Global Kill Chain Evidence Atlas: Architecture and Design Specification

Full report

On this page

Executive Atlas Concept

The integration of artificial intelligence and algorithmic decision-making into military systems has fundamentally altered the tempo, scale, and operational architecture of modern combat1. Traditional, linear engagement cycles—historically conceptualized as Find, Fix, Track, Target, Engage, Assess (F2T2EA)—are rapidly evolving into highly distributed, multi-domain "kill webs." This paradigm shift, most prominently articulated in initiatives such as the United States Defense Advanced Research Projects Agency (DARPA) Mosaic Warfare concept, emphasizes the rapid, algorithmic composability of disaggregated sensors, command nodes, and automated effectors3. As these technologies proliferate, assessing the global landscape of military artificial intelligence requires moving beyond simplistic binary categorizations of "autonomous" versus "manual" weapons. The Global Kill Chain Evidence Atlas serves as a rigorous, interactive, and transparent visualization engine designed to map this sociotechnical evolution. Its primary mandate is to render the intersection of artificial intelligence, human-control architectures, and international military doctrine visually compelling and exhaustively credible, without exposing sensitive operational targeting coordinates or generating tactical hazards5. The platform operates not as a static repository of military capabilities, but as a dynamic evidentiary graph. It distinguishes verified operational deployments from theoretical doctrine and defense manufacturer marketing claims by strictly enforcing open-source intelligence (OSINT) verification protocols6. By synthesizing frameworks such as the NATO Admiralty Code for intelligence evaluation7 with international legal principles outlined in instruments like the Political Declaration on Responsible Military Use of Artificial Intelligence and Autonomy8, the Atlas provides policymakers, researchers, and the public with a mechanism to interrogate the exact nature of AI in warfare. It addresses the critical epistemological challenge of modern defense analysis: public descriptions of AI-assisted command and sensing systems frequently leave human-control arrangements unspecified10. The Atlas forces clarity onto this ambiguity, utilizing WebGL-powered geospatial and topological visualizations to trace every technological assertion back to its foundational source documentation.

Information Architecture

The structural integrity of the Atlas relies on a deep, interconnected ten-layer relational graph. This architecture ensures that individual hardware or software systems are never presented in an informational vacuum; rather, they are continuously contextualized within the operating nation's declared policies, international legal obligations, and the qualitative strength of the underlying evidence.

The user navigation flow is designed as a process of progressive disclosure, allowing researchers to descend from macro-level geopolitical trends down to forensic documentation. Users initiate their inquiry at the World overview, utilizing a global map to identify clusters of technological adoption or policy adherence. Selecting a specific Country transitions the interface to a national profile, detailing the state's strategic posture. From the national hub, users explore the Doctrine layer to understand the state's self-imposed constraints, which in turn provides context for the network of operational Systems. Clicking a specific platform reveals its Function within the kill chain and the specific Human-control mechanisms intended to govern it. Should a user question the validity of a highly advanced algorithmic capability, they seamlessly pass through the Evidence layer to evaluate the Admiralty Code rating before arriving at the unalterable Source layer containing the archived proof.

Country Profile Template

The country profile operates as an exhaustive, neutral consolidation of a nation's public posture regarding military artificial intelligence and automated engagement. A core editorial mandate of the Atlas is to prevent the inference of secret or hyper-advanced capabilities derived solely from standard defense procurement budgets, patent filings, or aspirational marketing language5. Therefore, the country template enforces strict evidentiary boundaries, explicitly separating verified field capabilities from theoretical research. The profile begins with the country's standardized name and a neutral overview summarizing its historical trajectory in military automation and digitalization. This is immediately followed by a detailed section on publicly stated doctrine. This section archives and interprets foundational texts—such as a nation's interpretation of "appropriate levels of human judgment" over the use of force10—and maps how these doctrines dictate operational limits. The template then details relevant military or defensive AI programs, such as participation in Joint All-Domain Command and Control architectures or Mosaic Warfare initiatives4, distinguishing broad research mandates from specific weaponized outputs. The core of the profile is a gallery of publicly documented systems, directly cross-referenced against a matrix of kill-chain functions where AI or automation is reportedly utilized. This reveals exactly where a state is applying algorithmic support, whether in the "Find" phase via satellite image computer vision, or the "Engage" phase via semi-autonomous loitering munitions. The human-control policy section aggregates the nation's official stance on operator supervision, including critical failsafe behaviors during communications degradation or electronic warfare jamming11. Furthermore, the profile documents the state's weapons-review and governance policy, particularly its adherence to Article 36 of Additional Protocol I to the Geneva Conventions, which mandates the legal review of new weapons to ensure they can be used in accordance with international humanitarian law18. This is supplemented by a record of alliances and multinational declarations, such as the state's endorsement of the REAIM Political Declaration on the Responsible Military Use of Artificial Intelligence and Autonomy8. Given that AI systems are highly proliferated through defense trade, the export and acquisition relationships section tracks the transfer of these technologies, acknowledging that exclusive control and operational oversight are often lost once a system is exported22. The profile concludes with a chronological timeline of the state's development, an evidence-quality summary providing a macro-level grade of the state's public transparency, and an explicit listing of important disputes or unknowns. The inclusion of an "Unknowns" section is vital; it actively highlights where the public record fails to explain critical command-and-control architectures. Finally, the profile provides direct links to the methodology, the primary sources, the date of the last editorial review, and a structured path for external experts to submit corrections.

