THE ALGORITHMIC GATEKEEPER

Full report

On this page

Executive Concept

The proliferation of artificial intelligence within military and intelligence workflows has initiated a profound shift in the architecture of decision-making. The contemporary defense discourse frequently relies on the assurance that a "human-in-the-loop" (HITL) or "human-on-the-loop" (HOTL) mechanism will prevent automated systems from executing unintended lethal actions. This focus on the terminal authorization of force, however, obscures a pervasive vulnerability embedded much earlier in the kill chain: the phenomenon of upstream human removal. When an algorithm filters thousands of sensor observations, discards the vast majority of ambiguous data, synthesizes the remaining fragments into composite tracks, and presents a highly constrained, ranked list of target nominations to an operator, the human is no longer making an independent decision. Instead, the operator is merely ratifying the machine's curated framing of reality. This document serves as the comprehensive architectural, technical, and editorial specification for an interactive WebXR experience titled The Algorithmic Gatekeeper, designed for deployment on KillChains.com. As an educational and non-operational product, this immersive simulation places the user inside a spatial intelligence workflow to expose the cognitive and structural vulnerabilities inherent in AI-enabled decision chains. By utilizing synthetic data and Three.js-powered spatial environments, the experience visually and cognitively demonstrates how upstream algorithmic filtering preconditions downstream human decisions. The primary mission of the product is to pose a profound epistemological question to operators, policymakers, and the public: Can a human meaningfully decide when an algorithm has already selected the evidence, framed the alternatives, ranked the candidates, and hidden everything below a statistical threshold?

Explanation of Upstream Human Removal

The integration of machine learning and computer vision into intelligence, surveillance, target acquisition, and reconnaissance (ISTAR) workflows fundamentally alters the traditional OODA loop (Observe, Orient, Decide, Act). In the modern combat environment, the observation and orientation phases are increasingly delegated to automated systems capable of ingesting terabytes of data across multiple domains1. This compression creates an AI-augmented decision environment where human cognition is rapidly outpaced by machine synthesis, leading to severe cognitive friction and automation bias2. Automation bias is defined as the human tendency to over-rely on automated decision support systems as a heuristic replacement for vigilant information seeking and processing2. Engineering psychologists classify the errors stemming from this bias into two categories: errors of commission, where users follow incorrect machine advice, and errors of omission, where users fail to act because the automation did not provide a prompt or alert2. Research demonstrates that in time-critical, high-workload environments, human operators will inevitably defer to machine outputs, often disregarding contradictory evidence or failing to search for underlying data provenance2. This vulnerability is amplified by the perceived pedigree of the automated system; when an algorithm presents a target nomination with a mathematically precise confidence score, the human operator naturally assumes the underlying data is equally precise, conflating algorithmic confidence with evidence quality5. The concept of meaningful human control is severely degraded by the architecture of choice presented to the operator. Choice architecture, a principle of behavioral economics, dictates that the presentation of alternatives inherently influences the decision-maker's final selection7. In intelligence systems, the algorithm acts as the ultimate choice architect. By determining the default option, ranking candidates, and obscuring discarded data behind opaque interfaces, the system exerts an implicit endorsement that manipulates the operator's decision space8. The human operator, lacking the time and cognitive bandwidth to manually audit the 179 distinct sensor feeds that generated a single composite track, is forced to operate within the epistemological boundaries drawn by the algorithm10. The operator does not decide; the operator merely approves the machine's framing.

Six-Case Evidence Matrix and Real-World Case Cards

To ground the interactive simulation in verified public research, the experience integrates six documented case studies and programmatic analogies. These cases illustrate the rapid evolution of sensor fusion, computer vision, and the risks of unchecked automation bias. The data must be presented to the user through interactive case cards that strictly avoid implying that any of these systems operate fully autonomously without doctrine-mandated human oversight.

