Algorithmic Warfare and the Evolution of the Kill Web: AI-Driven Decision-Making, Autonomous Swarms, and Strategic Stability

Algorithmic Warfare and the Evolution of the Kill Web: AI-Driven Decision-Making, Autonomous Swarms, and Strategic Stability

The Architecture of the Kill Web and Joint All-Domain Command and Control

The character of modern warfare is undergoing a profound structural transformation, shifting away from linear, monolithic "kill chains" toward dynamic, disaggregated, and multidimensional "kill webs." A traditional kill chain operates as a sequential, tightly coupled process—find, fix, track, target, engage, and assess—often binding a specific sensor directly to a specific shooter. If a single node in this chain is neutralized by adversary action, the entire sequence collapses, rendering the system fragile in contested environments1. The kill web, conversely, leverages ubiquitous connectivity, cloud computing, and artificial intelligence (AI) to create a decentralized mesh network of sensors, command nodes, and effectors distributed across the air, land, sea, space, and cyber domains1. This architecture manipulates time and space by allowing for the rapid, dynamic pairing of the optimal weapon to the optimal target, maximizing force lethality, operational resilience, and the utilization of latent capacity across the joint force1. At the center of this transformation is the integration of machine learning (ML) and advanced algorithms into military decision-making cycles. The United States Department of Defense (DoD) Joint All-Domain Command and Control (JADC2) initiative embodies this strategic shift. JADC2 aims to connect sensors from all branches of the armed forces—Air Force, Army, Marine Corps, Navy, and Space Force—into a unified, AI-powered battle network3. Guided by the overarching operational paradigm of "Sense, Make Sense, and Act," JADC2 relies on algorithmic processing to ingest data at machine speed, rendering the battlefield transparent and accelerating the commander's decision cycle far beyond human cognitive limits5. This networked approach transcends individual services, manifesting in initiatives like the Air Force's Advanced Battle Management System (ABMS), the Army's Project Convergence, and the Navy's Project Overmatch3. To bridge these service-specific architectures, the DoD emphasizes Combined JADC2 (CJADC2), incorporating allied and partner nation capabilities into a modular, open-architecture framework6. Integrating these disparate systems relies on specialized software integration tools, such as the System-of-systems Technology Integration Tool Chain for Heterogeneous Electronic Systems (STITCHES) and the C2 Incident Management Emergency Response Application (C2IMERA), which facilitate machine-to-machine cueing and automate the translation of data across incompatible legacy networks1.

The automation of the kill web fundamentally alters the execution of warfare. However, delegating tactical decision-making to algorithms introduces unprecedented technical and strategic complexities. Automating the kill web involves solving NP-complete mathematical problems under extreme temporal constraints, deploying autonomous drone swarms via decentralized consensus protocols in communication-denied environments, and securing neural networks against physical and digital adversarial attacks10. Furthermore, the deployment of AI-driven kill webs raises profound ethical, legal, and strategic questions regarding the threshold of human judgment and the risk of algorithmic escalation spirals, often termed "flash wars"14.

Multi-Sensor Data Fusion and Automated Target Recognition

The foundational capability of the kill web is the ability to achieve persistent, high-fidelity battlespace awareness. The initial phases of erecting a kill web involve saturating an area with sensors to achieve persistent monitoring—a modern evolution of historical efforts like Operation Igloo White, which utilized acoustic and seismic sensors along the Ho Chi Minh trail2. Today, the "Sense" and "Make Sense" phases of the JADC2 framework depend entirely on multi-sensor data fusion (MSDF). MSDF is the process of combining inputs from disparate, geographically distributed sensing modalities to produce a verified, single intelligence picture that no individual source could generate independently2.

The Modalities and Architectures of Multi-Sensor Fusion

Modern Automated Target Recognition (ATR) systems leverage data-driven machine learning techniques and deep neural networks to process massive volumes of data from electro-optical (EO), mid-wave infrared (IR), synthetic aperture radar (SAR), acoustic, and radio frequency (RF) sensors2. Relying on a single sensor modality introduces critical vulnerabilities. For example, optical systems are heavily degraded by weather conditions, camouflage, and darkness, while radar systems, despite their all-weather capabilities, may struggle with precise spatial resolution or highly granular material identification20. Multi-sensor fusion mitigates these inherent weaknesses by exploiting complementary phenomenological strengths. Advanced fusion engines operate across three primary echelons within the kill web architecture. Data-level fusion combines raw signals or pixels from multiple sensors prior to feature extraction, requiring precise spatial and temporal coregistration20. Feature-level fusion extracts independent characteristics from separate sensors—such as structural edges from SAR and material composition from optical sensors—and concatenates them into a joint feature vector for classification20. Finally, decision-level fusion allows individual ATR algorithms to produce independent target predictions, which are subsequently reconciled using probabilistic methods, voting schemes, or Bayesian inference models21. The operational value of this integration is clearly demonstrated in counter-camouflage, concealment, and deception (CC\&D) applications. A prime example is the fusion of Hyperspectral Imaging (HSI) with Foliage-Penetration Synthetic Aperture Radar (FOPEN SAR)20. HSI operates across hundreds of contiguous, narrow spectral bands (typically ranging from 0.4 to 2.5 ![][image1]m), providing detailed material identification and background characterization based on chemical reflectance20. However, HSI lacks surface penetration capabilities. FOPEN SAR, utilizing lower-frequency UHF and VHF bands, can penetrate heavy foliage and detect the physical geometric structures of concealed threat objects20. By fusing these inputs at the feature level, the kill web can utilize HSI to identify synthetic camouflage netting while SAR confirms the presence of the metallic chassis of an adversary command vehicle hidden beneath it20. Similar fusion dynamics dictate maritime domain awareness. In highly contested maritime environments, fusion engines continuously cross-reference Automatic Identification Systems (AIS), SAR, EO imagery, and RF emissions17. A vessel engaged in illicit activity, spoofing, or sanctions evasion may disable its AIS transponder—a tactic known as "going dark." Multi-sensor fusion algorithms immediately detect the cessation of AIS data, cue SAR satellites to confirm the physical presence of the vessel regardless of cloud cover, utilize RF detection to locate non-cooperative emissions, and apply historical behavioral analytics to determine if the vessel's trajectory violates operational rules17. To handle dynamic tracking and reduce computational latency in these environments, platforms often utilize Information Filters (a variant of the Kalman Filter) to merge multi-sensor measurements21. In the context of Bearings-Only Tracking (BOT)—often required in submarine and stealth aircraft environments to maintain passive operations—fusion significantly enhances target observability, preventing filter divergence and improving state estimations21. Moreover, programs like the Multi-Spectral Sensor Surveillance System (M4S) demonstrate how airborne platforms integrate non-developmental organic sensors (e.g., PLAID ESM, AN/APS-137B SAR, and AN/AAQ-49 passive mid-wave FLIR) through closed-loop control and distributed microprocessor architectures to dramatically reduce prosecution timelines against time-critical targets18. To train these complex models, the Air Force Research Laboratory relies on synthetic data, scale-model measurements, and real sensor data to build high-fidelity performance models capable of operating across vast and unpredictable operating conditions22.