System Record Template and Data-Field Specification

The system card is the atomic unit of technical analysis within the Atlas. Because public discourse frequently conflates narrow algorithmic automation with generalized lethal autonomy, the system template enforces standardized wording to maintain absolute precision23. The data-field specifications require rigorous adherence to established engineering and legal taxonomies15. The record begins with the system name, its country or operator, the manufacturer or developer, and its primary category (e.g., Unmanned Surface Vessel, AI Decision Support System, Air Defense Artillery). The deployment status utilizes strict tags—Development, Demonstrated in Testing, Fielded, or Outdated/Superseded—to clarify operational reality. The public sensors field documents the hardware inputs, such as electro-optical/infrared (EO/IR), acoustic arrays, or synthetic aperture radar (SAR), which are critical for understanding how the system perceives its environment13. The AI or automated function field requires editors to utilize highly specific terminology to distinguish capabilities. An "AI-enabled" system might utilize neural networks solely for predictive maintenance; an "Algorithmically automated" system may follow pre-programmed GPS waypoints; while "Automatic target recognition" (ATR) implies the software classifies objects as potential threats based on sensor data25. The position in the kill chain field precisely maps the system's function to the F2T2EA cycle. Crucially, the documented autonomy and control mode fields define the system's operational independence. The control mode must be classified as operator-supervised engagement (human monitors and can abort), human-authorized engagement (human must explicitly approve force), autonomous target selection (the system selects but may not engage), autonomous engagement (full lethal autonomy), or publicly unspecified engagement authority10. The human intervention or abort information field details the specific mechanical or software failsafes allowing operators to override the system, a critical factor in determining legal compliance and risk26. To counter the narrative of infallible technology, the public limitations field documents known vulnerabilities, such as high false-positive rates in ATR algorithms, susceptibility to environmental noise, or the propagation of data-fusion errors in complex battle-management software11. The system record explicitly segregates manufacturer claims from independent evidence. Assertions made in defense contractor marketing materials are isolated and clearly labeled, while OSINT, NGO reports, or UN investigations that corroborate or refute those claims are given primary weight6. The record concludes with the system's source quality rating, a list of known unknowns, relevant governing policies, and a timestamp of the last editorial review.

Evidence States and Credibility Design

The fundamental challenge of visualizing military AI lies in the uneven quality of the underlying data. Without a rigorous epistemological framework, the Atlas risks making a defense manufacturer's unverified marketing claim appear equivalent to an independently verified, field-tested operational fact. To prevent this, the Atlas employs a highly visible, multidimensional evidence-state system derived from a synthesis of the NATO Admiralty Code7 and modern OSINT verification standards pioneered by organizations such as Bellingcat31. The Admiralty Code evaluates intelligence across two independent axes: Source Reliability, graded from A (Completely reliable) to F (Reliability cannot be judged), and Information Credibility, graded from 1 (Confirmed by other sources) to 6 (Truth cannot be judged)7. The Atlas adapts this logic to generate thirteen distinct evidence states:

1. Officially documented (e.g., published military doctrine).

2. Court- or treaty-recorded (e.g., UN Security Council investigations).

3. Independently corroborated (e.g., multi-source OSINT with chain-of-custody).

4. Operator-described (e.g., first-hand accounts from military personnel).

5. Manufacturer-described (e.g., defense contractor brochures).

6. Credibly reported (e.g., investigative journalism).

7. Demonstrated in testing (e.g., proving ground footage).

8. Alleged (e.g., unverified social media claims from conflict zones).

9. Disputed (e.g., conflicting accounts between state actors and independent analysts).

10. Operational status unspecified.

11. Control mode unspecified.

12. AI use not publicly established.

13. Outdated or superseded.

The visual design framework of the Atlas ensures that these states are immediately comprehensible without relying on color alone, maintaining strict accessibility compliance34. Verified facts (States 1, 2, 3\) are represented by solid, high-contrast borders, fully opaque node fills, and heavy connection lines within the graph view. Tooltips hovering over these nodes display a verified badge alongside the underlying Admiralty rating (e.g., "A1 \- Confirmed"). Conversely, manufacturer claims (State 5\) are visually demoted. They are represented by dashed borders, fifty-percent opacity fills, and distinct geometric shapes. An overarching watermark persistently labels the node as "Manufacturer Claim: Unverified Operationally." This visually reduces the weight of the information, forcing the user to acknowledge the lack of independent confirmation. Allegations and unspecified data (States 8, 10, 11, 12\) utilize dotted borders and hollow nodes, communicating the fragility of the underlying data. Disputed data (State 9\) is highlighted with a highly visible hazard pattern, prompting the user to investigate the conflicting source reports. By embedding the credibility of the information directly into the visual syntax of the map and the system network, the Atlas inherently guards against the propagation of false authority and automation bias11.

