90 Seconds: Human Command Versus Machine-Speed Warfare

Full report

On this page

The integration of artificial intelligence, advanced sensor fusion, and autonomous effectors has precipitated a structural revolution in military operations, fundamentally altering the temporal dynamics of the modern battlefield. Traditional combat engagements relied on sequential human cognition to process information, assess threats, and authorize lethal force—a cycle measured in minutes or hours. Contemporary automated defensive architectures compress this sequence into a matter of seconds or milliseconds1. To comprehend this paradigm shift without resorting to the fallacy of artificial intelligence as a magical entity that independently understands the nuances of war, one must deconstruct the automated kill chain into its constituent technical processes. Algorithms do not possess an inherent understanding of warfare, morality, or strategic intent5. Instead, modern defensive automation represents an orchestrated architecture of statistical probabilities, kinematic filtering, and preprogrammed boolean logic. The interactive educational simulation titled “90 Seconds: Human Command Versus Machine-Speed Warfare” is designed to elucidate this complex architecture. By analyzing a high-intensity maritime defense scenario across three distinct operational modes, the experience demonstrates how human agency is not necessarily erased by automation, but rather temporally displaced. Decision-making is shifted from the chaotic moment of engagement to the deliberate, calculated environment of pre-mission planning, shaping how humans design, bound, and authorize the parameters of lethal force long before a hostile actor ever enters the battlespace2.

System-Level Kill-Chain Diagram

The foundation of the interactive experience is a visual and conceptual deconstruction of the automated kill chain. This system architecture moves from environmental observation to lethal action through a highly integrated sequence of discrete technological components. \[ENVIRONMENTAL STIMULI\] │ ▼

1. FIND (Sensors collect radar, IR, acoustic, visual, and electronic data) │ ▼

2. FIX (Algorithms detect anomalous behavior and filter noise via EKF) │ ▼

3. TRACK (Data-fusion combines observations via JPDA or MHT) │ ▼

4. IDENTIFY (Classification models assign probabilities using Dempster-Shafer theory) │ ▼

5. PRIORITIZE (Systems rank threats based on kinematics and asset vulnerability) │ ▼

6. AUTHORIZE (Rule-based logic filters allowable responses against pre-granted ROE) │ ▼

7. ENGAGE (Automated effectors execute the approved kinetic or electronic response) │ ▼

8. ASSESS (Battle-damage assessment loops back into the system to confirm neutralization)

The illusion of a sentient machine making independent combat choices obscures the reality of how these sequential automated systems function. The process begins with the "Find" phase, wherein diverse sensors—such as active phased-array radar, passive infrared, acoustic, and electro-optical systems—continuously sweep the environment. These systems generate massive volumes of raw data regarding radar cross-sections, thermal signatures, and radio-frequency emissions7. As sensor data arrives, the "Fix" phase utilizes algorithms to isolate anomalous signals from background clutter, such as environmental noise, wave patterns, or civilian traffic. Advanced mathematical models, particularly the Extended Kalman Filter (EKF), are utilized to handle nonlinear dynamics, smoothing noisy sensor measurements into a coherent positional estimate by maintaining a state vector and error covariance matrix9. To maintain continuous custody of an object during the "Track" phase, systems employ complex correlation algorithms. The Joint Probabilistic Data Association (JPDA) filter evaluates multiple potential paths for a moving object, predicting future positions and correlating new sensor blips to established tracks using Mahalanobis distance gating, even when the object temporarily disappears from view11. Alternatively, Multiple Hypothesis Tracking (MHT) delays hard assignment decisions by carrying multiple track hypotheses forward in time until subsequent data resolves the ambiguity12. During the "Identify" phase, classification models assign probabilistic identities to the tracked object rather than absolute certainties. Utilizing frameworks like the Dempster-Shafer theory of evidence, the system fuses conflicting multi-modal data—for instance, when radar returns suggest a missile but infrared suggests a drone—to calculate a basic belief assignment and confidence threshold16. This explicitly accounts for uncertainty and incomplete information. The "Prioritize" phase subsequently utilizes rule-based logic algorithms to rank verified threats based on kinematic urgency. Time-to-impact, heading, altitude, and the vulnerability of the defended asset are calculated to determine which target poses the most immediate existential risk18. The critical nexus of human-machine interaction occurs in the "Authorize" phase. The system queries its preprogrammed constraints, often termed doctrine statements. If the target's parameters match the authorized criteria set previously by a human commander, the system generates an authorization to fire6. Upon authorization, the "Engage" phase cues automated effectors—such as kinetic interceptors, directed high-power microwaves, or close-in weapon systems—assigning a fire-control solution and launching without further human mechanical input22. Finally, the "Assess" phase employs battle-damage assessment algorithms to evaluate changes in the target's trajectory, thermal bloom, or radar signature, feeding this data back into the loop to determine if the threat was destroyed or if re-engagement is required3.

