Ongoing

CADE — Combined Arms Decision Exercise

AI Operator / Design Authority · 2026 · Capstone development cycle · 2 people · 6 min read

Designed and built a battalion staff decision exercise using Claude, ChatGPT, and Gemini as governed production engines. A two-person team produced the first executable version in one week. CADE has run three times with an average of 19 participants per session — no comparable exercise format existed for this context.

Overview

A two-person team built a battalion-level combined arms decision exercise in one week using AI as a governed production engine. I served as design authority: I framed the training problem, set success criteria, approved all source material, and decided what changed after each live execution. Claude, ChatGPT, and Gemini accelerated research, drafting, and artifact revision. CADE has run three times with an average of 19 participants per session. The scenario is based on a real operation — relieving the remnants of a brigade trapped behind enemy lines — which gave the staff authentic operational context to work inside.

Problem

Battalion staffs need a repeatable way to practice combined arms decision-making when training time is compressed, doctrinal baselines vary across participants, and language friction slows coordination. JCATS and DXTRS require days to configure and trained operators throughout. CPX requires preparation runway that compressed events don't have. CADE was designed to deliver decision-quality staff training without any of that overhead.

Constraints

  • Training time is compressed, so preparation overhead cannot consume the event.
  • Participant proficiency and doctrinal familiarity may vary across the staff.
  • Mixed-language delivery requires plain-language phrasing and visual-forward support.
  • Controllers need runtime products that work under pressure without designer dependency.
  • AI output must remain subordinate to source truth and human design authority.

Approach

I used AI as a governed production engine. Human design authority set the training problem, success markers, turn structure, adjudication logic, and controller discipline. AI accelerated research, drafting, artifact generation, and revision inside a workflow anchored to OPORD-quality source truth.

AI Operator Skill Demonstrated

  • Framed the training problem before using AI.
  • Constrained AI output with source truth and acceptance criteria.
  • Separated AI production from human design authority.
  • Converted live execution evidence into product changes.
  • Designed artifacts for human runtime usability, not just document completeness.

Turn Rhythm

01

Situation

Establishes the operational update and decision context for the turn.

02

Clarification

Allows the staff to resolve essential uncertainty before deliberation begins.

03

Deliberation

Forces cross-functional coordination and tradeoff discussion inside a time box.

04

Decision

Requires explicit commitment instead of deferral or open-ended discussion.

05

Brief to CDR

Turns staff reasoning into a commander-facing recommendation.

06

Adjudication

Applies consequences and sets conditions for the next turn.

Product Architecture

01

Source-governed scenario package

Maintains scenario narrative, operational graphics, control measures, and annex logic as coherent source truth.

Controls AI drift by forcing every downstream product to trace back to an approved source layer.
02

Turn engine

Defines the decision rhythm, phase transitions, time boxes, and required staff commitments.

Turns training intent into repeatable workflow logic that AI can support without changing the exercise purpose.
03

Adjudication model

Connects observable integration behavior to outcome bands so consequences are credible and comparable.

Converts qualitative staff behavior into structured decision support without outsourcing judgment to vague impressions.
04

Controller delivery package

Consolidates execution logic, role aids, decision prompts, timing cues, and review support into one runtime artifact.

Designs AI-produced artifacts for real human use under time pressure.
05

Learning capture layer

Structures turn-level review so reasoning, expected outcomes, coordination gaps, and next adjustments are captured.

Closes the loop between AI-assisted production, live execution, and evidence-driven iteration.

Key Decisions

Use AI as the production engine, not the design authority

Reasoning:

AI made speed possible, but the exercise required human judgment to define the training problem, approve source material, set success criteria, and decide what changed after execution.

Alternatives considered:
  • Let AI generate the exercise structure directly
  • Use AI only for editing and formatting

Anchor all products to OPORD-quality source truth

Reasoning:

A single authoritative scenario layer reduced drift across orders, turn materials, controller prompts, and review products.