Map and Globe Interactions

The Atlas utilizes the deck.gl framework, a WebGL/WebGPU-powered architecture capable of rendering large-scale spatial and topological data with high performance36. However, to strictly comply with the mandate against displaying active tactical unit positions, exact current operational coordinates, or strike locations useful for future operations, all geospatial data within the Atlas is aggressively abstracted. Map interactions are restricted to country-level centroids, broad regional bounding boxes, or abstract topological representations. The primary interaction mode utilizes deck.gl's GlobeView for macro-level planetary exploration, allowing users to rotate and zoom fluidly before seamlessly transitioning to a MapView utilizing the Web Mercator projection for regional analysis38. Selecting a country polygon isolates that nation's data, triggering an overlay panel that displays the Country Profile Template while dimming the surrounding globe. To facilitate complex analysis, the Atlas features robust multidimensional filtering. Users can filter by date using a temporal slider, watching the map animate the diffusion of specific policies or weapon systems over time. Users can filter by kill-chain function to isolate all nations publicly fielding algorithmic decision support in the targeting phase, or filter by domain to view capabilities across air, land, sea, space, and cyber environments. Crucially, users can filter by control mode, highlighting the stark contrast between nations legally committing to human-in-the-loop policies versus those actively researching human-out-of-the-loop full autonomy14. The integration of the evidence state framework into the map interface provides the most powerful analytical tool. By filtering by evidence state, users can instantly strip away all "Manufacturer Claims" and "Allegations," watching the map physically shrink to display only "Independently Corroborated" and "Officially Documented" facts. When users elect to open system relationships, the interface transitions from geographic space into an abstract node-link graph. This topological view visualizes the DARPA Mosaic Warfare concept of kill webs, demonstrating how individual sensor and shooter nodes theoretically integrate across domains without revealing their physical deployment locations3. To support research and journalism, the Atlas allows users to save a view and share a view, generating a canonical URL that captures the exact camera angle, filter state, and selected entities. Every geospatial node features an open source trail button, providing a deep-link directly to the underlying OSINT archive. Finally, to ensure universal access, a persistent toggle allows users to switch to an accessible table mode, translating all WebGL visual data into a highly structured, screen-reader-compliant HTML data table34.

Historical Timeline Design

The timeline layer maps the chronological evolution of artificial intelligence in military applications, correlating policy shifts with hardware deployments. By presenting legal governance and technological advancement on a unified temporal axis, the timeline visualizes the inherent lag between algorithmic demonstration and the establishment of international humanitarian constraints. The timeline tracks a wide array of event types: the publication of foundational doctrinal texts, major policy changes, initial system announcements, controlled demonstrations, procurement contracts, actual fielding milestones, reported combat use, UN or NGO investigations into unintended engagements, internal legal reviews, and the signing of international declarations, such as the initial REAIM Summit in The Hague in 2023 and its subsequent milestones8. Furthermore, the timeline explicitly records corrections or changed assessments, ensuring that the historical record remains transparent when earlier intelligence is proven false. Every timeline event is structured as a robust data object containing exact or bounded dates. When precise days are unknown, the system accepts bounding ranges (e.g., Q3 2021\) without forcing a false precision. The event defines the event type, the primary country or actor, and the specific system or doctrine in question. Each node displays a factual-status label directly linked to the Evidence State taxonomy, followed by an objective, neutral short description of the occurrence. Crucially, the timeline exposes the methodological rigor behind the event by displaying the source class (e.g., the specific Admiralty Code rating)33 and a qualitative confidence score regarding the event's historical certainty. Related events are hyperlinked, allowing users to trace the causal chain from a 2012 policy directive10 to a 2018 procurement contract, to a 2022 reported deployment. Every node also displays a last verification timestamp, indicating the most recent editorial audit of the underlying source material.