Case 1: Project Maven and the Maven Smart System (MSS)

Project Maven exemplifies the transition from basic computer vision to comprehensive, multi-domain battle management. Originally established to reduce the human burden of processing tactical unmanned aerial system video, the program evolved into the Maven Smart System, which standardizes highly heterogeneous data through an ontology layer10. The system acts as the ultimate algorithmic gatekeeper, utilizing AI to identify patterns, detect objects, and route prioritized target information directly to command nodes. The human operator is relegated to a highly curated interface where they review ranked recommendations, effectively operating within the epistemological boundaries drawn by the machine11.

Case 2: Tactical Intelligence Targeting Access Node (TITAN)

The TITAN program mirrors the trajectory of Project Maven but is optimized for the tactical edge. Designed as a scalable, next-generation ground station, TITAN leverages machine learning to ingest massive volumes of sensor data across all domains14. By automating the data fusion process, TITAN dramatically compresses the traditional intelligence cycle. However, this compression relies entirely on the premise that the underlying algorithms accurately classify and correlate raw signals before presenting the synthesized intelligence to the human analyst, masking the underlying uncertainty of the raw data.

Case 3: Space Development Agency (SDA) Tracking Layer

The SDA Tracking Layer introduces an unprecedented scale of sensor data generation. As a proliferated network of low Earth orbit satellites, it is designed to provide global, continuous missile warning and tracking, specifically targeting advanced hypersonic threats16. The requirement to generate "fire control-quality tracks" from space necessitates advanced edge processing and data fusion across multiple orbital regimes16. The sheer volume and velocity of this data mandate algorithmic filtering; human operators cannot manually correlate infrared signatures from dozens of satellites in real time, forcing complete reliance on the machine's synthesis.

Case 4: Integrated Battle Command System (IBCS)

The Integrated Battle Command System demonstrates the absolute necessity, and inherent risk, of multi-domain sensor fusion. By establishing a "connect any sensor to any shooter" architecture, IBCS aggregates disparate data streams into a single composite track20. While this significantly enhances the situational awareness of the air defense battalion, the reliance on automated sensor fusion means that any upstream misclassification by a contributing sensor—such as an F-35 or a Patriot radar—is inherited and potentially amplified by the composite track, leaving the human operator with a pristine but potentially flawed representation of the battlespace.

Case 5: Civilian Analogy — Healthcare Risk Algorithm Bias

To clarify the human-factors problem without equating civilian systems to lethal weapons, this 2019 study on algorithmic bias serves as a vital pedagogical tool. The study revealed that a widely deployed healthcare algorithm exhibited severe racial bias because it utilized healthcare costs as a proxy for healthcare needs23. Because Black patients historically incurred lower costs due to systemic unequal access to care, the algorithm systematically underrated the severity of their illnesses. The human physicians downstream were entirely dependent on the machine's flawed framing. The algorithm hid the truly sick patients below the presentation threshold, vividly demonstrating how a flawed upstream proxy variable dictates and restricts the downstream human decision space23.

Case 6: Historical Error — The 2003 Patriot Fratricides

The 2003 Patriot Missile fratricides provide the historical anchor for the catastrophic consequences of automation bias in military systems. During Operation Iraqi Freedom, the Patriot system's automated tracking algorithms misclassified a British Tornado as an anti-radiation missile, and subsequently, a US F/A-18 as a tactical ballistic missile26. Operating in a tightly coupled, time-compressed environment, the human operators had roughly one minute to challenge the machine's authoritative classification. Due to complex user interfaces, severe cognitive overload, and an inherent trust in the system's automated processing, the operators committed fatal errors of commission, authorizing the engagements3. The Defense Science Board subsequently identified these flaws not merely as isolated operator errors, but as systemic failures in human-machine integration, software design, and the psychological realities of supervisory control30.

Complete Interactive Storyboard

The user journey through The Algorithmic Gatekeeper is designed as a six-to-nine-minute spatial simulation utilizing WebXR and Three.js. The interaction design leverages the psychological impact of spatial compression and expansion—first demonstrating the overwhelming volume of raw data, and then manipulating the user's perception through algorithmic filtering. To maintain the necessary 90 frames per second on standard standalone VR hardware, the raw data visualization utilizes advanced rendering techniques, specifically instanced rendering and per-mesh face octree spatial data structures for rapid raycasting against thousands of objects31.