Vulnerabilities in SAR-ATR: The Threat of Adversarial Attacks

As the kill web increasingly relies on deep learning architectures for SAR-based Automated Target Recognition, a critical and systemic vulnerability emerges: the threat of adversarial examples13. Adversarial attacks involve the deliberate injection of mathematically calculated, visually imperceptible perturbations into input data, misleading the neural network into assigning an incorrect label with exceptionally high confidence25. While adversarial attacks on standard optical images (such as the Fast Gradient Sign Method or FGSM) are well-documented in civilian computer vision, attacking SAR systems requires highly specialized, physics-oriented techniques13. SAR images are not passive optical reflections captured by a lens; they are generated through active electromagnetic backscattering and complex phase-history imaging equations13. Consequently, simply applying optical domain attacks to SAR data often fails to produce physically realizable results. Standard optimization-based attacks, such as the Carlini & Wagner (C\&W) method, typically minimize mean-squared reconstruction errors to ensure stealth. However, in the SAR domain, this mathematical minimization inadvertently causes smooth target edges and blurry weak scattering centers, which appear unnatural to radar analysts and distinct from true SAR phenomenology26. To generate stealthier and more physically plausible attacks against the kill web, researchers have developed advanced frameworks utilizing U-Net Generative Adversarial Networks (GANs)26. In this architecture, a generator network learns the separable features of the target to synthesize adversarial perturbations in real-time, replacing slow iterative searching processes. Simultaneously, a discriminator network forces the output to approximate true SAR characteristics, ensuring the generated adversarial image retains sharp edges and explicit weak scattering centers26. Furthermore, physics-oriented adversarial attacks extend these digital vulnerabilities directly into the physical world. Rather than perturbing digital pixels after the data has been collected, adversaries can deploy physical adversarial scatterers on the battlefield13. By utilizing algorithms like SAR-BagNet to identify the highly salient regions of a target recognized by the kill web's classifiers, adversaries can calculate the precise electromagnetic, structural, and textural parameters required for physical scatterers24. When deployed around a physical target, these scatterers interact dynamically with the incoming radar pulse. This embeds adversarial perturbations directly into the raw echo signals before they are even processed by the satellite or aircraft, effectively blinding the kill web's ATR engines to the true nature of the target and causing recognition accuracy to plummet13. Additionally, the threat of transferable Targeted Adversarial Attacks (TAA) in a "black-box" scenario—where the adversary does not have access to the underlying structure or weights of the kill web's specific neural network—poses a significant challenge27. Using Contrastive Learning-based Targeted Adversarial Attacks (CL-TAA), attackers can generate perturbations on a surrogate model that successfully transfer to, and deceive, unseen target models27. Defending the kill web's fusion architecture against these sophisticated, physics-informed adversarial attacks remains a primary challenge for algorithmic stability.

Dynamic Weapon-Target Assignment (WTA) Optimization

Once hostile targets are conclusively identified, localized, and tracked by the multi-sensor fusion layer, the kill web must determine the optimal allocation of kinetic and non-kinetic effectors to neutralize the threat. This decision-making process is governed by the Weapon-Target Assignment (WTA) problem, a classical and highly complex challenge residing at the intersection of operations research and military strategy10.

The Mathematics and Complexity of WTA

The core objective of WTA is to assign a set of ![][image2] available weapons (interceptors, missiles, electronic warfare assets) to a set of ![][image3] incoming targets (aircraft, cruise missiles, ground vehicles) such that the expected survival value of the defended friendly assets is maximized, or conversely, the expected damage inflicted upon the enemy is maximized10. The problem is formally classified as a nonlinear integer programming problem and is notoriously NP-complete10. This mathematical classification dictates that as the number of weapons and targets in the battlespace increases linearly, the computational time required to find an exact, optimal solution grows exponentially10. The fundamental formulation of the WTA problem defines a binary decision variable ![][image4], where ![][image5] if weapon ![][image6] is assigned to target ![][image7], and ![][image8] otherwise11. The parameters governing the calculation include ![][image9] (the strategic or monetary value of the defended asset ![][image10]), ![][image11] (the calculated probability that target ![][image7] will destroy asset ![][image10]), and ![][image12] (the lethality probability that weapon ![][image6] will successfully destroy target ![][image7])11. In a purely static environment (Static WTA or SWTA), all inputs, spatial relationships, and probabilities are known in advance11. However, the kill web operates in a strictly dynamic and chaotic environment (Dynamic WTA or DWTA). It implements a continuous "shoot-look-shoot" (SLS) engagement policy, where the status of surviving targets, the availability of friendly weapons, and the shifting geometry of the battlespace are continuously updated based on real-time damage assessments generated by the sensor fusion layer11. Due to the NP-complete nature of the problem, exact mathematical solvers completely fail in real-time tactical scenarios, as they cannot compute solutions within the seconds required for air defense10. Consequently, the kill web must rely on heuristic and metaheuristic combinatorial optimization algorithms to find near-optimal solutions rapidly. Common algorithms applied to this problem include Genetic Algorithms (GA), Tabu Search (TS), Simulated Annealing (SA), and Variable Neighborhood Search (VNS), which utilize evolutionary search mechanisms and local neighborhood exploration to navigate the massive decision space11.