Comparison Mode

The side-by-side comparison mode accommodates up to four countries or individual weapon systems simultaneously. This feature is explicitly engineered to support neutral, factual inquiries and actively subverts the simplistic "most advanced," "most lethal," or "best military" rankings that dominate superficial defense analysis. Such rankings are analytically shallow, highly subjective, and frequently rely on unverified manufacturer bravado rather than operational reality. Instead, the comparison dimensions are designed to evaluate entities across precise, legally and technically significant axes. When comparing countries, the matrix aligns their overarching doctrine, highlighting differing interpretations of human responsibility. When comparing systems, the matrix evaluates the primary intended function within the kill chain, the specific types of sensors utilized to gather environmental data24, and the algorithmic methodologies employed for decision support (e.g., neural networks used for trajectory prediction versus computer vision used for target classification)2. The most critical comparison dimensions focus on human-machine interaction. The matrix contrasts the engagement authority delegated to each system, exposing the exact threshold at which a machine can legally or practically apply force. This is paired with an analysis of human supervision requirements, determining whether an operator must physically authorize a strike, or if they merely retain a veto power over an autonomous process10. Operating constraints and known limitations are placed side-by-side, revealing vulnerabilities such as algorithmic brittleness, susceptibility to electronic warfare, or the risks of data-fusion error propagation11. To maintain epistemological integrity, the matrix includes an evidence quality row, directly comparing the Admiralty Code aggregates40 supporting the profile of each subject. It also contrasts the public deployment status, governance adherence, and the date coverage of the available data. By structuring the comparison around these dimensions, the Atlas supports nuanced, neutral questions: Which functions are publicly documented? What level of authority is delegated? What evidence supports that assessment? Which facts remain unknown? What governance controls are publicly described?

Twelve Interactive Explainers

To ensure users comprehend the complex sociotechnical, algorithmic, and legal concepts underpinning the Atlas, twelve interactive explainers are seamlessly embedded throughout the interface. These focused modules combine brief animations, simulations, and explanatory text to clarify the nuances of military AI.

1. AI-Enabled Does Not Necessarily Mean Autonomous: A 30-second interactive flowchart where the user toggles AI functionality for navigation versus targeting. It visually demonstrates how a drone utilizing advanced machine learning for flight stabilization and route planning can still require manual human authorization for weapons release10. Artifact: "Definitions of Autonomy" infographic card.

2. Human-In, On, and Out-of-Loop: A 45-second slider interaction moving along the control spectrum. The visualization transitions from a human physically pressing a trigger (in-the-loop), to a human monitoring a system with an active abort button (on-the-loop), to a system selecting and engaging targets entirely independently (out-of-the-loop)10. Artifact: "Control Spectrum" visualization.

3. F2T2EA: A 60-second step-by-step interactive cycle. Clicking a specific phase (e.g., "Fix" or "Track") highlights the specific types of sensors, computer vision algorithms, and decision-support systems that operate within that narrow window. Artifact: "Kill Chain Phases" diagram.

4. Kill Chain Versus Kill Web: A 45-second interaction where users click to "break" a node in a traditional linear kill chain. The visualization then animates the system dynamically re-routing communication protocols into a resilient, distributed multi-domain web, illustrating the core thesis of Mosaic Warfare3. Artifact: "Architecture of a Kill Web" GIF.

5. Sensor Confidence Versus Positive Identification: A 60-second simulated targeting reticle view. The user adjusts a slider to introduce environmental noise (such as rain, fog, or ego-noise from the drone's own motors)27. As noise increases, the AI's confidence score fluctuates wildly, demonstrating why a high probabilistic confidence score generated by an algorithm does not equate to the legal standard of Positive Identification required by international humanitarian law. Artifact: "Limits of Algorithmic PID" card.

6. Manufacturer Claim Versus Independent Verification: A 45-second split-screen interaction. The left side displays a sleek, buzzword-heavy PR brochure for a hypothetical weapon. The right side reveals the rigorous, step-by-step OSINT process5—geolocation, chronolocation, and cross-referencing—stripping away unverified claims to reveal the actual, limited fielded capability. Artifact: "Information Credibility" guide.

7. Demonstration Versus Operational Fielding: A 30-second interactive timeline illustrating the "valley of death" in defense acquisition. It shows the standard five-to-fifteen year gap between a viral, highly edited prototype test on a proving ground and actual, integrated, ruggedized military fielding15. Artifact: "Defense Acquisition Lifecycle" chart.

8. Defensive Automatic Engagement: A 45-second scenario simulation of systems like the Iron Dome or Phalanx CIWS. It visually demonstrates why split-second saturation attacks by incoming munitions necessitate human-supervised, but highly automated, defensive kinetic responses, and how these differ legally and ethically from offensive autonomous targeting2. Artifact: "Time-Critical Defense" graphic.

9. Terminal Autonomy: A 45-second visualization of a "fire-and-forget" missile. The user authorizes the launch at a designated geographic box, and the missile utilizes onboard sensors to navigate, identify, and track a pre-selected target signature without further human input during its terminal flight phase10. Artifact: "Semi-Autonomous Munitions" card.

10. Data-Fusion Error Propagation: A 60-second network graph simulation. The user introduces a False Positive at a single, low-level sensor node. The animation shows how that error cascades, compounds, and is treated as absolute truth as it moves up through an AI-enabled battle-management system, illustrating the dangers of automation bias2. Artifact: "Error Cascades in AI" visualization.

11. Communications-Loss Behavior: A 45-second scenario where an advanced drone loses its communication link to the human command center due to electronic jamming. The user explores the pre-programmed failsafe behaviors—loiter in place, return to base, or proceed to autonomous engagement—based on the constraints outlined in directives like DoD 3000.0919. Artifact: "Failsafe Mechanisms" diagram.

