2026_05_07_DISCORD_ProposersDay_final__-__Read-Only.pdf

Approved for Public Release, Distribution Unlimited

Eric Davis, DARPA STO 2026/05/07

DARPA DISCORD

Disruption through Intelligent Strategies, Counter Options, and Resilient Defenses

OUSW(R&E) Ender’s Foundry

Approved for Public Release, Distribution Unlimited 2

Schedule of Proposer’s Day

EventEndStart

Security and Logistics9:4009:30

DISCORD Overview10:209:40

Stakeholder Brief10:3510:20

Stakeholder Brief10:5010:35

Proposal Submission Guidance11:0510:50

Submit questions / LUNCH12:3011:05

DARPA Connect12:4512:30

Lightning Talks15:1512:45

BREAK15:3015:15

Answer Questions / Closing Remarks16:0015:30

Optional No Host Happy Hour @ El Rey (4201 Wilson Blvd, Arlington, VA 22203)16:00

Approved for Public Release, Distribution Unlimited 3

Context | Department of War AI Acceleration Strategy

OUSW(R&E) and CDAO are partnering for a number of pace-setting projects intended to be outcome-oriented vehicles for rapid development of enabling foundational AI capabilities.

DISCORD (referred to as Ender’s Foundry by OUSW(R&E) right) is intended to provide capabilities to stay ahead of AI-enabled adversaries in warfighting.

Figure Source: https://media.defense.gov/2026/Jan/12/2003855671/-1/-1/0/ARTIFICIAL-INTELLIGENCE-STRATEGY-FOR-THE-DEPARTMENT-OF-WAR.PDF AI: Artificial Intelligence OUSW(R&E): Office of the Under Secretary of War for Research and Engineering CDAO: Chief Digital and Artificial Intelligence Office

Approved for Public Release, Distribution Unlimited 4

OUSW(R&E) Ender’s Foundry Goals

• Ensure US stays ahead of AI-enabled adversaries

• Build on success of SCEPTER: accelerate M&S environments to 100,000x real-time to support real-time execution

• Generate a tactics “factory” capable of executing 10,000+ tactical experiments daily

• Develop a counter-UxS proving ground

• Exploit vulnerabilities in adversarial COAs and architecture at machine speed

DARPA DISCORD Goals

• Collapse deliberate planning and execution to a new AI-native approach suited to the tempo of adaptation fights

• Identify non-linear "team-breaking" vulnerabilities in adversarial mission architectures

• Furnish commanders with a portfolio of strategies that offer options to a wide range of possible futures

Translating OUSW(R&E) Intent to DARPA Impact

Phase 1: Simulated Proving Ground Phase 2: LVC Proving Ground

OUSW(R&E): Office of the Under Secretary of War for Research and Engineering

AI: Artificial Intelligence

SCEPTER: Strategic Chaos Engine for Planning, Tactics, Experimentation and Resiliency

M&S: Modeling and Simulation

UxS: Unmanned Systems COAs: Courses of Action LVC: Live, Virtual. Constructive

Approved for Public Release, Distribution Unlimited 5

Current Planning Paradigm is Outpaced and Outmatched Russia-Ukraine War: Visualizing the Battlefield, RAND Arroyo Center

Doctrine fails to adapt with the battlefield. Sophisticated air defenses are consistently overwhelmed by saturation attacks from low-cost drones. A single interceptor can cost 10-100x more than its target. Static, kinetic, attrition-based planning is unsustainable.

Tactical Brittleness

In Ukraine, both sides have shown that success hinges on identifying and exploiting weaknesses in an adversary's communications, sensing, or C2. However, these discoveries are ad-hoc and opportunistic.

Shallow Insight

Adversaries in Ukraine adapt methods of attack in hours. The deliberate process delivers a single, "optimized" COA to commanders with no time to formally explore or validate a portfolio of options in response to real time data on a dynamic enemy.

The Illusion of Choice

The Russo-Ukrainian War provides daily evidence that the character of conflict has changed. Adversaries are demonstrating an ability to adapt tactics and technology in under two weeks. The deliberate planning process fails to match this tempo. This creates three critical limitations.

C2: Command and control COA: Course of Action

Approved for Public Release, Distribution Unlimited 6

DISCORD program goals

TECHNICAL AREAS

TA 1: Asymmetric Advantage Engine

CHALLENGES APPROACH

Current surrogate model acceleration is insufficient for real-time replanning. Current models are slow to retrain and assume multiple training examples to adapt.