Commercial Algorithms and the Capability Marketplace

To scale the WTA problem across massive, multidimensional forces, the Defense Advanced Research Projects Agency (DARPA) initiated the Adapting Cross-Domain Kill-Webs (ACK) program1. Traditional command and control structures are heavily stovepiped; an Air Force commander facing a rapidly developing maritime threat may have limited visibility into the available capabilities of Navy destroyers or Space Force electronic warfare assets1. The ACK software framework resolves this structural insularity by operating the kill web as a decentralized "Capability Marketplace"1. Adapting bidding algorithms originally developed for commercial e-commerce, the system treats military domain commanders simultaneously as consumers (who need a specific target engaged) and suppliers (who possess the sensors, weapons, and support elements required to achieve the effect)1. Through a specialized, standardized "bid and offer" language, capability providers offer their assets to the network based strictly on the effects they can generate, rather than the specific technological mechanics of the asset1. This abstraction is critical for operational security, allowing domains to offer support without exposing sensitive operational details, classified sources, or specialized methods1. The architecture utilizes "Virtual Liaisons"—a service abstraction layer operating within the marketplace—to rapidly evaluate and compare diverse cross-domain options1. During a demonstration with the Advanced Battle Management System (ABMS), the ACK decision aid analyzed thousands of permutations in real-time to select the optimal sensor-to-shooter pair to counter incoming aerial threats8. Once the optimal combination was identified, the software immediately sent an automated command "play" to the C2IMERA system and ground-based integrated fire control systems, facilitating instantaneous machine-to-machine cueing over Link-16 to scramble interceptors1. By pairing the right sensor and weapon together regardless of service branch, the algorithmic marketplace increases lethality, creates resilience through rapid substitutions if a node is lost, and capitalizes on latent capacity1.

The Integration of LLMs into Tactical Guidance

While classical metaheuristic optimization algorithms excel at maximizing numerical probabilities, they exhibit severe limitations in highly dynamic, uncertain tactical environments28. Traditional algorithms lack contextual understanding, often leading to a phenomenon known as "assignment switching" or chattering, where minor fluctuations in spatial metrics cause the algorithm to continuously reassign weapons to different targets, wasting fuel, time, and tactical advantage28. Recent advancements have demonstrated the viability of embedding Large Language Models (LLMs) directly into the dynamic WTA and cooperative missile guidance decision loops28. Formulating the tactical decision process not as a pure math problem, but as a contextual reasoning task, the LLM ingests global mission state data—such as threat direction, the strategic priority of defended assets, closing velocities, and geometric timing—and generates feasible interceptor-target allocations without relying on brittle, predefined weighting parameters28. The system queries the LLM with scenario data, receives a raw text response, and parses it using Regular Expressions (RegEx) to extract the structured assignment vector28. This assignment vector is then fed directly into classical kinematic guidance laws, such as Proportional Navigation Guidance (PNG)28. The PNG acceleration command relies on aligning the interceptor's velocity vector with the Line-Of-Sight (LOS) vector, expressed mathematically as: ![][image13] Where ![][image14] is the navigation constant, ![][image15] is the relative velocity vector between the interceptor and target, ![][image16] is the LOS angular rate vector, and ![][image17] is the unit LOS vector28. By inserting an LLM reasoning engine into this loop, the kill web bridges the gap between rigid numerical optimization and adaptive, human-level tactical prioritization. This hybrid approach preserves the proven physics of classical guidance laws while leveraging generative AI to yield a more stable, context-aware, and strategically coherent assignment paradigm28.

Autonomous Drone Swarms as Edge Nodes

A highly networked, centralized kill web demands resilience; relying solely on exquisite, multi-million-dollar platforms (such as advanced fighter jets or destroyers) creates vulnerable single points of failure. To counter adversaries leveraging overwhelming mass—most notably the People's Republic of China's extensive anti-access/area denial (A2/AD) capabilities in the Indo-Pacific—the DoD launched the Replicator Initiative30. The Replicator Initiative aims to rapidly field thousands of all-domain, attritable autonomous (ADA2) systems within an aggressive 18-to-24-month timeframe32. Overseen by the Defense Innovation Unit (DIU), Replicator prioritizes platforms that are "small, smart, cheap, and many"33. Selected systems in the first tranches include the AeroVironment Switchblade 600 loitering munition, the Anduril Ghost-X, and the Performance Drone Works C-100 UAS34. By deploying these relatively low-cost, expendable platforms, the kill web pushes computational power to the tactical edge. These autonomous systems act simultaneously as sensors collecting ISR data and shooters engaging targets, creating a pervasive, highly resilient network capable of penetrating adversarial theater-level bubbles31.

Decentralized Task Allocation: The Consensus-Based Bundle Algorithm (CBBA)

Operating thousands of autonomous drones collaboratively in contested, communication-denied environments renders traditional, centralized command-and-control impossible. Drone swarms must operate autonomously as decentralized mesh networks, resolving complex task allocation organically and dynamically12. The standard algorithmic framework enabling this capability is the Consensus-Based Bundle Algorithm (CBBA)12. CBBA is a distributed, market-based protocol that provides provably good approximate solutions for multi-agent, multi-task allocation38. The algorithm consists of two alternating phases:

1. Bundle Building Phase (Auction): Each agent (drone) utilizes a greedy strategy to sequentially add tasks to its local bundle. It calculates bids based on a scoring function that adheres to the principle of Diminishing Marginal Gain (DMG)—the mathematical reality that the marginal utility of adding additional tasks to an agent's queue decreases monotonically as the bundle grows12.

2. Consensus Phase: Agents broadcast their winning bids, execution paths, and associated knowledge states to neighboring agents. Through local communication and a strict set of rule-based deconfliction protocols, the swarm resolves conflicting claims, guaranteeing a globally consistent and conflict-free task assignment across the network12.

The mathematical formulation for evaluating a task's utility highly prioritizes temporal constraints and spatial geometry. In evaluating the time sensitivity of each task, the reward incorporates a temporal decay factor. The score for an agent ![][image6] bidding on task ![][image7] is calculated as: ![][image18] Where ![][image19] represents the intrinsic strategic value of task ![][image7], ![][image20] is the temporal decay coefficient, ![][image21] is the earliest required start time of the task, and ![][image22] is the earliest feasible arrival time based on the drone's velocity, physical location, and current spatial bundle trajectory39.