Scene-by-Scene Interactive Narrative

The core of the educational experience places the user inside a simulated Combat Information Center (CIC) of a fictional defensive installation guarding the Strait of Oakhaven. This installation is tasked with protecting a coastal civilian population center and an allied naval task force. A swarm of low-flying, radar-evading loitering munitions, followed immediately by high-speed anti-ship cruise missiles, breaches the defensive perimeter. The total engagement window, from first radar detection to physical impact on the allied assets, is exactly ninety seconds. The simulation allows the user to experience the exact same ninety-second event under three different command-and-control architectures. In the Human-Directed mode, operators review every sensor report and approve each individual action. At ninety seconds to impact, early warning radars detect forty-five distinct anomalies, flooding the operator's screen with unclassified radar tracks. At seventy-five seconds to impact, the operator attempts to manually interrogate each track using optical sensors. The simulation demonstrates the Hick-Hyman Law, which dictates that an operator's decision time increases logarithmically as the number of choices and complexity expands, rapidly inducing cognitive overload25. By fifty seconds to impact, the operator successfully identifies three cruise missiles hidden within the drone swarm, but radio communication delays with the battery commander consume another fifteen seconds to verify rules of engagement. At twenty seconds, the operator manually authorizes engagement on the first missile. At zero seconds, the interceptors destroy one missile, while the remaining missiles and drones pass through the defensive perimeter. The human was perfectly in the loop and in control, but physically and cognitively incapable of matching the operational tempo2. In the AI-Assisted mode, artificial intelligence correlates detections and recommends actions, but a human must authorize the final engagement. At ninety seconds, sensors detect the same forty-five anomalies, and the AI-assisted system instantly correlates the data. By eighty-five seconds, the system filters out bird flocks and sea clutter, presenting the operator with twelve highly probable threats, color-coded by urgency. At seventy seconds, the AI prompts the operator with a recommendation for immediate engagement of three inbound high-speed tracks. The operator, experiencing a reduced cognitive burden but still acting as the final authorization bottleneck, reviews the AI's confidence scores and pushes the authorization button at sixty seconds. The system executes the engagements, destroying the cruise missiles by forty seconds to impact. However, the operator struggles to manually approve engagements for the remaining nine loitering munitions before they cross the threshold, allowing several objects to pass through the defensive area before a decision is made. This mode illustrates the latency trap of human-on-the-loop systems, where machine-speed attacks are met with human-speed bureaucracy29. In the Preauthorized Automation mode, the system acts automatically when programmed criteria are satisfied. At ninety seconds, the anomalies are detected. By eighty-eight seconds, the automated system, operating under a pre-established doctrine, fuses the data, calculates the kinematics, and identifies three cruise missiles and nine drones1. One second later, the system checks the established parameters against its logic engine, confirming that target velocities and headings intersect the defended asset, thus meeting the criteria to engage. At eighty-six seconds to impact, the system automatically allocates kinetic interceptors to the missiles and directs electronic effectors at the drone swarm22. By seventy-five seconds, all primary and secondary targets are destroyed. At zero seconds, the human operator, who merely monitored the system without intervening, watches the battle-damage assessment confirm total threat neutralization. The simulation underscores that the actual human decision was not removed; it was made weeks prior when the commander defined the precise geometric and kinematic conditions under which the system was authorized to defend itself2.