Stage 1 — The Raw Data Wall

The visitor spawns in a vast, darkened spatial void. Surrounding them is a colossal, 360-degree cylindrical wall comprising exactly 10,000 synthetic observations. These observations are rendered as a chaotic mosaic of abstract image tiles, scrolling alphanumeric sensor logs, radar heatmaps, and fluctuating confidence values. The environment is intentionally overwhelming. The spatial audio engine, leveraging the Web Audio API's PannerNode and browser-specific Head-Related Transfer Functions (HRTFs), generates a precise but chaotic soundscape of overlapping alerts, simulating a high-stress command post32. The visual data contains duplicate observations of identical events, stale tracks with visibly outdated timestamps, highly confident but contradictory reports flashing in proximity, and a massive volume of benign, low-relevance background noise. The Prompt: High-contrast text appears in the center of the user's vision: "Identify five critical anomalies. You have 30 seconds." The visitor attempts to locate anomalies by physically turning, pointing, and clicking using standard WebXR ray-based pointer events33. It is structurally and cognitively impossible to succeed. The application silently records user metrics: items registered in the primary viewport, items missed, dwell time per fixation, source diversity accessed, and the sheer volume of unreviewed evidence. Real eye tracking and biometric data are strictly prohibited. The objective is to induce the visceral reality of cognitive saturation, establishing the absolute necessity for machine assistance in modern intelligence workflows.

Stage 2 — The Machine Filter

At the conclusion of the 30-second stress test, an abrupt, low-frequency alarm sounds, and the chaotic environment freezes. A deep, calming auditory tone washes over the spatial environment, signaling a state change. A massive digital sweep line descends across the Raw Data Wall. The Prompt: "Activating Algorithmic Gatekeeper." Instantly, 9,988 of the chaotic data tiles shatter and dissolve into the darkened background. The spatial audio cacophony is instantly muted. The remaining twelve tiles glide forward smoothly, perfectly organizing into a clean, hierarchical list floating at eye level. The system assigns neat, authoritative confidence scores to each item (e.g., "98% Confidence: Hostile Track"). The user's field of view is drastically narrowed, and the psychological relief is palpable. The user is now instructed to review these twelve pristine items and make a final determination on the top-ranked candidate. The experience intentionally mirrors the epistemic closure experienced by real-world operators; it does not reveal whether the filter's output is actually correct, forcing the user to rely entirely on the machine's unverified confidence score6.

Stage 3 — Threshold Control

Having experienced the relief of automation, the visitor is now granted access to the internal mechanics of the machine filter. A spatial control panel materializes, featuring a series of sliders and toggle switches representing the hidden architecture that dictates what survives the filter.

As the user manipulates these sliders, the visual interface responds dynamically. Lowering the thresholds causes the wall to flood with hundreds of candidates, flashing warnings of a severe "False-Positive Burden," accompanied by a rising swell of spatial audio. The user realizes that there is no perfect mathematical setting; every adjustment trades one type of catastrophic operational risk for another. The user is eventually forced to lock in a configuration and authorize a synthetic action based on their customized gatekeeper.

Stage 4 — What the Human Never Saw

This stage constitutes the educational and emotional climax of the simulation. Upon the user authorizing the action, the environment turns stark, clinical white. Time is artificially rewound to the moment just before the decision. The Prompt: "Did you make a decision, or did you merely approve the machine's framing?" The environment reconstructs the exact interface the user just authorized, but now, the 9,988 hidden items are rendered in a ghostly, luminescent red hue behind the primary interface. The system highlights specific, catastrophic failures of the algorithm that the user just implicitly trusted:

1. The Flawed Lineage: A high-confidence (98%) target nomination is visually traced backward through the fusion engine. The system reveals that it relied on two supposedly independent sensor reports that actually shared the same flawed origin point, demonstrating the severe danger of circular reporting masked by fusion.

2. The Hidden Contradiction: A machine-generated summary of a sector is shown alongside a discarded, low-confidence human intelligence report. The text summary explicitly omitted the human report, which directly contradicted the primary radar track.

3. The Stale Track: A track presented as current and actionable was actually based on stale data that bypassed the user's staleness tolerance due to an upstream metadata parsing error.