12. Why Public Information is Incomplete: A 30-second explainer on the "iceberg" nature of classified military data. Users interact with a transparency slider, demonstrating how OSINT methodologies only capture the unclassified ripples of highly classified programs, emphasizing that absence of evidence does not equal absence of capability5. Artifact: "The OSINT Iceberg" card.

Viral and Shareable Outputs

To maximize the educational impact of the Atlas and combat the proliferation of defense-related disinformation on social media, the platform is designed to facilitate the generation of high-fidelity, nuanced shareable outputs. These artifacts are dynamically generated utilizing HTML5 Canvas rendering and exported as optimized images or embeddable HTML snippets. The suite of safe shareable artifacts includes the Country evidence card, the System comparison matrix, the "What public evidence shows" summary card, the Human-control spectrum diagram, Historical timeline clips, Source-trail documentation cards, Saved spatial atlas views, and Embedded interactive mini-maps. Crucially, the design of these viral outputs incorporates strict contextual preservation constraints. Every generated artifact is cryptographically stamped and visually watermarked to retain its epistemological context, preventing the image from being repurposed as propaganda. Every artifact must persistently display the exact Factual-Status label associated with the data (e.g., prominently displaying "Manufacturer Claim" over a rendered system card). The artifact must include the generation date to prevent outdated information from circulating indefinitely, the exact source count supporting the visualization, and a persistent, scannable link back to the Atlas methodology page. Furthermore, fields where data is lacking are not omitted; rather, they are explicitly rendered with an "Unknown" label, visually highlighting the gaps in public knowledge. Finally, every geospatial or abstract map artifact must include a clear, unalterable disclaimer stating that the visual abstractions do not reveal operational locations or tactical targeting data.

Search and SEO Architecture

To ensure the Atlas serves as a definitive counterweight to defense marketing and misinformation, it employs a highly structured Search Engine Optimization (SEO) architecture. This strategy prioritizes the creation of high-value, canonical hubs over the generation of thousands of low-value, automated programmatic pages. Public page patterns are established with permanent, semantic URLs for Country profiles (e.g., /country/iso-code), System profiles (e.g., /system/standardized-name), Doctrine explainers (e.g., /doctrine/dod-3000-09), Comparison pages (e.g., /compare?entities=\[id1,id2\]), Historical timeline events, Control-mode definitions, and Evidence-methodology pages. The backbone of the SEO strategy is the implementation of rigorous structured data. All pages utilize JSON-LD schemas44. For system profiles, the Atlas utilizes customized schemas mapping to defense hardware ontologies, ensuring that search engine crawlers can programmatically parse the distinction between an "AI-enabled" sensor and an "Autonomous" weapon46. A dense, semantic internal linking structure connects timeline events to specific systems, systems to operating countries, and evidence ratings back to the methodology hub, seamlessly distributing page authority across the domain. Correction and update handling is treated as a core SEO function. Pages prominently feature HTML \<time\> tags for dateModified. When significant factual corrections occur based on new OSINT, the page utilizes HTTP headers and on-page notices to flag superseded information. Rather than deleting the page and generating a 404 error, the historical URL is retained for transparency, educating the user on how and why the evidentiary assessment evolved.

Accessibility and Performance Requirements

The Atlas adheres strictly to WCAG 2.2 AA standards, ensuring that highly complex WebGL data visualizations do not become barriers to access for users with disabilities34. While the deck.gl framework provides exceptional high-performance rendering for spatial and graph data36, HTML canvas elements are inherently opaque to assistive technologies like screen readers. To resolve this, the Atlas implements parallel DOM overlays and hidden semantic HTML tables that sync their state perfectly with the WebGL canvas. When a user filters the 3D globe, the hidden semantic table updates simultaneously, providing full screen-reader summaries of the spatial data distribution34. All interactive elements, including timeline scrubbing, map filtering, and comparison matrix adjustments, are fully navigable via standard keyboard controls. Visual accessibility is maintained through high-contrast modes and the non-color evidence state design, ensuring that differing line weights, dash arrays, and distinct iconography convey meaning entirely independent of hue. To ensure performance across diverse hardware, the Atlas incorporates a "Reduced Motion" toggle that disables complex particle animations and fluid globe rotations. For mobile environments or devices running in low-power mode, the application detects hardware capabilities and defaults to progressive loading strategies, aggressive clustering of map nodes, and a static SVG fallback if WebGL context creation fails. Finally, a dedicated CSS print view strips all interactive elements to format the data for physical printing, while a core "No-JavaScript" fallback allows users in low-bandwidth environments, or researchers operating on high-security networks that disable JavaScript execution, to access the raw text, tables, and critical OSINT source links.