Mode Comparison and System Metrics

To quantify the operational disparities between the three command architectures, the simulation provides three dynamic mode-comparison tables that update in real-time as the scenario unfolds. These tables highlight the inverse relationship between immediate human mechanical control and system efficacy in a time-compressed environment.

Mode Comparison: Temporal Dynamics and Workload

Mode Comparison: Data Processing and Uncertainty

Mode Comparison: Risk and Governance

User Interactions and Pre-Engagement Architecture

To ensure the educational experience deeply emphasizes the pre-engagement responsibilities of human commanders, users are required to configure the system before the ninety-second simulation begins. These branching decisions dictate the entire outcome of the Preauthorized Automation mode, demonstrating that human control shapes the system long before the tactical event occurs. The user must establish confidence thresholds, setting the algorithmic certainty required to release a weapon. If the user sets the threshold too high, demanding ninety-nine percent certainty, the system will hesitate during the simulation, allowing targets to pass through the perimeter due to minor radar anomalies. Conversely, setting the threshold too low risks fratricide or civilian casualties6. The user is also tasked with defining geographic boundaries by drawing kill boxes on the tactical map. The automated system will only engage targets that enter these specific coordinates. A poorly drawn boundary may fail to protect the naval task force, while an overly broad boundary may endanger civilian maritime traffic21. Furthermore, the user must select approved target categories, toggling which specific profiles the system is authorized to engage, such as unmanned aerial systems, fixed-wing aircraft, or supersonic missiles. Operating time windows must be established, restricting the automated mode to function solely during high-risk nocturnal periods or specific threat windows. Finally, the simulation incorporates an abort condition interaction known as the kill switch. During the scenario, the user holds a physical or digital abort mechanism. If the system locks onto a civilian airliner that has strayed into the airspace due to a degraded transponder, the user has precisely three seconds to recognize the algorithmic error and break the automated kill chain, demonstrating the immense pressure placed on human supervisory control20.

What the Machine Saw: A Forensic Replay

Following the conclusion of the simulation, the user enters a forensic replay feature titled "What the Machine Saw." This sequence dissects the engagement, stripping away the polished user interface to reveal the raw mathematical and algorithmic reality of the system from four distinct perspectives. The replay begins with the perspective of the individual sensors. The screen displays pure, chaotic environmental data. Radar returns bounce off ocean waves, causing multipath interference and ghost tracks. Thermal sensors are blinded momentarily by atmospheric moisture. The simulation illustrates that to a raw sensor, there is no conceptual understanding of a missile or a drone; there are only fluctuating points of energy, Doppler shifts, and radio-frequency noise11. The perspective then shifts to the fused AI model. The simulation visually represents the Extended Kalman Filter and Joint Probabilistic Data Association algorithms in action. Bounding boxes appear around the noise, and the system projects predictive cones of probability ahead of each blip using Mahalanobis distance parameters. When a track disappears behind a terrain feature, the visualizer shows Multiple Hypothesis Tracking maintaining a ghost track moving forward, waiting to correlate the data when the object re-emerges from the blind spot13. The third perspective places the viewer back into the shoes of the human operator. The operator does not see the raw mathematical variance or the complex probability densities; they see simplified, sanitized iconography. Red diamonds denote hostile entities, while green circles denote friendly assets. This stark contrast demonstrates how the human-computer interface inherently shapes and occasionally biases the human's trust and understanding of the battlespace, hiding the machine's underlying uncertainty behind authoritative symbology. The final perspective reveals the automated rule set. The graphical interface dissolves into cascading lines of boolean logic and execution code. The simulation highlights specific commands evaluating track confidence, kinematic profiles, and location matrices before executing a fire control solution. The replay pauses to explicitly state that this specific line of code executing the lethal strike was authorized by a legal officer and a combatant commander during a planning meeting weeks prior, reinforcing the central thesis of the experience2.