Addressing Scalability and Battlefield Damage

While the baseline CBBA provides a robust theoretical framework, it faces severe scalability limits in massive, heterogeneous swarms due to heavy communication overhead39. To overcome this, architects developed the Two-Level Clustered CBBA (TLC-CBBA). In the first layer, the swarm utilizes graph-theoretic centrality measures to establish a three-tiered communication hierarchy, selecting highly connected drones as core backbone nodes39. In the second layer, a resource-balanced and distance-aware K-medoids algorithm groups the remaining drones into local sub-clusters39. The standard CBBA is executed locally within these small sub-clusters, while the "cluster heads" coordinate lightweight inter-cluster consensus. To prevent highly capable drones from being assigned all the tasks, TLC-CBBA incorporates a nonlinear load-balancing penalty that decays exponentially as an agent nears its maximum capacity, promoting an equitable distribution of workload39. Furthermore, in highly confrontational environments, individual drones will inevitably be destroyed or jammed. Enhanced dynamic variants, such as CBBA with Partial Replanning (CBBA-PR) and Damage-Reallocation models, utilize heartbeat-hold mechanisms for real-time condition monitoring12. When a drone is destroyed, recalculating the entire swarm's mission matrix from scratch is computationally prohibitive and tactically dangerous. Instead, CBBA-PR triggers targeted partial resets, releasing only the lowest-bid tasks from the damaged agent's bundle for rebidding by the surviving swarm12. This targeted reset balances rapid convergence with high coordination quality, ensuring mission continuity with minimal latency12.

Human Judgment and the Legal Framework of Autonomous Weapons

As the kill web accelerates the tempo of warfare, it inevitably strains traditional paradigms of human oversight. The delegation of lethal targeting decisions to software algorithms—particularly in the context of autonomous swarms executing decentralized engagement logic—has ignited intense international debate regarding International Humanitarian Law (IHL), the Law of Armed Conflict (LOAC), and the ethical limits of machine agency16.

Deconstructing DoD Directive 3000.09: Myths and Realities

In the United States, the development and employment of these systems are governed by DoD Directive 3000.09, Autonomy in Weapon Systems, originally published in 2012 and significantly updated in January 202345. The directive strictly defines a Lethal Autonomous Weapon System (LAWS) as one that, "once activated, can select and engage targets without further intervention by an operator," contrasting it with semi-autonomous systems that require an operator to select specific targets14. A pervasive and persistent myth surrounding U.S. military policy is that the DoD explicitly bans fully autonomous weapons or legally requires a "human-in-the-loop" for tactical engagements14. This is fundamentally false. The phrase "human-in-the-loop" does not appear anywhere in the text of Directive 3000.09; this omission was deliberate and calculated14. The U.S. government consistently rejects the civil society demand for "meaningful human control" (championed by advocacy groups), arguing that it falsely implies a requirement for continuous, real-time tactical supervision14. Such continuous oversight is operationally impossible in modern warfare, where communications are routinely jammed, and hypersonic weapon speeds dictate microsecond response times that far exceed human cognitive capacity14. Instead of demanding a human in the tactical loop, the directive requires that autonomous systems be designed to allow commanders and operators to exercise "appropriate levels of human judgment over the use of force"45. Accountability is situated at the operational and command levels, acknowledging that once a system is authorized and deployed, continuous tactical intervention may be impossible14. This distinction is vividly illustrated by systems currently in deployment and development. The Phalanx Close-In Weapon System (CIWS) has operated defensively aboard naval vessels since the 1980s. When incoming anti-ship missiles outpace human reaction times, operators switch the CIWS into automatic mode, allowing the radar to autonomously detect, track, and destroy threats14. The requisite "human judgment" occurs when the commander authorizes the system's activation based on the broader threat environment, not during the split-second tactical firing sequence14. Similarly, with Next-Generation AI-Guided Missiles or Collaborative Combat Aircraft (CCA), a human operator executes judgment by launching the weapon under defined Rules of Engagement (ROE)14. Once the missile is over the horizon and enters a communications-denied environment, it uses computer vision to autonomously identify and strike a target. The human has no ability to intervene or abort, yet the commander remains legally accountable for the deployment14.

Weapons Review, Article 36, and the Nuclear Exception

While Directive 3000.09 does not ban autonomous weapons, it mandates highly rigorous testing. Autonomous systems must undergo a standard legal review (Section 4), ensuring they can discriminate between combatants and non-combatants, minimizing civilian harm in accordance with the Law of Armed Conflict14. Furthermore, Article 36 of Additional Protocol I to the Geneva Conventions imposes a peacetime obligation on contracting parties to determine whether the employment of any new weapon or method of warfare would violate international law44. For novel autonomous architectures, Directive 3000.09 introduces two additional layers of mandatory senior-level review—involving the Undersecretary of Defense for Policy and the Chairman of the Joint Chiefs of Staff—before a system enters formal development, and again before it is fielded45. However, certain established defensive systems and non-lethal autonomous platforms are explicitly exempted from this stringent secondary review14. There is one critical exception to this paradigm. While the DoD directive avoids "human-in-the-loop" terminology for conventional weapons, the 2022 Nuclear Posture Review makes a distinct exception. It explicitly dictates that the United States will maintain a human "in the loop" for all critical actions regarding the presidential initiation or termination of nuclear weapons, isolating the highest echelons of strategic deterrence from the volatility of AI automation14.

Strategic Stability, Escalation Dynamics, and the "Flash War" Phenomenon

The most profound systemic risk associated with an AI-driven kill web is its potential to radically undermine crisis stability and inadvertently spark catastrophic conflict49. The primary danger lies not in science-fiction scenarios of "killer robots" achieving sentience, but in the structural compression of decision timelines, the amplification of automation bias, and the unpredictable interactive complexity of adversarial algorithms16.

The Risk of Algorithmic Flash War

In financial markets, high-frequency trading (HFT) algorithms routinely execute trades in fractions of a millisecond. In 2010, the unexpected, recursive interaction of competing trading algorithms caused the "Flash Crash," temporarily erasing approximately one trillion dollars of market value in minutes15. Defense strategists and wargaming experts at RAND and CSIS warn that the widespread militarization of AI could easily replicate this catastrophic dynamic on the battlefield, resulting in an uncontrollable "Flash War"15. If both the United States and a peer competitor—such as China, which is heavily pursuing a doctrine of "intelligentized warfare" emphasizing unmanned swarms—deploy highly automated kill webs to a contested region, the interaction of these systems may become entirely decoupled from human intent51. An AI-driven sensor fusion system might misclassify an adversary's defensive posture as an imminent offensive launch due to algorithmic hallucination or an adversarial physical scatterer16. The system's dynamic WTA algorithms, operating under diminishing time constraints to maximize the survival of friendly assets, could automatically deploy edge nodes (swarms) to aggressively jam or preemptively engage the perceived threat16. The adversary's automated systems, detecting this deployment at machine speed, would instantly initiate a counter-response. This recursive loop of auto-retaliation creates a kinetic fait accompli—a war triggered, expanded, and escalated far beyond the threshold of political intervention before human leadership is even notified16. Because AI-driven decision systems (AI-DDS) optimize for short-term tactical efficiency without the nuance of diplomatic context, their pursuit of immediate battlefield advantage strips away the friction necessary for de-escalation16.