Alternatives considered:
  • Allow each artifact to evolve independently
  • Use loose narrative summaries as the source layer

Move from fragmented runtime artifacts to a unified Controller Package

Reasoning:

Controllers need fast, reliable access under pressure. Consolidating execution logic, role aids, decision prompts, and review structure made the framework more portable.

Alternatives considered:
  • Keep the runbook as the center of gravity
  • Use separate support products for each controller function

Use deterministic adjudication bands tied to observable behavior

Reasoning:

Controller-to-controller variance weakens outcome credibility. Observable bands reduce discretionary drift while preserving human oversight.

Alternatives considered:
  • Let controllers adjudicate primarily by judgment
  • Use fully scripted outcomes detached from staff behavior

Use plain-language and visual-forward delivery

Reasoning:

CADE must work when language friction is present. Plain wording and visual aids reduce avoidable misunderstanding without diluting the training logic.

Alternatives considered:
  • Keep doctrinal language dense and assume translation will solve it
  • Simplify the exercise itself instead of improving delivery aids

Tech Stack

  • Claude
  • ChatGPT
  • Gemini
  • Markdown
  • PPTX/HTML briefing workflows
  • OPORD-grounded planning workflows

Result & Impact

  • 3
    Live executions
  • ~19 per session
    Avg participants
  • 1 week
    Time to first executable version
  • 4 of 4 on first execution
    Validation criteria met
  • 5 documents, zero simulation infrastructure
    Controller Package

CADE is the first exercise of its kind for this context. No equivalent format existed that could deliver decision-quality battalion staff training inside a compressed window without simulation infrastructure. AI made a one-week production timeline possible. Human governance made the result operationally credible.

Evidence Signals

Staff operated inside the structure

Confirmed across all three executions. S1, S2, S3, S4, XO, Medical, Engineer, Fires, and additional staff positions operated inside the framework without designer support during turns.

Controllers executed with delivered artifacts

Confirmed on first execution. Identified fragmented artifacts as a failure point — led directly to the unified Controller Package architecture in subsequent runs.

Consequences created meaningful decision pressure

Confirmed. Resource depletion, casualties, timing constraints, and degraded options from earlier turns visibly shaped staff behavior in later turns.

Review captured learning before the next cycle

Confirmed. Protected AAR windows held across all sessions. Execution observations fed directly into the v3.0 redesign.

Where CADE Fits

  • Training time is compressed and setup overhead must stay low.
  • The primary objective is decision behavior under pressure.
  • The staff needs to practice cross-functional integration, not isolated section work.
  • Language friction or uneven proficiency is present or anticipated.
  • The event owner needs a repeatable format that can be facilitated from a Controller Package.

Designed Against Failure Modes

Controller discipline degrades without the designer present.

Role boundaries, pre-execution guidance, and the Controller Package keep facilitation behavior constrained.

Adjudication becomes inconsistent across controllers.

Outcome bands and observable behavior categories reduce discretionary variance.

Scenario details drift across AI-generated artifacts.

OPORD-quality source truth governs every downstream product.

Language friction slows decision windows.

Plain-language phrasing and visual-forward aids reduce avoidable comprehension burden.

Consequences fail to carry forward.

Persistent state tracking keeps resources, timing, casualties, and option degradation connected across turns.

Boundaries

  • CADE does not replace high-fidelity simulation when technical modeling is the primary objective.
  • CADE does not certify doctrinal competency or grade students on a pass/fail basis.
  • CADE does not fully automate design judgment through AI.
  • CADE is strongest when the desired outcome is decision behavior under constrained conditions.

Learnings

  • AI is most useful when the operator defines the problem, boundaries, and acceptance criteria before generation begins.
  • Source truth is the control mechanism that keeps a multi-artifact AI workflow coherent.
  • A training product is not finished when the documents look good; it is finished when controllers can use it under pressure.
  • Execution evidence is more valuable than document-only review for improving CADE.