Real-World Context: Sourced Public Examples

To anchor the theoretical simulation in reality, the experience explores established public systems, carefully delineating verified capabilities from manufacturer claims and analytical interpretations, while strictly avoiding claims of full autonomy where reliable public evidence is absent. The United States Navy's Aegis Combat System serves as a primary example of naval defensive automation. Verified public facts establish that Aegis is an integrated naval weapons system utilizing the AN/SPY-1 radar and Standard Missiles to track and engage enemy targets7. Manufacturer and government claims assert that the system possesses an "Auto-Special" mode designed to defeat multi-regiment bomber raids and high-speed anti-ship missiles, allowing for hands-off engagement from initial detection to kill assessment1. However, analyst interpretation clearly defines Auto-Special not as independent artificial intelligence, but as a human-starts-the-loop architecture. Operators develop multiple doctrine statements dictating exactly how the system will react to specific parameters; the human physically sets a fire integrity switch to enable the mode, after which the system automatically executes the pre-authorized logic2. The fictional simulation elements in the educational experience abstract the Aegis doctrine statements to fit the Strait of Oakhaven scenario, ensuring current operational vulnerabilities are not disclosed. In the realm of automated counter-rocket defense, the Israeli Iron Dome system provides crucial context. Verified facts confirm Iron Dome is a mobile air defense system designed to intercept short-range rockets, utilizing the ELM-2084 radar and a Battle Management and Weapon Control center18. Government claims cite a success rate exceeding ninety percent, largely attributed to the system's trajectory prediction algorithms that enable selective defense—calculating impact points and ignoring rockets destined for unpopulated areas to conserve interceptors18. Analysts interpret Iron Dome's operations as necessitating near-full automation during high-volume saturation attacks due to the extraordinarily short flight times of incoming rockets, leaving human operators primarily in a supervisory, confirm-or-abort role38. The precise neural network architectures used to rapidly discriminate between credible threats and debris remain classified capabilities. For automated counter-drone systems, Anduril's Lattice platform highlights the integration of distributed sensor fusion. Verified facts indicate Lattice is a software platform that integrates disparate sensors and systems into a single network layer for public safety and defense applications40. Manufacturer claims describe Lattice as utilizing machine learning to autonomously detect, classify, and track targets, seamlessly connecting sensors to effectors like the Anvil kinetic interceptor to compress the kill chain from minutes to seconds3. Analyst interpretations suggest that systems like Lattice represent a shift toward software-defined warfare and Joint All-Domain Command and Control (JADC2), where decentralized mesh networking prevents localized electronic attacks from blinding the platform43. The simulation strictly separates these claims from the fictional narrative, noting that while the software automates the tracking pipeline, public doctrine still mandates human authorization for lethal engagements against complex targets3.

Failure and Uncertainty Scenarios

Automated systems are exceptionally brittle when operating outside their precise training parameters or pre-established rules. The simulation educates users on these vulnerabilities by referencing historical failures and technological uncertainties. The most profound historical example of automated system failure is the United States Patriot missile fratricide incidents during the 2003 invasion of Iraq. Patriot batteries, operating in highly automated modes due to the expectation of rapid tactical ballistic missile attacks, misidentified a British Tornado GR4 and a U.S. Navy F/A-18 fighter jet as hostile anti-radiation missiles. The automated system recommended engagement, and the human operators authorized the lethal strikes45. Investigations, including the Defense Science Board Task Force report, revealed a catastrophic intersection of human and machine failures. The system's automated logic was originally adapted from Cold War-era anti-ballistic missile systems designed to fight high-altitude nuclear wars without human intervention, rendering the algorithms ill-suited for the complex, cluttered tactical airspace of Iraq47. Furthermore, operators suffered from automation bias—an unwarranted, uncritical trust in the machine's output45. Operators lacked the time to question the sensors, and critical safety mechanisms, such as the Identification Friend or Foe (IFF) transponders, suffered concurrent technical failures46. This scenario concretely demonstrates that placing a human on the loop is entirely insufficient if the human lacks the time, situational awareness, or organizational culture required to effectively challenge the machine's algorithmic conclusion. In modern multi-domain environments, systems face profound uncertainty generated by electronic warfare and adversarial machine learning. Adversaries utilize sensor spoofing and adversarial patch attacks to feed poisoned data into optical sensors or radar arrays8. If an adversary successfully deploys an adversarial algorithm that subtly alters a drone swarm's radar signature to mimic a flock of birds, the underlying Kalman filters and Dempster-Shafer models will process the poisoned data as truth. Consequently, the subsequent automated prioritization and authorization sequence will execute flawlessly upon a fundamentally false premise, bypassing the defensive architecture entirely. Furthermore, the simulation highlights the limitations of machine learning in chaotic environments. Reinforcement learning relies on statistical probabilities derived from vast training datasets. In an unprecedented combat scenario that the artificial intelligence has never encountered, the system will gravitate toward the most statistically plausible action based on past data, potentially ignoring rare but critical variables, leading to unintended escalation or mission failure5.