The user is forced to physically walk around and inspect the discarded, ghostly data. The realization sets in: the machine's certainty was an illusion generated by the mathematical discarding of uncertainty. The user's downstream decision was entirely preconditioned by the upstream algorithmic thresholds.

Stage 5 — Provenance-Aware Control

In the final stage, the user is empowered to redesign the gatekeeper to mitigate automation bias, utilizing architectural constraints based on human-factors engineering rather than statistical thresholds.

The user selects their desired architecture, and the scenario replays. The interface is slightly more visually complex, but the user is now structurally equipped to spot the circular reporting and the hidden contradiction. The user makes a highly informed decision to reject the flawed algorithmic recommendation. The experience concludes with a powerful message on the absolute necessity of data provenance and systemic doubt in AI integration.

The Algorithmic Funnel

To ensure the user maintains situational awareness of the filtering process throughout the simulation, a persistent 3D visual metaphor is maintained in the user's peripheral environment, accessible via a quick-glance gesture: The Algorithmic Funnel. This is rendered as a narrowing, translucent three-dimensional tunnel. As synthetic data points move through the tunnel, they encounter virtual meshes representing algorithmic filters. Crucially, discarded data points do not disappear. Instead, they are deflected outside the main path, remaining visible as a massive, hovering cloud of rejected context. The pipeline stages are explicitly labeled, and when the user gazes at a specific narrowing point, an interactive tooltip reveals the mechanics of the transformation:

This visual model actively combats the psychological tendency to assume that what is presented on the screen is the entirety of reality, constantly reminding the user of the sheer volume of data discarded to produce the current view.

The Choice-Set Explainer

Following the primary narrative, a standalone interactive module demonstrates how algorithms influence human choice without ever possessing the explicit authority to execute lethal force. This module is grounded in behavioral economics and choice architecture research, which demonstrates that the presentation of a choice significantly alters the outcome7. The user participates in a rapid-fire A/B/C test involving a simplified scenario (e.g., dispatching a reconnaissance drone to one of several sectors).

• Condition A (Raw Choice): The human receives all 10 possible sectors on a flat map with equal visual weighting.
• Condition B (Algorithmic Ranking): The human receives a list of the 10 sectors, but an AI has ranked them 1 through 10, applying bright red colors to the top three and graying out the bottom seven.
• Condition C (Algorithmic Framing): The human receives only 1 recommended option, flanked by two highly flawed decoy alternatives. The other 7 options are hidden entirely behind an "Advanced Options" dropdown.

After the user makes their selections, the module immediately explains the psychological manipulation. It reveals aggregated data demonstrating how presentation order, color-coding (confidence language), and omission (decoy effects and default options) reliably push users toward the algorithm's preferred choice, even when the human formally retains absolute authority. The module emphasizes that whoever controls the choice architecture controls the decision7. The user is explicitly informed that they were not manipulated deceptively for data collection, but rather to personally experience the mechanics of choice architecture.

Country and System Integration Recommendations

To ground the interactive specification in modern military realities without utilizing classified data, the product references standard integration frameworks, specifically focusing on data interoperability and multi-domain command and control (CJADC2). The most critical framework for this context is the NATO Federated Mission Networking (FMN) initiative. FMN is a governed conceptual framework designed to enable rapid instantiation of mission networks by federating NATO organizations and partner capabilities34. The current operational baseline, FMN Spiral 4, introduces rigorous service-level requirements, specific security characteristics, and data-centric security (DCS) protocols35. When algorithms fuse data across national boundaries—for example, integrating a UK sensor feed with a US processing node—the preservation of data provenance is paramount. The Algorithmic Gatekeeper must recommend that any algorithmic filtering system integrated into FMN support Confidentiality Metadata Based Access Control (CMBAC) standardized in STANAG 567836. Furthermore, the system must preserve the origin tags of fused data across XMPP and messaging services, ensuring that downstream commanders can audit the confidence scores of multi-national composite tracks36. Additionally, as algorithms increasingly move to the edge (e.g., onboard SDA satellites or TITAN ground vehicles) to reduce latency18, the product recommends architectures that transmit a lightweight cryptographic hash or summary of the discarded data's parameters. This allows the downstream human interface to mathematically prove how much data was omitted, preserving trust and auditability even in bandwidth-constrained environments.