Editorial Workflow

The integrity, authority, and safety of the Atlas rely entirely on its editorial workflow. This pipeline adapts the rigorous standards of the Berkeley Protocol on Digital Open Source Investigations and Bellingcat methodologies to the specific, highly technical domain of military artificial intelligence6. The workflow proceeds through eight distinct phases:

1. Draft: Analysts collect initial data points from news reports, government PDFs, and defense exhibitions.

2. Source Review: All primary links are immediately archived utilizing services like the Wayback Machine or Auto-Archiver tools31 to permanently prevent link rot and the silent deletion of defense claims.

3. Evidence Classification: The Admiralty Code7 is formally applied. Multiple independent analysts must achieve consensus on the rating, ensuring a rigorous distinction between source reliability and information credibility.

4. Legal/Safety Review: Before publication, content is audited to ensure it does not inadvertently expose sensitive operational targeting data, violate privacy norms, or provide actionable tactical instructions to combatants5.

5. Publication: The validated data is pushed to the live database, instantly updating the graph network and JSON-LD schemas.

6. Periodic Verification: Automated link-checking scripts run continuously, while high-profile systems undergo semi-annual expert audits.

7. Correction: If new evidence surfaces, the system is updated, triggering the SEO correction protocols.

8. Archive or Supersession: Outdated systems are gracefully retired but remain in the database for historical context.

The editorial rules governing this workflow are unyielding. Claims of active operational deployments require a minimum source threshold of two independent, corroborating sources (e.g., a UN investigative report cross-referenced with verified satellite imagery) to achieve a "Verified" state. Primary-source preference dictates that official government documents50 or direct, authenticated video evidence51 supersede secondary journalism. Allegations, classified claims, or unverifiable assertions from conflict zones are explicitly handled as "Alleged" or "Unknown"; the Atlas does not attempt to verify the unverifiable. Manufacturer-content rules dictate strict isolation; any claim of an advanced AI capability made by a developer is assumed to be theoretical or demonstrated only in highly controlled testing until independent field verification exists. Exact-date requirements demand bounding logic rather than guessing, and in the event of an evidence disagreement between sources, the system defaults to the lower confidence rating and displays the hazard/disputed visual state, forcing the user to acknowledge the friction in the public record. Finally, the removal of stale claims is executed transparently; debunked evidence is not deleted, but its evidence state is downgraded to "Proven false or misleading" (Admiralty F6)52, serving as an educational record of military disinformation.

Required Deliverables: Implementation Roadmap, Risks, and Acceptance

The execution of the Global Kill Chain Evidence Atlas is phased to manage immense technical complexity and rigorous editorial capacity.

MVP, Second Release, and Advanced Roadmap

MVP (Minimum Viable Product): Focuses on the core architecture and a limited, high-impact dataset. Features include the deployment of the basic Country and System templates, a Web Mercator 2D map utilizing deck.gl38, and the application of the Standardized Evidence States to a curated list of the top 20 most prominent global AI military systems. The MVP will launch with 3 of the 12 interactive explainers and integrate major doctrinal documents, including DoD Directive 3000.09 and the REAIM Political Declaration9. Second Release: Expands geographic scope, temporal depth, and analytical capability. Features include the activation of the 3D Globe visualization, the full Historical Timeline layer, the Comparison Mode matrix accommodating up to four systems, and the completion of 8 of the 12 explainers. The Shareable Artifacts generation engine will be fully implemented. The dataset will expand to over 100 systems, integrating transcripts and reports from the UN CCW Group of Governmental Experts on LAWS17. Advanced Roadmap: Introduces predictive modeling, deep systems analysis, and community integration. Features include the rollout of the abstract System-of-Systems "Kill Web" topology graphs3, completion of all 12 explainers, and advanced accessibility features including full WebGPU porting and parallel DOM mapping34. The data pipeline will open to a crowdsourced OSINT integration portal, governed by a rigorous triage and validation queue for independent researchers.

Risks and Mitigations

Acceptance Criteria

The Global Kill Chain Evidence Atlas will be considered ready for public deployment when the following criteria are met:

1. The information architecture seamlessly connects all ten defined layers, allowing a user to navigate from a macro-level treaty down to a specific OSINT source document without encountering dead ends or orphan nodes.

2. The deck.gl map visualizations successfully render abstracted country and regional data without inadvertently plotting specific tactical coordinates or revealing vulnerable infrastructure.

3. The Evidence States design passes rigorous WCAG 2.2 AA accessibility audits, proving that color contrast and non-color differentiation (shapes, line weights) successfully communicate data veracity.

4. The system architecture fundamentally prevents a user from viewing a manufacturer's capability claim without explicitly seeing its unverified evidentiary status attached.

5. All twelve interactive explainers are fully functional, performant, and consistently load in under three seconds on standard broadband connections.

6. The shareable artifact generator successfully stamps all output images with the generation date, factual-status label, source counts, and the required operational disclaimer.

7. The canonical SEO structure is actively generating valid JSON-LD schema objects for all major entities, correctly differentiating AI categories for search crawlers.