Safeguards and Human Control

The delegation of lethal force to automated processes invokes profound legal, ethical, and humanitarian obligations. International Humanitarian Law (IHL) mandates that all attacks must strictly adhere to the principles of distinction, proportionality, and precaution52. Under Article 36 of the 1977 Additional Protocol I to the Geneva Conventions, states are obligated to conduct rigorous legal reviews of all new weapons, means, and methods of warfare to determine if their employment would violate international law in some or all circumstances54. However, automated systems and AI-enabled decision-support tools challenge traditional Article 36 reviews. When a system utilizes dynamic machine learning, its behavior can evolve unpredictably post-deployment, potentially rendering pre-deployment legal reviews obsolete and creating a severe accountability gap52. To mitigate these risks during targeting, militaries employ Collateral Damage Estimation (CDE) methodologies within the Joint Targeting Cycle. However, there is a distinct risk that AI decision-support systems reduce complex proportionality assessments into mere numerical computations, sidelining the ethical deliberation and legal reasoning essential for maintaining human control57. If an algorithmic database, such as the Modernized Integrated Database (MIDB), relies on outdated category codes that incorrectly catalog a civilian structure as a military administrative building, the automated kill chain will mathematically assume the principle of distinction has been met, bypassing the rigorous legal scrutiny required for a high-risk strike19. The International Committee of the Red Cross (ICRC) has forcefully advocated for the preservation of meaningful human control over all weapon systems. The ICRC position asserts that autonomous weapon systems designed or used to target humans directly without human intervention violate the dictates of public conscience and fundamentally undermine human dignity, reducing victims to mere data points53. The ICRC argues that compliance with IHL cannot be delegated to machine processes; it is the human commander who remains legally and morally accountable for the strike. Therefore, human operators must understand and bound the spatial and temporal scope of any automated weapon to ensure predictable and lawful outcomes53. Consequently, defense policies such as U.S. Department of Defense Directive 3000.09 mandate rigorous hardware and software verification and validation (V\&V), alongside realistic developmental and operational testing, to ensure commanders can exercise appropriate levels of human judgment over the use of force throughout the weapon's lifecycle62.

Glossary of Automated Warfare

Conclusion

The compression of the military kill chain from hours to seconds represents an irreversible evolution in the character of modern warfare. As demonstrated by the interactive simulation, attempting to force human operators to manually process the sheer volume of data generated by multi-domain saturation attacks guarantees cognitive collapse and defensive failure. Conversely, delegating total unrestricted authority to algorithmic systems introduces the peril of context-blind execution, automation bias, and catastrophic systemic failure, as evidenced by historical fratricide incidents. The resolution to this strategic paradox lies not in debating whether humans should be involved in warfare, but when they are involved. Defensive automation demands that human oversight transition from the chaotic, split-second tactical edge to the deliberate, strategic planning phase. By establishing precise geographic boundaries, strict confidence thresholds, and rigorous rules of engagement, commanders shape the battlefield environment before the adversary arrives. When observing an integrated naval defense system seamlessly track, identify, and destroy an incoming supersonic threat without a tactical operator ever touching a firing console, one must confront the reality of modern combat architecture. When a machine acts in two seconds, was the human removed from the decision—or was the human decision made weeks before the event?