Accessibility Specification

The interactive experience must be designed in strict accordance with the W3C WebXR Accessibility User Requirements (XAUR)38. Immersive environments frequently exclude individuals with physical or cognitive disabilities; The Algorithmic Gatekeeper is designed to serve as a benchmark for inclusive spatial design. To support users with physical disabilities, the application implements motion-agnostic interaction. Users must be able to interact with the Raw Data Wall and the Threshold Control panel without requiring full six-degrees-of-freedom (6DoF) physical movement. The system supports alternative mappings, allowing gaze-tracking, traditional keyboard/mouse inputs, and specialized adaptive controllers to manipulate the 3D environment seamlessly38. For deaf or hard-of-hearing users, the cacophony of the Raw Data Wall—which relies heavily on spatialized sound—is augmented by a spatialized subtitle system. Text descriptors of the audio (e.g., \[High-pitched radar lock, 45 degrees left\]) are visually anchored in the 3D space relative to the user's field of view, ensuring equal access to environmental cues39. Furthermore, to accommodate users with cognitive disabilities who may experience severe distress from the intentional cognitive overload in Stage 1, the application includes a mandatory pre-launch personalization menu. This menu allows users to cap the rendering volume, slow the time limit, or apply a high-contrast mode to the abstract data tiles38.

Viral Outputs and Shareable Artifacts

To maximize the educational reach of the simulation, the conclusion of the experience generates highly customized, shareable artifacts that summarize the user's specific journey and vulnerability to automation bias. The primary artifact is a dynamically generated Result Card titled: WHAT THE ALGORITHM LET ME SEE.

Additional viral and educational outputs include:

• A six-second collapsing-data-wall WebGL loop suitable for social media embedding.
• A fifteen-second “what was hidden” reveal video snippet.
• A friend challenge URL, allowing colleagues to experience the exact same synthetic data seed and threshold constraints for direct comparison.
• A classroom threshold debate guide for academic integration.
• An embeddable evidence funnel interactive widget.
• A text-only accessible audit report, providing a screen-reader-optimized breakdown of the scenario's data lineage.
• A source-provenance share card detailing the dangers of circular reporting.

The system ensures that these artifacts strictly utilize abstract, synthetic data, explicitly disclaiming any connection to real military operations or target processing.

Analytics and Measurement

Data collection is strictly limited to non-personally identifiable interactions designed exclusively to measure the educational efficacy of the product. The application explicitly prohibits the collection of raw gaze data, precise headset pose streams, biometric markers, geolocation, or personal behavioral profiles.

Editorial, Safety Controls, and Acceptance Criteria

Given the highly sensitive nature of military command and control analogies, the product adheres to strict editorial and safety controls to ensure it remains a purely educational tool regarding algorithmic assurance and human-factors engineering. The non-negotiable rules governing the application dictate that the simulation must contain zero live intelligence, real target data, or operational coordinates. The application must not utilize any facial recognition technology or simulate the targeting of real individuals. The system must not optimize weapon-target assignments or provide functional target-nomination algorithmic code. The user interface must never assume or imply that a high-confidence algorithmic score equates to ground-truth positive identification. Furthermore, the narrative must explicitly reject the implication that AI support equates to autonomous lethal force.

Release Roadmap and Acceptance Criteria

The development and deployment of The Algorithmic Gatekeeper follows a structured, multi-phase release roadmap:

• Phase 1: Prototyping (Weeks 1-4): Develop the WebXR instanced rendering engine and octree spatial data structures. Finalize the synthetic data generation script to ensure repeatable, deterministic scenarios.
• Phase 2: Narrative Integration (Weeks 5-8): Implement the spatial audio design and UI elements for the Machine Filter and Threshold Control stages. Integrate the Six-Case Evidence Matrix into the educational codex.
• Phase 3: Accessibility & Optimization (Weeks 9-11): Implement motion-agnostic controls, spatial subtitles, and finalize rendering optimizations.
• Phase 4: Beta Testing (Weeks 12-13): Deploy to a closed group of human-factors analysts and UI/UX designers to tune the difficulty and emotional resonance of the narrative reveal.
• Phase 5: Launch: Public deployment on KillChains.com alongside the release of the embeddable widgets and challenge codes.