Automation Bias and the "Fog of Certainty"

While the traditional "fog of war" refers to the confusion born of limited information, the integration of AI into JADC2 systems risks creating an equally dangerous "fog of certainty"16. AI models, particularly generative systems and complex classifiers, provide probabilistic recommendations that project an illusion of total mathematical certainty to the operator16. This phenomenon engenders automation bias, where commanders under immense cognitive overload inherently trust the machine's output over their own judgment16. Operating as a "tactical general," strategic leaders equipped with hyper-connected JADC2 networks might bypass hierarchical chains of command to execute tactical strikes based entirely on an algorithm's target prioritization16. However, as previously established, computer vision and SAR-ATR models are highly susceptible to adversarial spoofing16. A corrupted training dataset or a well-placed physical adversarial scatterer could cause the algorithm to "hallucinate" a phantom fleet, prompting a commander to authorize a multi-domain strike against a non-existent threat, thereby initiating an escalatory spiral based entirely on synthetic data16.

Mitigating Escalation: Circuit Breakers and Strategic Latency

Addressing the systemic risks of algorithmic instability requires profound structural shifts in strategic doctrine and system design. Scholars and defense analysts propose several safeguards to prevent the inadvertent escalation of the kill web:

1. Strategic Latency: A deliberate design principle emphasizing the intentional preservation of time, organizational friction, and human judgment within command structures16. While the tactical execution of a hypersonic missile intercept or a CIWS engagement necessitates full, out-of-the-loop automation, the authorization to deploy offensive swarms into ambiguous, politically sensitive, or nuclear-adjacent theaters must retain latency. This ensures diplomatic avenues for de-escalation remain open16.

2. Battlefield Circuit Breakers: Drawing directly from the regulatory solutions implemented by the SEC after the 2010 financial flash crash, military AI systems could be programmed with automated, hard-coded safeguards. If an escalation metric—such as a sudden geometric expansion of the battlefront, an anomalous spike in autonomous engagement rates, or the introduction of novel weapon types—exceeds a predefined threshold, the system temporarily restricts autonomous operations16. This forced shutdown acts as a fail-safe, demanding manual human intervention before hostilities expand16.

3. Explainable AI and Open Algorithms: To manage escalation in the shadow of nuclear weapons, states must balance countering adversary algorithms with the risk of unintentionally blinding a rival, which could trigger a "dead-hand" automated nuclear response50. Prioritizing explainable AI, verifiable human oversight models, and establishing dedicated bilateral hotlines for rapid de-escalation in the event of autonomous miscalculation are critical steps for maintaining twenty-first-century arms control50.

Conclusion

The transition from the linear kill chain to the AI-driven kill web represents a fundamental paradigm shift in the architecture of global military power. By leveraging advanced machine learning for multi-sensor fusion, military networks can achieve unparalleled situational awareness, systematically stripping away the adversary's ability to maneuver unseen across the electromagnetic spectrum. Concurrently, the application of complex combinatorial optimization and LLM-guided reasoning to the Weapon-Target Assignment problem ensures that effectors are deployed with maximal efficiency. At the tactical edge, consensus-based protocols like CBBA empower decentralized, attritable swarms to project overwhelming mass, realizing the strategic goals of the Replicator Initiative in contested environments. Yet, this massive leap in capability is counterbalanced by severe, potentially existential systemic fragilities. The vulnerability of ATR algorithms to physics-oriented adversarial attacks demonstrates that the very sensors driving the kill web can be weaponized against it. Furthermore, the structural demands of algorithmic warfare—prioritizing machine speed, efficiency, and automated engagement over deliberate human oversight—threaten to rapidly erode crisis stability. While DoD Directive 3000.09 establishes a robust legal framework focused on command accountability and appropriate human judgment, the physical reality of intersecting, adversarial AI systems engaging at machine speed remains deeply unpredictable. To maintain deterrence in the era of intelligentized warfare, military establishments must pursue a highly deliberate equilibrium. They must harness the computational superiority of the kill web to dominate the tactical environment, while rigidly enforcing fail-safes, battlefield circuit breakers, and strategic latency to ensure that the ultimate decision to cross the threshold of war remains squarely in the hands of human beings. Failure to balance these competing imperatives risks surrendering the pace and scope of geopolitical escalation to the unforgiving, recursive logic of the algorithm.

Works cited

1. ACK \- DARPA, https://www.darpa.mil/research/programs/adapting-cross-domain-kill-webs

2. How the kill web manipulates time and space \- Military Embedded Systems, https://militaryembedded.com/comms/communications/how-the-kill-web-manipulates-time-and-space

3. Joint All-Domain Command and Control \- Wikipedia, https://en.wikipedia.org/wiki/Joint\_All-Domain\_Command\_and\_Control

4. Pathways to Implementing Comprehensive and Collaborative JADC2 \- CSIS, https://www.csis.org/analysis/pathways-implementing-comprehensive-and-collaborative-jadc2

5. Summary of the Joint All-Domain Command and Control Strategy \- Department of War, https://media.defense.gov/2022/Mar/17/2002958406/-1/-1/1/SUMMARY-OF-THE-JOINT-ALL-DOMAIN-COMMAND-AND-CONTROL-STRATEGY.pdf

6. Combined Joint All-Domain Command & Control | Lockheed Martin, https://www.lockheedmartin.com/en-us/capabilities/multi-domain-operations.html

7. Army Fleshing Out Joint All-Domain Command, Control \- National Defense Magazine, https://www.nationaldefensemagazine.org/articles/2021/6/10/army-fleshing-out-joint-all-domain-command-control

8. Creating Cross-Domain Kill Webs in Real Time \- DARPA, https://www.darpa.mil/news/2020/cross-domain-kill-webs

9. Automatic Target Recognition on Synthetic Aperture Radar Imagery: A Survey, https://www.researchgate.net/publication/346967157\_Automatic\_Target\_Recognition\_on\_Synthetic\_Aperture\_Radar\_Imagery\_A\_Survey

10. Exact and Heuristic Methods for the Weapon Target Assignment Problem \- DSpace@MIT, https://dspace.mit.edu/entities/publication/8c3b5dbd-451a-4b36-b895-39fedeb34e02