Works cited

1. The Blind Spot in Robot-Enabled Warfare: Deeper Implications of the IED Challenge, https://researchcentre.army.gov.au/library/australian-army-journal-aaj/volume-7-number-2-winter/blind-spot-robot-enabled-warfare-deeper-implications-ied-challenge

2. Reality in Autonomous Systems: It starts the Loop | The Cove \- Australian Army, https://cove.army.gov.au/article/reality-autonomous-systems-it-starts-loop

3. Anduril and Palantir: AI-Enabled Transformation of US Defense \- Lodi 411, https://lodi411.com/lodi-eye/anduril-and-palantir-ai-enabled-transformation-of-us-defense

4. Operation Epic Fury: The First AI War and The Revolution in Military Affairs \- lab45, https://static.lab45.id/2026/03/28/467/266operation-epic-fury-the-first-ai-war-and-the-revolution-in-military-affairssh.pdf

5. Can Complementary Learning Methods Teach AI the Laws of War? \- CIP, https://internationalpolicy.org/publications/can-complementary-learning-methods-teach-ai-the-laws-of-war/

6. Trust in the Machine: AI, Autonomy, and Military Decision Making with Lethal Consequences \- DTIC, https://apps.dtic.mil/sti/trecms/pdf/AD1121116.pdf

7. Aegis Combat System \- Wikipedia, https://en.wikipedia.org/wiki/Aegis\_Combat\_System

8. Scale-Adaptive Adversarial Patch Attack for Remote Sensing Image Aircraft Detection, https://www.mdpi.com/2072-4292/13/20/4078

9. Kalman Filtering in Sensor Fusion, https://sensorfusionauthority.com/kalman-filter-sensor-fusion

10. KALMAN FILTERING \- Theory and Practice Using MATLABw, https://d1.amobbs.com/bbs\_upload782111/files\_40/ourdev\_647449X5ZB54.pdf

11. A Multiple-Detection Joint Probabilistic Data Association Filter \- ResearchGate, https://www.researchgate.net/publication/258792149\_A\_Multiple-Detection\_Joint\_Probabilistic\_Data\_Association\_Filter

12. Adaptive Multi Target Tracking Through Optimal Data Association Technique Selection, https://search.proquest.com/openview/6834a754400bf8859b8b416e63b9a02f/1?pq-origsite=gscholar\&cbl=18750\&diss=y

13. Multiple hypothesis tracking for multiple target tracking \- SciSpace, https://scispace.com/pdf/multiple-hypothesis-tracking-for-multiple-target-tracking-3fzibddaoe.pdf

14. DATA ASSOCIATION ALGORITHMS FOR MULTISENSOR-MULTITARGET TRACKING \- MacSphere, https://macsphere.mcmaster.ca/bitstreams/3b218a30-1a1b-4c33-b676-52d4fd36e8f0/download

15. Multi-Information-Assisted Joint Detection and Tracking of Ground Moving Target for Airborne Radar \- MDPI, https://www.mdpi.com/2072-4292/17/12/2093

16. Dempster-shafer Theory: A Comprehensive Guide for 2025 \- Shadecoder, https://www.shadecoder.com/topics/dempster-shafer-theory-a-comprehensive-guide-for-2025

17. Multisensor Data Fusion in IoT Environments in Dempster–Shafer Theory Setting: An Improved Evidence Distance-Based Approach \- PMC, https://pmc.ncbi.nlm.nih.gov/articles/PMC10255415/

18. Iron Dome \- Wikipedia, https://en.wikipedia.org/wiki/Iron\_Dome

19. The New Battlespace: How Geospatial AI, Outdated Intelligence, and the Illusion of Oversight Are Reshaping Military Targeting, https://projectgeospatial.org/geospatial-frontiers/new-battlespace-how-geoai-outdated-intelligence-and-the-illusion-of-oversight-reshaping-military-targeting

20. Lethal Autonomous Weapon Systems: Reconciling the Myth of Killer Robots and the Reality of the Modern Battlefield by Mackenzie \- DukeSpace, https://dukespace.lib.duke.edu/bitstreams/edc98653-21cd-4c3c-9782-e4e6c07f59d1/download

21. A Four-Phase Approach to Developing Ethically Permissible Autonomous Weapons, https://www.armyupress.army.mil/Journals/Military-Review/Online-Exclusive/2018-OLE/May/Four-Phase-Approach/

22. A Counter to Drone Swarms: High-Power Microwave Weapons \- DroneShield, https://www.droneshield.com/blog/a-counter-to-drone-swarms-high-power-microwave-weapons