8. The editorial workflow has been successfully stress-tested by analysts using historical, highly complex, and contradictory data points, proving the methodology can handle real-world ambiguity and evidence disagreement.

Works cited

1. What is Mosaic Warfare? | CNAWS, https://cnaws.in/explainers/what-is-mosaic-warfare

2. Algorithmic targeting in the Iranian–Israeli confrontation: Technical realities, legal thresholds, and the boundaries of human control. \- F1000Research, https://f1000research.com/articles/14-1200

3. DARPA's Mosaic Warfare \- Multi Domain Ops, But Faster \- Breaking Defense, https://breakingdefense.com/2019/09/darpas-mosaic-warfare-multi-domain-ops-but-faster/

4. DARPA Tiles Together a Vision of Mosaic Warfare, https://www.darpa.mil/news/features/mosaic-warfare

5. Open-Source Intelligence in the Social Media Era Graham Overcash \- Moritz College of Law, https://moritzlaw.osu.edu/sites/default/files/2025-03/Graham%20Overcash%20Blog%202.pdf

6. Integrating OSINT into Justice Processes | Institute for War and Peace Reporting \- IWPR, https://iwpr.net/global-voices/integrating-osint-justice-processes

7. Admiralty code \- Wikipedia, https://en.wikipedia.org/wiki/Admiralty\_code

8. Building Consensus on the U.S. Framework for a Political Declaration on the Responsible Military Use of Artificial Intelligence and Autonomy \- State Department, https://2021-2025.state.gov/building-consensus-on-the-u-s-framework-for-a-political-declaration-on-the-responsible-military-use-of-artificial-intelligence-and-autonomy/

9. Political Declaration on Responsible Military Use of Artificial Intelligence and Autonomy, https://www.state.gov/bureau-of-arms-control-deterrence-and-stability/political-declaration-on-responsible-military-use-of-artificial-intelligence-and-autonomy

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

11. The risks and inefficacies of AI systems in military targeting support, https://blogs.icrc.org/law-and-policy/2024/09/04/the-risks-and-inefficacies-of-ai-systems-in-military-targeting-support/

12. Globalizing Responsible AI in the Military Domain by the REAIM Summit \- Just Security, https://www.justsecurity.org/85440/globalizing-responsible-ai-in-the-military-domain-by-the-reaim-summit/

13. \[2007.02106\] Automatic Target Recognition on Synthetic Aperture Radar Imagery: A Survey, https://arxiv.org/abs/2007.02106

14. Exploring the 2023 U.S. Directive on Autonomy in Weapon Systems \- CEBRI, https://cebri.org/revista/en/artigo/114/exploring-the-2023-us-directive-on-autonomy-in-weapon-systems

15. ISO/IEC/IEEE 24748-7 \- Systems and software engineering \- Life cycle management \- Part 7, https://standards.globalspec.com/std/13055852/iso-iec-ieee-24748-7

16. The effect of source reliability and information credibility on judgments of information quality in intelligence analysis | Judgment and Decision Making \- Cambridge University Press & Assessment, https://www.cambridge.org/core/journals/judgment-and-decision-making/article/effect-of-source-reliability-and-information-credibility-on-judgments-of-information-quality-in-intelligence-analysis/E67548E8010A47345C3439D45D9EC6B3

17. Lethal autonomous weapons systems : resolution / adopted by the General Assembly, https://digitallibrary.un.org/record/4071100

18. Challenges to Article 36 Reviews Posed by Autonomous Weapons Systems (AWS) (Chapter 8\) \- Cambridge University Press & Assessment, https://www.cambridge.org/core/books/international-law-and-weapons-review/challenges-to-article-36-reviews-posed-by-autonomous-weapons-systems-aws/A76B5E5318AF5B4C843F2ACD78F6BEC1

19. DoD Directive 3000.09, November 21, 2012; Incorporating Change 1, May 8, 2017, https://ogc.osd.mil/Portals/99/autonomy\_in\_weapon\_systems\_dodd\_3000\_09.pdf

20. DARPA's Mosaic Warfare — Multi Domain Ops, But Faster \- DSIAC, https://dsiac.dtic.mil/articles/darpas-mosaic-warfare-multi-domain-ops-but-faster/

21. Political Declaration on Responsible Military Use of Artificial Intelligence and Autonomy, https://2021-2025.state.gov/bureau-of-arms-control-deterrence-and-stability/political-declaration-on-responsible-military-use-of-artificial-intelligence-and-autonomy/

22. Review of the 2023 US Policy on Autonomy in Weapons Systems | Human Rights Watch, https://www.hrw.org/news/2023/02/14/review-2023-us-policy-autonomy-weapons-systems

23. DOD Is Updating Its Decade-Old Autonomous Weapons Policy, but Confusion Remains Widespread \- CSIS, https://www.csis.org/analysis/dod-updating-its-decade-old-autonomous-weapons-policy-confusion-remains-widespread