11. Weapon Target Assignment with Combinatorial Optimization Techniques \- The Science and Information (SAI) Organization, https://thesai.org/Downloads/IJARAI/Volume2No7/Paper\_7-Weapon\_Target\_Assignment\_with\_Combinatorial\_Optimization\_Techniques.pdf

12. Consensus-Based Bundle Algorithm (CBBA) \- Emergent Mind, https://www.emergentmind.com/topics/consensus-based-bundle-algorithm-cbba

13. SAR-PATT: A Physical Adversarial Attack for SAR Image Automatic Target Recognition, https://www.mdpi.com/2072-4292/17/1/21

14. Autonomous Weapon Systems: No Human-in-the-Loop Required, and Other Myths Dispelled \- War on the Rocks, https://warontherocks.com/autonomous-weapon-systems-no-human-in-the-loop-required-and-other-myths-dispelled/

15. Flash Wars: Where could an autonomous weapons revolution lead us?, https://ecfr.eu/article/flash\_wars\_where\_could\_an\_autonomous\_weapons\_revolution\_lead\_us/

16. (PDF) AI-Enabled Military Decision-Making and Escalation Risk: Human-Machine Command Authority in Great Power Competition \- ResearchGate, https://www.researchgate.net/publication/401893273\_AI-Enabled\_Military\_Decision-Making\_and\_Escalation\_Risk\_Human-Machine\_Command\_Authority\_in\_Great\_Power\_Competition

17. What Is Multi-Sensor Fusion? \- Windward, https://windward.ai/glossary/what-is-multi-sensor-fusion/

18. Program Overview of the Multi-Spectral Sensor Surveillance System (M4S) \- DTIC, https://apps.dtic.mil/sti/tr/pdf/ADA399443.pdf

19. Double Weight-Based SAR and Infrared Sensor Fusion for Automatic Ground Target Recognition with Deep Learning \- MDPI, https://www.mdpi.com/2072-4292/10/1/72

20. Multisensor Fusion with Hyperspectral Imaging Data: Detection and Classification \- MIT Lincoln Laboratory, https://archive.ll.mit.edu/publications/journal/pdf/vol14\_no1/14\_1multisensorfusion.pdf

21. Multi-Sensor Data Fusion for Target Tracking Using Machine Learning Techniques, https://www.researchgate.net/publication/364294894\_Multi-Sensor\_Data\_Fusion\_for\_Target\_Tracking\_Using\_Machine\_Learning\_Techniques

22. AFRL Fusion-based Target Recognition Systems \- YouTube, https://www.youtube.com/watch?v=AJk8NcAAH0E

23. FUSION-BASED TARGET RECOGNITION SYSTEMS \- Air Force Research Laboratory, https://afresearchlab.com/wp-content/uploads/2023/02/AFRL\_FBTRS\_FS\_0223.pdf

24. Physics-oriented adversarial attacks on SAR image target recognition \- OpenReview, https://openreview.net/forum?id=DvmRl0K62A

25. Adversarial Attack and Defence through Adversarial Training and Feature Fusion for Diabetic Retinopathy Recognition \- PMC, https://pmc.ncbi.nlm.nih.gov/articles/PMC8201392/

26. Adversarial Attack for SAR Target Recognition Based on UNet-Generative Adversarial Network \- Semantic Scholar, https://pdfs.semanticscholar.org/5ab4/0ccc000fa15d806c9b55cd4a06398fa5c376.pdf

27. Transferable Targeted Adversarial Attack on Synthetic Aperture Radar (SAR) Image Recognition \- MDPI, https://www.mdpi.com/2072-4292/17/1/146

28. Generalized Intelligence for Tactical Decision-Making: Large Language Model–Driven Dynamic Weapon Target Assignment \- arXiv, https://arxiv.org/html/2511.10207v1

29. (PDF) Weapon-Target Assignment Strategy in Joint Combat Decision-Making Based on Multi-Head Deep Reinforcement Learning \- ResearchGate, https://www.researchgate.net/publication/374693028\_Weapon-Target\_Assignment\_Strategy\_in\_Joint\_Combat\_Decision-Making\_based\_on\_Multi-head\_Deep\_Reinforcement\_Learning

30. Autonomous Multi-Domain Adaptive Swarms-of-Swarms (AMASS): A Game-Changer in Countering Anti-Access/Area-Denial (A2/AD) Systems \- https://debuglies.com, https://debuglies.com/2023/09/27/autonomous-multi-domain-adaptive-swarms-of-swarms-amass-a-game-changer-in-countering-anti-access-area-denial-a2-ad-systems/

31. Pentagon Plans Dynamic Drone Swarms to Penetrate Enemy Defenses, https://www.missiledefenseadvocacy.org/air-defense-news/pentagon-plans-dynamic-drone-swarms-to-penetrate-enemy-defenses/

32. Scaling the Future: How Replicator Aims to Fast-Track U.S. Defense Capabilities, https://warontherocks.com/scaling-the-future-how-replicator-aims-to-fast-track-u-s-defense-capabilities/

33. Hicks unveils DOD's new 'Replicator' initiative to counter China via autonomous tech, https://defensescoop.com/2023/08/28/hicks-unveils-dods-new-replicator-initiative-to-counter-china-via-autonomous-tech/

34. Deputy Secretary of Defense Kathleen Hicks Announces Additional Replicator All-Domain Attritable Autonomous Capabilities \- Department of War, https://www.war.gov/News/Releases/Release/Article/3963289/deputy-secretary-of-defense-kathleen-hicks-announces-additional-replicator-all/

35. Replicator Initiative Continues Unmanned System Development | Federal Budget IQ, https://federalbudgetiq.com/insights/replicator-initiative-continues-unmanned-system-development/

36. Deep Dive: Pentagon's Replicator Initiative Raises Questions \- Inkstick Media, https://inkstickmedia.com/deep-dive-pentagons-replicator-initiative-raises-questions/

37. Event Driven CBBA with Reduced Communication \- arXiv, https://arxiv.org/pdf/2509.06481

38. Consensus-Based Bundle Algorithm (CBBA) \- Aerospace Controls Laboratory \- MIT, https://acl.mit.edu/projects/consensus-based-bundle-algorithm

39. A Two-Level Clustered Consensus-Based Bundle Algorithm for Dynamic Heterogeneous Multi-UAV Multi-Task Allocation \- PMC, https://pmc.ncbi.nlm.nih.gov/articles/PMC12610533/

40. A Dynamic Task Allocation Algorithm for Heterogeneous UUV Swarms \- PMC \- NIH, https://pmc.ncbi.nlm.nih.gov/articles/PMC8951437/