Explore new methods for accelerated, compositional function lifting. Incorporate non-linear time series analysis methods to learn from fewer examples and single-shot.

TA 2: Human-Machine Teaming

AI-Native processes move at machine speed but are bottlenecked by the human components of decision and execution due to the limits of explainable AI.

Alternatives to generating a universally explainable AI or interface. Learn profiles of individuals through limited engagements enabling adaptation to specific humans.

TA 3: Live Adaptation Experimentation Simulation fails to address warfighter skepticism and sim-to-real gap, including integration uncertainties.

Integrate DISCORD into remote procedure and object frameworks and conduct LVC testing with warfighters.

C2: Command and Control AI: Artificial Intelligence TA: Technical Area LVC: Live, Virtual, and Constructive

SCEPTER: Strategic Chaos Engine for Planning, Tactics, Experimentation and Resiliency

APIs: Application Programming Interfaces

DISCORD aims to create an AI-Native process collapsing planning and execution, to out-pace and out-think adaptive adversaries in real-time.

Approved for Public Release, Distribution Unlimited 7

Takens’ Theorem and the Observational Inefficiency Problem

Delforge, 2020

Takens’ Delay Embedding Theorem (1981)

Let a 𝑑-dimensional dynamical system evolve on a smooth attractor, and let 𝑦 𝑡 be a generic smooth scalar observation of the system.

Then the delay embedding map

𝐘 𝑡 ൌ 𝑦 𝑡 𝑦 𝑡 െ 𝜏 …𝑦 𝑡 െ 𝑘 െ 1 𝜏 is an embedding (i.e., a one-to-one, smooth reconstruction of the attractor) for 𝑘 2𝑑 1.

Thus, the geometry of the original attractor (e.g., the Lorenz attractor) can be reconstructed from time-delayed observations of a single variable.

Takens’ theorem shows that a single observed time series implicitly contains the full state information via time delays.

Unlike Reinforcement Learning or Inverse Reinforcement Learning, you don’t need many trajectories or rich observations to reconstruct the underlying dynamics. Sugihara et al., 2020

Approved for Public Release, Distribution Unlimited 8

Takens Embedding Theorem and Causality

Sugihara et al., 2020

Using Takens' embedding theorem, we reconstruct separate delay-embedded manifolds from each observed time series (e.g., 𝑀from 𝑋 𝑡 ,𝑀from 𝑌 𝑡 .(If the variables are part of the same dynamical system, these manifolds are different projections of the same underlying attractor. As a result, their geometries are linked: neighborhoods on one manifold correspond to neighborhoods on the other.

Think of each delay-embedded manifold as a geometric “fingerprint” of a variable’s history. To test causality, we pick a point on 𝑀, find nearby points (similar pasts of 𝑌), and look at the corresponding times in 𝑋. If those 𝑋 values are also similar, it means 𝑌’s history reliably tells us what 𝑋 was doing.

When this reconstruction gets better with more data (and works in one direction more than the other) it indicates a causal influence.

Approved for Public Release, Distribution Unlimited 9

TA 1 – Asymmetric Advantage Engine

Challenge: The SCEPTER program succeeded in accelerating physics-based modeling. While suitable for deliberate planning, it is insufficient for real-time replanning in adaptive environments.

Insight: DISCORD will learn from the geometry of joint system behavior. Using nonlinear time series techniques, based on Takens' Embedding Theorem (such as RQA), TA 1 will reconstruct an adversary's dynamic state from short windows of simulated and live sensor data, revealing hidden features while reducing the training time and data requirements by up to three orders of magnitude.

DISCORD solutions will reason in advance to generate a portfolio of strategies as a starting point and will then adapt estimated models with real data to refine COAs in the portfolio accounting for adversarial surprise.

Approach and Metrics: TA 1 solutions will be given commander’s intent from TA 2 to optimize against, using surrogate models as approximate policy oracles.

Successful TA 1 solutions will not only be faster than real time but will use that speed up to efficiently search for tactics which use recovered architectural insight of vulnerabilities to generate COAs which maximize commander’s intent providing advantage over baseline COAs. Successful

Trauth et al. 2019 RQA is an advanced nonlinear data analysis technique that is often described as “data-efficient”.

RQA reveals new features like determinism, recurrence rate, and laminarity between interacting components. RQA has shown more than a 100x improvement when used to enhance inverse reinforcement learning.

