10× less leverage.
Note prompts — click to add
+ Which of our deployment locations have assured connectivity to IL5/IL6, and which do not?+ Have we characterized the DIL profile (bandwidth, latency, jitter, outages) each node must tolerate?+ Is the same model artifact deployable across fixed, mobile, airborne, and afloat nodes?Select the physical and network location where the ensemble will actually run.
Single choice
Trinidy — Afloat, airborne, and ground-mobile C2 nodes all need the same model behavior the fixed JOC sees — with no guarantee of bandwidth back to a central enclave. Trinidy is the inference substrate for disconnected, intermittent, and limited (DIL) operations; the same stack runs in the fixed JOC and on the mobile node.
Map operational planning artifacts the model must consume or produce
Why This Matters
APEX replaced the legacy JOPES construct under CJCSI 3122.06, and any decision-support system that claims to assist deliberate or crisis action planning must produce outputs that compose with APEX artifacts. A model that cannot round-trip through a FRAGO or update the commander's running estimate is a demo, not a deployable capability.
Note prompts — click to add
+ Which planning artifact is the primary hand-off — what format do we export to, and who validates it?+ Does our model need to read prior OPLAN revisions to understand commander's intent?+ How are CCIRs encoded so the model can recognize when a COP change meets a CCIR threshold?Identify which doctrinal planning artifacts the AI surface must read from or write into.
Select all that apply
Define mission-partner and coalition release requirements
Identify the release posture the outputs must support — a US-only output is a different system than a REL TO FVEY one.
Select all that apply