41. Auction-based distributed task allocation algorithm for drone swarms Dron sürüleri için müzakere tabanlı dağıtık görev \- Semantic Scholar, https://pdfs.semanticscholar.org/d0de/bd522187960c6453124e5eb1269684dd7335.pdf

42. Heterogeneous Multi-UAV Mission Reallocation Based on Improved Consensus-Based Bundle Algorithm \- MDPI, https://www.mdpi.com/2504-446X/8/8/345

43. Artifical inteligence (AI) in Warfare: Risks & Legal Challenges | ICRC \- YouTube, https://www.youtube.com/watch?v=ef8\_qqhRlnw

44. Artificial Intelligence and the Article 36 Legal Review \- Scholarship Commons, https://scholarship.law.slu.edu/cgi/viewcontent.cgi?article=2465\&context=lj

45. ARTIFICIAL INTELLIGENCE DoD Directive 3000.09: Autonomy in Weapon Systems \- Carahsoft, https://static.carahsoft.com/concrete/files/2417/3887/5530/Guidance\_DoD\_Directive\_3000.09\_-\_Autonomy\_in\_Weapon\_Systems.pdf

46. Pentagon updates guidance for development, fielding and employment of autonomous weapon systems | DefenseScoop, https://defensescoop.com/2023/01/25/pentagon-updates-guidance-for-development-fielding-and-employment-of-autonomous-weapon-systems/

47. Department of Defense Directive 3000.09 \- Wikipedia, https://en.wikipedia.org/wiki/Department\_of\_Defense\_Directive\_3000.09

48. United States, Use of Autonomous Weapons \- How does law protect in war? \- ICRC, https://casebook.icrc.org/case-study/united-states-use-of-autonomous-weapons

49. Don't Sweat the AGI Race \- RAND Corporation, https://www.rand.org/content/dam/rand/pubs/perspectives/PEA4100/PEA4188-1/RAND\_PEA4188-1.pdf

50. Algorithmic Stability: How AI Could Shape the Future of Deterrence \- CSIS, https://www.csis.org/analysis/algorithmic-stability-how-ai-could-shape-future-deterrence

51. Preventing a flash war: Countering the risk of AI-driven escalation on the battlefield \- CERL, https://www.penncerl.org/the-rule-of-law-post/preventing-a-flash-war-countering-the-risk-of-ai-driven-escalation-on-the-battlefield/

52. Emerging Warfare Dynamics in the Algorithmic Age—Artificial Intelligence, Mass, and Deception in Systemic Conflict \- U.S. Naval War College Digital Commons, https://digital-commons.usnwc.edu/cgi/viewcontent.cgi?article=8507\&context=nwc-review

53. AI Arms Race: How Autonomous Systems Are Reshaping Deterrence and Escalation Dynamics | Atlas Institute for International Affairs, https://atlasinstitute.org/ai-arms-race-how-autonomous-systems-are-reshaping-deterrence-and-escalation-dynamics/

[image1]: <data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAAwAAAAbCAYAAABIpm7EAAAA0klEQVR4XmNgGAWDEcgDsS6aGDMQ86KJgQEPEB8A4qVAzIgk3gDEF4FYGEkMDJSA+DkQFyGJcQPxHgZMQ8DAD4h/ArENkhjMkHQkMTiYBMR3gVgcSQybIXBAkgaYh9cAMQuSeCsDpiFgoAnEb4G4CkkM2RAQuxuIxWCSIKv/M0BMhIFgIP7LADFEB4h7GZBCCuT+f0D8HojbgXgFENcB8VwgPgXEy4HYEKYY3WpJKA0CIBNBEQbjgwHM/cgRhhfgDDpcAGQy1rSCC4DcCUqRIwYAAEHeJrl0xYuMAAAAAElFTkSuQmCC>

[image2]: <data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAAwAAAAbCAYAAABIpm7EAAAA10lEQVR4Xu3RrQ9BURjH8bNhY/MS2EwQaDKaYmMjKJIiois2CknX/AMUTVA0RRU0xaYKApspvufec66zO0Ui+G2f3Z3nuffc8yLEP7+UMKpIqrEPOdQQ1y/pBDHFCCc0sUQbA1xQct4mFXSQxRVrRFQvgSN6amylhQzqeKBo9GT9jK5RczLBHjGj1sAdBaPm5KMPQthgAa+qyeccW/Hak5N3a00J+9SGCGCMqG7qDZu/lndwQx5l9I2eNctOGDOQNA5YYSZcy/IL+wLdkTcuD8Hjbvzz3TwBLFQieXz1O1wAAAAASUVORK5CYII=>

[image3]: <data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAABIAAAAaCAYAAAC6nQw6AAABF0lEQVR4Xu3SMUtCURjG8TcyKChDdAlpDYIghwgEF7eECBHXoCkRnIpwMKLP0Fa4uLcVEdEg+Akc2tWxiKAhiJb+rx4Pr/c63vE+8MN7n8M913vOEYkTJ7okUcKmu1/CHsqmW8AWqu5X72eyiltcYYRjPOAULXzgEHe4xAkGOJdADlDDLr7xinU3toEh3rHvOk0HXZn8CZ+6TCap4A9FM7aNTzRNpw93cY+E6X1u8IaM6Y7wi4LpppM3TOezhp6E36KTD5A1nU7whR3T+Uzfcma6eZ+wgmdHry9kdu38+sz7BDu5brvuom6OHou2BBb8Gn2kTafn6gd50+lDT3jBI3JmbJxlCcxMFpGS8MHTXjdED22cOJHmH2DYLhijlc5IAAAAAElFTkSuQmCC>

[image4]: <data:image/png;base64,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>

[image5]: <data:image/png;base64,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>

[image6]: <data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAAcAAAAaCAYAAAB7GkaWAAAAnElEQVR4XmNgGHjADcSFQKyGLgECRUD8H4jT0SVAQASIHYCYFU0cN2AGYmMgtoGy4QBkxAQgrgXi00DciyzpCsQ1QMwHxAeAeCUDku5MINYHYksg/gbEETAJZNAAxE+AWBFNHOyFq0A8BYgZ0eQYPID4FxC7ALE6A8QUOJjBAHGpMAMklECOhAM/Boh9G4C4gAGL0TxALIAuOBIAANr5E9moi3bFAAAAAElFTkSuQmCC>