Acceptance Criteria for Release:

1. Performance Baseline: The WebXR experience must maintain a minimum of 90 frames per second on standard standalone VR hardware during the rendering of the 10,000-item Raw Data Wall.

2. Accessibility Compliance: The product must pass a comprehensive audit against W3C XAUR guidelines, proving functional playability without 6DoF tracking.

3. Educational Efficacy: Beta testing must demonstrate that at least 70% of users successfully identify the concepts of "circular reporting" and "automation bias" in a post-experience questionnaire.

4. Safety and Compliance Review: A final code and asset review confirming the absolute absence of real-world geospatial, intelligence, or sensitive sensor data.

Works cited

1. Reshaping Air Power Doctrines: Creating AI-Enabled 'Super-OODA Loops' \- Shift Paradigm., https://theairpowerjournal.com/reshaping-air-power-doctrines-creating-ai-enabled-super-ooda-loops/

2. Automation Bias in Intelligent Time Critical Decision Support Systems, https://maritimesafetyinnovationlab.org/wp-content/uploads/2023/02/Automation-Bias-in-Intelligent-Time-Critical-Decision-Support-Systems.pdf

3. Automation bias: a systematic review of frequency, effect mediators, and mitigators \- PMC, https://pmc.ncbi.nlm.nih.gov/articles/PMC3240751/

4. Automation bias: a systematic review of frequency, effect mediators, and mitigators, https://www.researchgate.net/publication/51230614\_Automation\_bias\_a\_systematic\_review\_of\_frequency\_effect\_mediators\_and\_mitigators

5. Automation bias: a systematic review of frequency, effect mediators, and mitigators, https://pubmed.ncbi.nlm.nih.gov/21685142/

6. Full article: The AI Commander Problem: Ethical, Political, and Psychological Dilemmas of Human-Machine Interactions in AI-enabled Warfare \- Taylor & Francis, https://www.tandfonline.com/doi/full/10.1080/15027570.2023.2175887

7. (PDF) Beyond nudges: Tools of a choice architecture \- ResearchGate, https://www.researchgate.net/publication/236302915\_Beyond\_nudges\_Tools\_of\_a\_choice\_architecture

8. Choice Architecture, Framing, and Cascaded Privacy Choices Idris Adjerid\*, Alessandro Acquisti+, George Loewenstein+ For consume, https://www.heinz.cmu.edu/\~acquisti/papers/Acquisti\_Choice\_Architecture\_Frame\_Cascaded\_Privacy\_Choices.pdf

9. Nudging for eco-friendly online shopping \- DiVA Portal, https://www.diva-portal.org/smash/get/diva2:1688722/FULLTEXT01.pdf

10. What Is Maven Smart System, and What Does It Do? \- CSIS, https://www.csis.org/analysis/what-maven-smart-system-and-what-does-it-do

11. Project Maven \- Wikipedia, https://en.wikipedia.org/wiki/Project\_Maven

12. Project Maven, https://dodcio.defense.gov/Portals/0/Documents/Project%20Maven%20DSD%20Memo%2020170425.pdf

13. PROJECT MAVEN | The Architecture of Algorithmic Warfare | by Mohamed Salah \- Medium, https://medium.com/@m.salah2405/project-maven-the-architecture-of-algorithmic-warfare-3ff147e7b520

14. Army Tactical Intelligence Targeting Access Node (TITAN) Ground Station Prototype \- Award, https://cpeisw.army.mil/2024/03/06/army-tactical-intelligence-targeting-access-node-titan-ground-station-prototype-award/

15. Palantir wins $178M Army deal for TITAN artificial intelligence-enabled ground stations, https://defensescoop.com/2024/03/06/palantir-army-titan-ground-station-award-178-million/