23. Epirus Leonidas High-Power Microwave: Directed Energy for cUAS, cUAS Swarms, Counter Electronics, https://www.epirusinc.com/electronic-warfare

24. Chapter: 7 Integrating Autonomy in Network-Centric Operations \- Read "Autonomous Vehicles in Support of Naval Operations" at NAP.edu, https://www.nationalacademies.org/read/11379/chapter/9

25. Human performance modeling \- Wikipedia, https://en.wikipedia.org/wiki/Human\_performance\_modeling

26. Engineering Psychology | ROAM, https://roam.libraries.psu.edu/system/files/e-books/Psychology-Engineering\_Psychology.pdf

27. Exploring triple tasks: A deeper dive into multitasking dynamics \- Athene Forschung, https://athene-forschung.unibw.de/doc/149944/149944.pdf

28. Human-AI Teaming Under Fire: Lessons from Ukraine's Human-in-the-Loop Combat AI Systems \- Digital Commons @ USF \- University of South Florida, https://digitalcommons.usf.edu/cgi/viewcontent.cgi?article=2583\&context=jss

29. Humans in the Loop \- wp0 | Vanderbilt University, https://wp0.vanderbilt.edu/lawreview/wp-content/uploads/sites/278/2023/03/Humans-in-the-Loop.pdf

30. The Intent Architect: Reclaiming the Cyber Initiative Through Symbiotic Autonomy, https://www.afcea.org/signal-media/cyber-edge/intent-architect-reclaiming-cyber-initiative-through-symbiotic-autonomy

31. The Drift: NavyCon Edition \- Defense News, https://www.defensenews.com/naval/the-drift/2020/07/23/the-drift-navycon-edition/

32. Aegis | Center for International Maritime Security, https://cimsec.org/tag/aegis/

33. Set JPDA Filter for Multitarget Tracking \- Chalmers Publication Library, https://publications.lib.chalmers.se/records/fulltext/146503/local\_146503.pdf

34. A Concept of Operations for Armed Autonomous Systems, https://ndia.dtic.mil/wp-content/uploads/2006/disruptive\_tech/canning.pdf

35. Maritime Security Dialogue: The Aegis Approach with Rear Admiral Tom Druggan \- CSIS, https://www.csis.org/analysis/maritime-security-dialogue-aegis-approach-rear-admiral-tom-druggan

36. ISRAEL'S AIR DEFENSE SYSTEM: STRUCTURE, COMBAT, https://usajournals.org/index.php/6/article/download/2429/2516/4938

37. Iron Dome \- RAFAEL Advanced Defense Systems, https://www.rafael.co.il/system/iron-dome/

38. How does Israel's Iron Dome use AI to improve system accuracy? \- Quora, https://www.quora.com/How-does-Israels-Iron-Dome-use-AI-to-improve-system-accuracy

39. DEADLY ALGORITHMS Destructive Role of Artificial Intelligence in Gaza War | SETA, https://media.setav.org/en/file/2025/02/deadly-algorithms-destructive-role-of-artificial-intelligence-in-gaza-war.pdf

40. Anduril's Lattice: A Trusted Dual Use Platform for Public Safety and Defense, https://www.anduril.com/news/anduril-s-lattice-a-trusted-dual-use-commercial-and-military-platform-for-public-safety-security

41. Anduril Lattice Mesh: The Neural Network Behind Fury UCAV and C-UAS Systems (2026), https://www.aomway.com/post/1961.html

42. Report: Anduril Industries Business Breakdown & Founding Story | Contrary Research, https://research.contrary.com/report/anduril

43. Anduril's Lattice Platform: Architecture, Accountability, and the Future of Autonomous Warfare in American Defense Strategy \- Foreign Affairs Forum, https://www.faf.ae/home/2026/4/21/andurils-lattice-platform-architecture-accountability-and-the-future-of-autonomous-warfare-in-american-defense-strategy

44. Hypergraph Centrality and Low-Observable MUM-T Architectures: Quantifying Drone Swarm Command via the Two-Seat Su-57D Platform \- debugliesintel, https://www.debugliesintel.com/hypergraph-centrality-and-low-observable-mum-t-architectures-quantifying-drone-swarm-command-via-the-two-seat-su-57d-platform/