24. \[2107.11453\] Automatic Detection Of Noise Events at Shooting Range Using Machine Learning \- arXiv, https://arxiv.org/abs/2107.11453

25. \[2411.03491\] An Application-Agnostic Automatic Target Recognition System Using Vision Language Models \- arXiv, https://arxiv.org/abs/2411.03491

26. Military AI Needs Technically-Informed Regulation to Safeguard AI Research and its Applications \- arXiv, https://arxiv.org/html/2505.18371v1

27. arXiv:2303.15042v1 \[eess.AS\] 27 Mar 2023, https://arxiv.org/pdf/2303.15042

28. Alert or Noise? Reducing False Positives with Active Behavioral Analysis for Cloud Security, https://arxiv.org/html/2508.12584v1

29. Policy Considerations of Open-Source Intelligence: A Study of Bellingcat's Online Investigation Patterns (2014-2024) \- Virtual Commons \- Bridgewater State University, https://vc.bridgew.edu/cgi/viewcontent.cgi?article=1202\&context=ijcic

30. UNCERTAINTY IN ANALYSIS OF HYBRID CONFLICT \- NATO, https://publications.sto.nato.int/publications/STO%20Meeting%20Proceedings/STO-MP-IST-190/MP-IST-190-13.pdf

31. Bellingcat's Online Investigation Toolkit: Home, https://bellingcat.gitbook.io/toolkit

32. Lessons from Building an Online Toolkit to Aid Open-Source Investigations, https://niemanreports.org/osint-open-source-investigations-bellingcat-volunteers/

33. Enhance your Cyber Threat Intelligence with the Admiralty System | SANS Institute, https://www.sans.org/blog/enhance-your-cyber-threat-intelligence-with-the-admiralty-system

34. Clarify how canvas fallback content relates to the accessibility tree · Issue \#7490 \- GitHub, https://github.com/whatwg/html/issues/7490

35. Military AI Needs Technically-Informed Regulation to Safeguard AI Research and its Applications \- arXiv, https://arxiv.org/html/2505.18371v2

36. WebGPU \- Deck.gl, https://deck.gl/docs/developer-guide/webgpu

37. Introduction \- Deck.gl, https://deck.gl/docs

38. Views and Projections | deck.gl, https://deck.gl/docs/developer-guide/views

39. 13176 – The bounds of canvas fallback content, as rendered on the canvas, are not provided by the user agent to an assistive technology. \- W3C, https://www.w3.org/Bugs/Public/show\_bug.cgi?id=13176

40. NATO AJP-2.1 Source Reliability and Information Credibility Scales \- ResearchGate, https://www.researchgate.net/figure/NATO-AJP-21-Source-Reliability-and-Information-Credibility-Scales\_tbl1\_328858953

41. Real-time vehicle target detection in inclement weather conditions based on YOLOv4 \- PMC, https://pmc.ncbi.nlm.nih.gov/articles/PMC10034385/

42. Solving May's Open Source Challenge by Bellingcat \- Adina, https://traceblog.medium.com/solving-mays-open-source-challenge-by-bellingcat-6cef92910055

43. \[2606.29026\] Preventing Error Propagation in Multi-Agent AI through Runtime Monitoring, https://arxiv.org/abs/2606.29026

44. JSON-LD 1.1, https://nhqc3s.hq.nato.int/apps/DCRA\_Report/id-29d4122b072148f5aaf4882ecc5d963c/elements/id-1c7c2c83-ad55-11ef-4529-005056901351.html

45. OGC Disaster Pilot JSON-LD Structured Data Engineering Report, https://docs.ogc.org/per/21-054.html

46. JSON-LD Schema Visualization \- ToDiagram, https://todiagram.com/use-cases/jsonld-schema-visualization

47. What is JSON-LD and how to implement Schema.org? \- FRT Digital, https://frt.digital/en/articles/what-is-json-ld-and-how-to-implement-schema-org

48. Guides & Handbooks \- Bellingcat's Online Investigation Toolkit, https://bellingcat.gitbook.io/toolkit/resources/guides-and-handbooks

49. Bellingcat \- GitHub, https://github.com/bellingcat

50. 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

51. Open source intelligence, with Bellingcat's Giancarlo Fiorella \- YouTube, https://www.youtube.com/watch?v=dVWyNUmUvMA

52. Admiralty code for the verification of information \- Enterprise Open Systems GmbH, https://eosgmbh.com/en/admiralty-code-for-the-verification-of-information/

53. Lethal autonomous weapons systems: Report of the Secretary-General \- UNDOCS \- Welcome to the United Nations, https://docs.un.org/en/a/79/88

54. Autonomous Weapon Systems and AI-enabled Decision Support Systems in Military Targeting: A Comparison and Recommended Policy Responses | SIPRI, https://www.sipri.org/publications/2025/other-publications/autonomous-weapon-systems-and-ai-enabled-decision-support-systems-military-targeting-comparison-and