Games and Theory

Platform

The Agentic Conflict Engine (ACE)

ACE is commercial software for AI-powered conflict simulation and strategic analysis. Defense commands, intelligence agencies, commercial strategy teams, and research institutions use ACE to test plans against adversaries that reason from assigned personas, objectives, and constraints, across hundreds to thousands of scenarios.

The status quo this replaces is expensive and slow: analytical and Title-10-class wargames commonly run hundreds of thousands to millions of dollars per game and take months to plan, organize, and execute (Jensen, Atalan, and Tadross, CSIS, 22 Feb. 2024)[CSIS 2024].

Demonstrated tempo: one 50-game corpus of complete 3-turn games, generated in about 33 wall-clock minutes at concurrency 50, per-game mean 24.7 minutes.

The traditional planning timeline shown above for comparison follows the cited CSIS framing.[CSIS 2024]

How It Works

  1. Define the question. A decision-maker states what they want to learn. ACE translates that into testable hypotheses.
  2. Design the experiment. ACE builds a structured exploration of the decision space using rigorous design-of-experiments methods: full replication ensembles, factorial stress arms, or controlled single-perturbation excursions run many times, whichever design resolves the question.
  3. Run the simulation. AI agents representing different actors execute strategies across hundreds to thousands of parallel scenarios. Each agent decides based on its assigned persona, objectives, and constraints.
  4. Analyze the results. ACE compares outcomes to show which strategies hold up and which break down, with statistical confidence across the full scenario set.
  5. Deliver findings. Automated reporting traces every result back to the original question. Output is briefing-ready, not a data dump.

Example key question

Does the plan hold if the adversary moves first?

Hypotheses tested against it

HYPOTHESIS 1
Option A holds across the adversary set.
Supported
HYPOTHESIS 2
The timing shift improves the median outcome.
Not supported
HYPOTHESIS 3
Constraint C binds before the decision point.
Condition not met

One key question decomposes into testable hypotheses; each fills with run progress, then a verdict stamps. Verdicts that fail stay on screen. Synthetic content, illustrative.

Key Capabilities

  • Designed experiments: replication ensembles, factorial stress arms, controlled excursions; outcome distributions, not single narratives
  • AI cognitive surrogates agents that reason based on assigned personas and strategic objectives
  • Massively parallel execution hundreds to thousands of scenarios running simultaneously
  • Hypothesis-driven analysis every simulation answers a specific question
  • Automated reporting briefing products generated directly from simulation results
  • Demonstrated tempo: one 50-game corpus of complete 3-turn games, generated in about 33 wall-clock minutes at concurrency 50, per-game mean 24.7 minutes
  • Defense, intelligence, commercial, and academic applications

Ask the record

The Retrieval Moment

A question drops into the field of decision points recorded across past exercises. Points that resemble the question light up, and the answer assembles with a provenance line back to each lit row.

Interactive: try it

How to read it: dim points are decision points recorded across past exercises. Gold points are the ones that match the selected question. Every gold point is traced by a line to the assembled answer, and the answer panel lists how many matched.

Select a question to assemble an answer.

Synthetic content, illustrative. The field, the similarity scores, and the ledger references are placeholders. Nothing here depicts a real corpus finding.

Philosophy

Surrogate, Not Twin

We don't claim to clone a specific adversary. We map how that kind of adversary can plausibly behave, then study the distribution.

Provenance

Two Records, One Timeline

Every run keeps a typed state record and the agent dialogue that produced it, in lockstep. Hover either stream to highlight its counterpart.

State record

t=004 · posture: hold · reserve: uncommitted
t=007 · posture: hold · indicator: 1 of 2
t=011 · posture: shift · reserve: committed
t=014 · effect: contested · supply: degraded
t=018 · posture: consolidate · reserve: spent

Agent dialogue

"Hold until the second indicator."
"One indicator is not enough. Wait."
"Second indicator. Commit the reserve."
"Supply is degrading. Expect the position to be contested."
"Consolidate. Nothing left in reserve."

Abstract placeholder content, illustrative.

Deployment

  • Unclassified platform. ACE is designed and maintained as entirely unclassified. Classification enters only at the data layer, when customers load their own scenarios and operational plans.
  • Deployable on any network (unclassified, CUI, or classified) with appropriate LLM endpoint configuration.
  • Object code delivery. Deployed as compiled software. Source code available for security review under controlled conditions.
  • Model-agnostic by design. ACE orchestrates large language models through configurable endpoints; customers select and provide the model endpoints appropriate to their environment (commercial API, government-hosted, or on-premise), configured with G&T at deployment. The platform ships no models of its own.
  • Container assurance. Working toward DoD Impact Level 4 (IL4) deployment readiness: hardened container packaging with a controlled image-promotion pipeline, SBOM per EO 14028, and immutable build and approval audit records. Compatible with customer-side accreditation uplift to IL5 and IL6 and with deployment in customer-accredited environments above IL6; the customer performs and owns higher-impact-level accreditation.
  • EAR99. No ITAR restrictions. No export license required for Five Eyes or NATO destinations.

Responsible Use of AI

Published research has raised real concerns about escalation bias when large language models act as autonomous decision-makers in crisis simulations. ACE does not use models that way: it uses them as a distributional instrument, generated across independent providers, with every action provenanced to a ledger row, which is the mitigation that research points toward, not away from. (Rivera et al., FAccT 2024; Atalan, Reynolds, and Jensen, CSIS, Feb. 2025.)[Rivera et al. 2024][CSIS 2025]

Conceptual diagram, illustrative. No figures are shown.

Patent Protection

ACE is built on patent-pending methods invented by G&T's founders at the University of Colorado Boulder. See Patents for details.

Sources
  1. Jensen, Benjamin, Yasir Atalan, and Dan Tadross. Center for Strategic and International Studies, 22 Feb. 2024. (Cost and timeline of analytical and Title-10-class wargames.)
  2. Rivera et al. ACM Conference on Fairness, Accountability, and Transparency (FAccT), 2024. arXiv:2401.03408. (Escalation risks from language models as autonomous decision-makers.)
  3. Atalan, Reynolds, and Jensen. Center for Strategic and International Studies, Feb. 2025. (Escalation-bias concerns in LLM crisis simulation.)

Consolidated citations for the whole site: References.

Every game is fictional. Every number is real.