45. CENTAUR WARFIGHTING: THE FALSE CHOICE OF HUMANS VS. AUTOMATION Paul Scharre\* \- Temple Sites, https://sites.temple.edu/ticlj/files/2017/02/30.1.Scharre-TICLJ.pdf

46. Understanding the errors introduced by military AI applications \- Brookings Institution, https://www.brookings.edu/articles/understanding-the-errors-introduced-by-military-ai-applications/

47. PATRIOT WARS \- Amazon S3, https://s3.amazonaws.com/files.cnas.org/documents/CNAS-Report-EthicalAutonomy5-PatriotWars-FINAL.pdf

48. AI Safety and Automation Bias \- CSET, https://cset.georgetown.edu/wp-content/uploads/CSET-AI-Safety-and-Automation-Bias.pdf

49. Lethal Autonomous Weapons Systems: The Overlooked Importance of Administrative Accountability \- Scholarly Commons | George Washington University Law School Research, https://scholarship.law.gwu.edu/cgi/viewcontent.cgi?article=2621\&context=faculty\_publications

50. Adversarial Machine Learning \- Joint Air Power Competence Centre, https://www.japcc.org/essays/adversarial-machine-learning/

51. Military Inertial Sensor Market Research Report 2033 \- Dataintelo, https://dataintelo.com/report/military-inertial-sensor-market

52. How Autonomous Weapon Systems Threaten International Humanitarian Law \- Giga- Hamburg, https://www.giga-hamburg.de/publications/giga-focus/how-autonomous-weapon-systems-threaten-international-humanitarian-law

53. Autonomous Weapon Systems and International Humanitarian Law: Selected Issues \- ICRC, https://www.icrc.org/sites/default/files/2026-03/4896\_002\_Autonomous\_Weapons\_Systems\_-\_IHL-ICRC.pdf

54. Implementing Article 36 Weapon Reviews in the Light of Increasing Autonomy in Weapon Systems | SIPRI, https://www.sipri.org/publications/2015/sipri-insights-peace-and-security/implementing-article-36-weapon-reviews-light-increasing-autonomy-weapon-systems

55. IMPLEMENTING ARTICLE 36 WEAPON REVIEWS IN THE LIGHT OF INCREASING AUTONOMY IN WEAPON SYSTEMS\* \- SIPRI, https://www.sipri.org/sites/default/files/files/insight/SIPRIInsight1501.pdf

56. Article 36 reviews and addressing Lethal Autonomous Weapons Systems, https://article36.org/assets/documents/publication\_files/LAWS-and-A36.pdf

57. The erosion of human(e) judgement in targeting? Quantification logics, AI-enabled decision support systems and proportionality assessments in IHL | International Review of the Red Cross | Cambridge Core, https://www.cambridge.org/core/journals/international-review-of-the-red-cross/article/erosion-of-humane-judgement-in-targeting-quantification-logics-aienabled-decision-support-systems-and-proportionality-assessments-in-ihl/31024C3473211FB84FD535BC815ADC03

58. GC REAIM Expert Policy Note Series \- The Hague Centre for Strategic Studies, https://hcss.nl/wp-content/uploads/2025/05/Dorsey.pdf

59. Autonomous weapons | ICRC, https://www.icrc.org/en/law-and-policy/autonomous-weapons

60. Autonomous Weapons Systems: Ensuring Human Control Based on Respect for the Dignity of Life | Soka Gakkai (global), https://www.sokaglobal.org/contact-us/media-room/statements/ensuring-human-control-of-aws-based-on-dignity-of-life.html

61. Autonomous Weapon Systems and International Humanitarian Law: Selected Issues \- ICRC, https://www.icrc.org/en/article/autonomous-weapon-systems-and-international-humanitarian-law-selected-issues

62. DoD Directive 3000.09, "Autonomy in Weapon Systems," January 25, 2023 \- Executive Services Directorate, https://www.esd.whs.mil/portals/54/documents/dd/issuances/dodd/300009p.pdf

63. AI Governance for Defense & EU AI Act | Modulos, https://www.modulos.ai/industries/defense/