16. Tracking \- Space Development Agency, https://www.sda.mil/tracking/

17. ON ORBIT \- Space Development Agency (SDA), https://www.sda.mil/on-orbit/

18. Getting on Track, https://csis-website-prod.s3.amazonaws.com/s3fs-public/2023-12/231218\_Dahlgren\_Getting\_Track\_0.pdf?VersionId=gyTyKePGJmFvnZmTgQY5.\_GidZ0jfGh4

19. Space Development Agency \- Wikipedia, https://en.wikipedia.org/wiki/Space\_Development\_Agency

20. Integrated Air and Missile Defense Battle Command System \- Wikipedia, https://en.wikipedia.org/wiki/Integrated\_Air\_and\_Missile\_Defense\_Battle\_Command\_System

21. Integrated Air and Missile Defense \- Wikipedia, https://en.wikipedia.org/wiki/Integrated\_Air\_and\_Missile\_Defense

22. Missile defence: Integrated Battle Command System (IBCS) demonstrates operational readiness in flight test, https://defence-industry.eu/missile-defence-integrated-battle-command-system-ibcs-demonstrates-operational-readiness-in-flight-test/

23. Dissecting racial bias in an algorithm used to manage the health of populations, https://www.ftc.gov/system/files/documents/public\_events/1548288/privacycon-2020-ziad\_obermeyer.pdf

24. Dissecting Racial Dias in an Algorithm Used to Manage the Health of Populations, https://just-tech.ssrc.org/citation/dissecting-racial-bias-in-an-algorithm-used-to-manage-the-health-of-populations/

25. Dissecting racial bias in an algorithm used to manage the health of populations \- RRAPP, https://rrapp.spia.princeton.edu/algorithms-can-replicate-or-remedy-racial-biases-in-healthcare-resource-allocation/

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

27. Patriot Wars \- CNAS, https://www.cnas.org/publications/reports/patriot-wars

28. Lessons from air defence systems on meaningful human control for the debate on AWS Ingvild Bode and Tom Watts \- Drone Wars UK, https://dronewars.net/wp-content/uploads/2021/02/DW-Control-WEB.pdf

29. Not by widgets alone \- Armed Forces Journal, https://armedforcesjournal.com/not-by-widgets-alone/

30. Machine Failing: How Systems Acquisition and Software Development Flaws Contribute to Military Accidents \- Texas National Security Review, https://tnsr.org/2024/10/machine-failing-how-systems-acquisition-and-software-development-flaws-contribute-to-military-accidents/

31. Octree injection for faster raycasting/raytracing \- Showcase \- three.js forum, https://discourse.threejs.org/t/octree-injection-for-faster-raycasting-raytracing/8291

32. Spatial Audio for WebXR: Perceptual Evaluation of Sound Localization Technologies on the Browser \- Luca Turchet, http://www.lucaturchet.it/PUBLIC\_DOWNLOADS/publications/conferences/Spatial\_Audio\_for\_WebXR-Perceptual\_Evaluation\_of\_Sound\_Localization\_Technologies\_on\_the\_Browser.pdf

33. Meet Immersive Web SDK: A New Era for Spatial Web Development \- Meta for Developers, https://developers.meta.com/horizon/blog/immersive-web-sdk-new-era-spatial-web-development/

34. Federated Mission Networking \- NATO's ACT, https://www.act.nato.int/activities/federated-mission-networking/

35. FMN spiral 4: what the latest NATO federation standard \- Corvus Intelligence, https://corvusintell.com/blog/interoperability/fmn-spiral-4-requirements/

36. Data Centric Security and Federated Mission Networking \- Isode, https://www.isode.com/whitepaper/data-centric-security-and-federated-mission-networking/

37. Algorithmic Warfare Cross Functional Team (AWCFT) 590\_0307588D8Z\_6\_0400\_PB\_2022, https://dtic.minsky.ai/590\_0307588D8Z\_6\_0400\_PB\_2022/text

38. XR Accessibility User Requirements \- W3C, https://www.w3.org/TR/xaur/

39. Report from W3C Workshop on Inclusive Design for Immersive Web Standards, https://www.w3.org/2019/08/inclusive-xr-workshop/report.html