[image7]: <data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAAoAAAAZCAYAAAAIcL+IAAAAtklEQVR4Xt3RMQ4BURDG8REkREIjEa1Sq9xQCIVaqVK4gJKWC4gLbKNFROcOSrXKCZTCf/Je8CYUEg1f8is232zmvV2R300afUS2sOngipktbHJoI2+Lj5NEDQ1x53yZAhYYYo0NssEESWCCun/WSxxRvk/4FDFGStxFdtgi8zxkU8EJI1vYTMUN6gtvo6t05ffW6q+7oGULzQA9cR97jgNKwYTPGUs0xa3thvUjMfZYoWq6/8wNVcsZk3eqOEgAAAAASUVORK5CYII=>

[image8]: <data:image/png;base64,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>

[image9]: <data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAABMAAAAaCAYAAABVX2cEAAABOElEQVR4Xu3SsSuFURjH8UcoukoyILdEGUySSSGDu0om/gApu5QMt6QoMvgHMNsMJEUm5W8wKCmLwWD2/fWc93WuF+87sby/+nTf95xzzz3nea5ZmTJ/l2aMYxqtYawJoxhIFhWJvryPNdyFZ0UbveEKlTCWmxo20WW+2XEYb8cRbtARxnKzjBFM4B2L0dwkDsNzP27xjKF0xQ+p4wmD0dgCVqL3WVxazkl1xXucoiWMqQF181Mn2cB29P5t+vCI9WhMJ9yzz83bcIY59GLHfPNkPk0PHswnFXV4F2PpCrOqeYNmsGTefV0502ldaRUvOMEF5htWeL1ezTuvU8qvtdMCXVmfX6MSHJg3RCfqbJwunrhe+jH9RYbNN87ULC+q6bV5U7pxji1MxYuKRjWNC60GZQpf5p/zAREmKWg0xWZ3AAAAAElFTkSuQmCC>

[image10]: <data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAAsAAAAbCAYAAACqenW9AAAA9UlEQVR4Xu3SoYsCQRTH8ScqKOqhGC+IeAjX/AMuarJpNPgPWLQYxSheunbRfigGu2C0mkwGsVmMd3B33+fOwDK7cggXDP7gAztvht03Mytyk4kgj6w74WaET/yg58yFpoEvvLgTYXnDDo9OPZAMVlgg4cwF8owj+masmy2jhqRdZNPCN6qIY4gx5hKyYdtvCQNUxFsUOJ0c1tjgXbyWNNpGFykzPsf2exLv7drCg3+BP1ct9l/GE7aYyYUjdC9jIt4edC91tE1d0ljiAzFT08U61vN9RdHUpYADOrZAmthjaup6Qefog34uagsm+sU/f9V7/j+/q+oqHbeR1QMAAAAASUVORK5CYII=>

[image11]: <data:image/png;base64,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>

[image12]: <data:image/png;base64,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>

[image13]: <data:image/png;base64,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>

[image14]: <data:image/png;base64,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>

[image15]: <data:image/png;base64,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>

[image16]: <data:image/png;base64,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>

[image17]: <data:image/png;base64,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>

[image18]: <data:image/png;base64,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>

[image19]: <data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAABEAAAAaCAYAAABRqrc5AAABNUlEQVR4Xu2SvyuFURjHv0IpShn8LjEoJauBTOIaGJQymkhZJX+AlFJSSmIx22wykMWkrEoZZLMqGfg89zkv572Xc7ur7rc+vb3f85zzvs/3PFJNNVWvehiCKWiGRhjIVVTQINzCNizDNdzBVlyUUh88wCbUBW8JPmEuvCfVAMfwDP2RvwGv8vZijcJCiVcssuIz+YEme9r7FbQEL9OqvN2cZuW/vRJ5PfAE+5GXVHbITOSNw7vyeXTCDuxBa+QXNSxvJ9tgBZfBy/KwltahK6wVgv8tu401uIcT+dW+KJ9Hm7zFXvkY/Dk7tqEduuFRv+cxCRcqD7tMlscHzJcuyK89OXw24h3y8N5gWvkvNsG5KgzfCBzCUcREtG6Z3CiRR0q7cCoP/0A/A1mVxuQHLcpb/u/6Ao1YMODTzd49AAAAAElFTkSuQmCC>

[image20]: <data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAABQAAAAaCAYAAAC3g3x9AAABN0lEQVR4XuWUPytGYRiHb6H8KyxKTBSLxWKwkpQ/A4tvYLBTprfkGxjtFiYlbLKQD2A1iEwymBSun/uh8zy9pfOcs71XXct9v+99nnPu3zlmLc9AsBY6cBvf8Qz74nY+O/iJ82kjlyl8xQNsS3pZdOM5PuBI3MpnE79wI23kMobPeGS+rMpoyLH5UA2vjDb8ZH7buv1m6Flv4XDaSFnCW5w1X4wWpD+njOKh/TNwDR9xxjwyio4ipCiVRsPecL1QWzQPucL+iy60gie4UKhHaMiH+atXDPMg3uEN9oeaLjJnPvQCe0L9jwl8wQZ2xq0fGuan1CAxaf67XdwPtQidSCdpTxsB1Ycs/lh04SmuFmqV0Ct5bTVlVCinl1bzJ67p8yuDlnGPy3iF03G7PL3msdrD8aTXynwDL+0sc3JebLcAAAAASUVORK5CYII=>

[image21]: <data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAABMAAAAaCAYAAABVX2cEAAABNUlEQVR4Xu3UrUtDYRTH8SMq+DKwiCIKooIiWMQk2CbDwTbQKqwaZtGgqFHEZDHrf6DJNDVMVqzCQBjYxGY0GfR7OM+Vx2eT3WuwuB982O55Xnbv2cMVaaedv08nZpFBfzCWKNO4xzGKqOAF296cWBlHHXvocLV1fKAQTYqTLpzhGRNeXe/oVeyxw/SihJFwQCfroguxjTX6qdcVpFzNz5jYDTRslhd7nA2vNownnHq1WIk2y3q1JbxjDYvYcnXtp86/FPvHGzIn9phRowdwizcs4ADLbmwFabENy+hz9a/or23iAee4wSpqYgtO0O3mzrjv+zhytabRRg+JHVyNLhr0rqP04EoSHpmfMooqJsOB30T7dy3Nj0zi7EqLfrWKNv4ROdxh/vtwsugbZAeHmArG/ls+AT4lKytlxRTBAAAAAElFTkSuQmCC>

[image22]: <data:image/png;base64,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>