TA: Technical Area SCEPTER: Strategic Chaos Engine for Planning, Tactics, Experimentation and Resiliency RQA: Recurrence Quantification Analysis COAs: Courses of Action

Phase 2Phase 1Metrics

>100>10Advantage Score – JP 5-0 weighted numerical comparison

>98%>75%Architectural Insight – Percentage of pre-identified vulnerabilities detected

Approved for Public Release, Distribution Unlimited 10

TA 2 – Human-Machine Teaming

Challenge: AI-Native processes move at machine speed enabling rapid discovery of dominating tactics in controlled experiments. Command post exercises with the SCEPTER tool have revealed the growing human bottleneck. Operators must iterate to translate commander’s intent into the tool to produce COAs and then work to understand the outputs.

Insight: DISCORD will create an AI that understands its role within a human-machine team. Instead of just presenting data, the system will actively modeling the team's shared goals, skills, and dependencies; including each individual’s expertise and knowledge. Interdependence analysis and transactive memory systems have shown the ability to improve machine common ground understanding, and improve teaming outcomes, including tempo.

The result is a true cognitive partner that makes the human commander more powerful, not a complex tool that overwhelms them; and learns individual warfighters in no more than 25-50 interactions as achieved during initial deliberate exploration during evaluation.

Approach and Metrics: TA 2 solutions will work with minimal data from warfighters using the system and build profiles rapidly.

Successful TA 2 solutions will have a measurable impact on the speed of command, allowing warfighters to come to decisions with DISCORD in 60s or less. They will do so by achieving >95% intent translation accuracy of natural language into weighted JP 5-0 criteria for TA 1.

TA: Technical Area AI: Artificial Intelligence SCEPTER: Strategic Chaos Engine for Planning, Tactics, Experimentation and Resiliency COAs: Courses of Action

JP 5-0: Joint Publication 5-0, Joint Planning, dated 16 June 2017 COGs: Centers of Gravity CSS: Combat Service Support

Phase 2Phase 1Metrics

< 60s< 5mSpeed of Command – Time required for a commander to orient to a tactical problem, comprehend options, and make a confident decision

>95%>80%Intent Translation Accuracy – Ability to correctly interpret commanders intent, automating weighting of JP 5-0 criteria

An example of the COA comparison process from JP 5-0.

Commanders must first establish criteria and weightings for those criteria to numerically compare COAs. Each commander will define different criteria and weighting, meaning there is never one best COA from an objective basis.

Approved for Public Release, Distribution Unlimited 11

TA 3 – Live Adaptation Experimentation

Challenge: DISCORD’s envisioned AI-native collapse of planning and C2 cannot be proven in simulation alone. Simulated exercises fail to address warfighter skepticism and the sim-to-real gap, including integration uncertainties.

Insight: DISCORD will integrate TA 1 and TA 2 solutions into real world capabilities in a LVC proving ground.

Live sensing and drones will be combined with simulated electronic warfare and kinetic effectors such as close in weapons systems. TA 3 focuses on moving DISCORD from a theoretical engine to one ready for validation by warfighters.

TA 3 focuses on the engineering requirements for integration of the software and hardware components with DISCORD’s partners as part of regularly scheduled demonstrations and exercises.

Approach and Metrics: TA 3 solutions must manage the integrated system’s real time planning requirements efficiently in simulated and LVC experiments.

Successful TA 3 solutions will ensure that time to effective response of the integrated prototype adds no more than additional 60s to TA 2 metrics; while iterating with TA 1 components over live data to ensure plan resilience under fire of 80% of the COAs for a commander’s intended advantage.

DISCORD will be tested in multi-domain UxS proving grounds with live, virtual, and constructive components. Existing testing facilities already implement live sensing capabilities that are fused and made available to Maven Smart Systems as standardized object models and feature large-scale UxS capabilities including aerial, surface, sub-surface, and ground capabilities that respond to standard remote procedure call systems like gRPC.

TA: Technical Area AI: Artificial Intelligence C2: Command and Control LVC: Live, Virtual, and Constructive

COAs: Courses of Action UxS: Unmanned System gRPC: Google Remote Procedure Call

Phase 2Phase 1Metrics

< 2m< 10mTime to Effective Response – End to end battlefield agility. Total time to sense plan-invalidating surprise through execution of new counter-plan

> 80%> 50%Plan Resilience Under Fire – Percentage of the original portfolio still viable after a sustained period of live engagement

Approved for Public Release, Distribution Unlimited 12

DISCORD Metrics

The DISCORD metrics are designed to provide quantitative, warfighter-oriented evidence that the program is systematically retiring the primary risks of the program.

Scheduled down selects will occur for failure to meet conditions of integration (Months 3, 6, and 18) or for failure to meet the specified metrics per TA (Months 12, 24).

Regular command post exercises will be executed as available with INDOPACOM and USSOCOM by the government team with performer technologies.

TA: Technical Area LVC: Live, Virtual, and Constructive INDOPACOM: US Indo-Pacific Command USSOCOM: US Special Operations Command

JP 5-0: Joint Publication 5-0, Joint Planning, dated 16 June 2017

Phase 2Phase 1Metrics

>100>10Advantage Score – JP 5-0 weighted numerical comparisonTA 1

>98%>75%Architectural Insight – Percentage of pre-identified vulnerabilities detected

< 60s< 5mSpeed of Command – Time required for a commander to orient to a tactical problem, comprehend options, and make a confident decision

TA 2

>95%>80%Intent Translation Accuracy – Ability to correctly interpret commanders intent, automating weighting of JP 5-0 criteria

< 2m< 10mTime to Effective Response – End to end battlefield agility. Total time to sense plan-invalidating surprise through execution of new counter-plan

TA 3

> 80%> 50%Plan Resilience Under Fire – Percentage of the original portfolio still viable after a sustained period of live engagement

Approved for Public Release, Distribution Unlimited 13

DISCORD Government Team

The DISCORD government team is organized to steer, ground, and validate the program from inception to transition. The team's structure combines operational leadership with deep technical expertise to guide performers and independently verify results.

The government advisory and experimentation team will define the critical c-UxS problem sets, platforms, capabilities, and adversaries for DISCORD’s tests.

IDA will translate these operational requirements into specific, intellectually challenging scenarios engineered to probe the limits of performers' systems.

MITRE will implement these scenarios in their SIMEX environment already used in UxS exercises by key combatant commands.

The DISCORD government team will build and implement these scenarios within their advanced LVC testbeds.

c-UXS: Counter Unmanned Systems IDA: Institute for Defense Analysis SIMEX: Simulation Experiment UxS: Unmanned System LVC: Live, Virtual, and Constructive

Approved for Public Release, Distribution Unlimited 14

DISCORD Program Schedule

(Month 3) Integration Test: At program kickoff, the Government Team will provide all performers with an integrated simulation environment, a baseline c-UxS scenario, and initial APIs for connecting TA 1 and TA 2 systems. By Month 3, all TA 1 and TA 2 performers must demonstrate a "steel thread" proof of concept, showing an end-to-end workflow that successfully exercises these APIs.

(Month 6) Midterm Evaluation: TA 1 and TA 2 performers will execute a dry run of the full evaluation process, demonstrating a Minimum Viable Product of their integrated capabilities within a live simulation that includes synthetic data injections of novel confounds. TA 3 performers will demonstrate a candidate system capable of receiving and acting upon TA 2 command structures in a live (hardware) demonstration.

(Month 12) Phase 1 Formal Evaluation: TA 1 and TA 2 performers will conduct a formal evaluation of their integrated systems, demonstrating quantitative progress against the Phase 1 program metrics. TA 3 performers will execute a live flight demonstration, showing their ability to adapt in real-time to simulated commands and achieve the required Phase 1 metrics. This event serves as the basis for down-selection to Phase 2.

(Month 13) Phase 2 Kick-Off: Successful performers will participate in a joint kick-off and workshop at a government test facility, interacting with joint UAS, USV, and UUV platforms to build a concrete technical roadmap for the final phase of the program.

(Month 18) End-to-End Integration Demonstration: Performers will demonstrate a fully integrated TA 1, TA 2, and TA 3 system operating in a joint UxS scenario at a government test facility, proving the end-to-end functionality of their solution.

(Month 24) Final Graduation Exercise: The program will culminate in a final, scored "graduation" exercise in a full LVC environment orchestrated by an organization on the government team. This final exam will include novel adversary tactics developed by IDA, live human operators from combatant commands, and formal scoring of the final Phase 2 metrics by MITRE to determine ultimate program success.

www.darpa.mil

Approved for Public Release, Distribution Unlimited 15