Algorithmic Warfare Cross-Functional Team (AWCFT) / Project Maven

Project Maven, officially the Algorithmic Warfare Cross-Functional Team (AWCFT), was established in a memorandum signed by Deputy Secretary of Defense Robert O. Work on 26 April 2017. The memorandum directed the Department of Defense to accelerate the integration of big data, machine learning, and computer vision into intelligence workflows, beginning with the processing, exploitation, and dissemination (PED) of full-motion video (FMV) from tactical and medium-altitude unmanned aerial systems. The program was initially funded at approximately $70 million, placed under the Under Secretary of Defense for Intelligence (USD(I)), and assigned a Director for Defense Intelligence (Warfighter Support) as its executive agent. Over the nine years that followed, an initiative conceived to reduce the backlog of unreviewed Predator and Reaper footage from the counter-ISIS campaign matured into the Department’s flagship artificial intelligence program of record, the central nervous system of the Combined Joint All-Domain Command and Control (CJADC2) architecture, and the analytical substrate on which the United States fought Operation Epic Fury against Iran in early 2026.

The trajectory from a $70 million pathfinder to what the Office of the Secretary of Defense publicly describes as a platform generating on the order of one thousand targeting recommendations per hour is the principal institutional story of machine learning in the American military. It is also a story about how an exploratory project assembled outside the conventional defense contracting base eventually consolidated under a single commercial prime contractor, how an initiative framed in its early years as a decision-support aid for human analysts has been progressively integrated into the kill chain, and how the ethical perimeter around military artificial intelligence has been renegotiated at each stage of the program’s expansion.

On the Name

A maven is a recognized master of a domain, someone whose expertise is so deep, so internally coherent, and so consistently demonstrated that the surrounding community treats their judgment as both authoritative and generative. The word comes from the Yiddish meyvn, meaning “one who understands,” and that etymology captures the essence: a maven does not merely know things; a maven comprehends the structure beneath the things.

In practice, a maven exhibits several intertwined traits that distinguish them from a simple specialist. A maven accumulates knowledge not as trivia but as a living architecture, a system of interlocking principles that allows them to interpret new information instantly and correctly. A maven becomes a gravitational center in their field, orbited by others for guidance, calibration, and clarity. A maven’s reputation is built not only on mastery but on the ability to transmit that mastery — through explanation, synthesis, or the uncanny ability to diagnose what others have overlooked.

Historically, the term has been applied to figures who shaped the intellectual or technical terrain around them. In archival contexts, a maven would be the person whose annotations become the reference standard, whose classifications stabilize a field, whose interpretations endure because they are structurally correct rather than merely fashionable. In technical domains, a maven would be the one who sees the hidden failure modes, the latent opportunities, the underlying geometry of a system. In cultural or artistic domains, a maven would be the one who can trace lineage, influence, and semiotic weight with precision.

A maven is not a guru, not a dilettante, not a polymath for its own sake. A maven is a deep specialist with panoramic awareness, someone who can move from micro-detail to macro-structure without losing coherence. The title is earned, not claimed, and it persists because the maven’s work continues to illuminate the field long after the immediate moment has passed.

The choice of the word as the code name for the Department of Defense’s first major artificial intelligence program carried an implicit claim. The AWCFT was to be the system that understood — that comprehended the structure beneath the imagery, that exhibited the accumulated and architected mastery that human analysts did not have the time or bandwidth to produce. Whether a computer vision model trained on four million labeled images of military vehicles is plausibly described by the Yiddish word meyvn is a question the program’s founders left to their audiences to resolve for themselves. The subsequent evolution of the program into the Maven Smart System, with its weapon-pairing recommendations and its sub-minute end-to-end targeting cycles, has made the claim considerably more ambitious than it was in 2017.

Origins and Mandate

The intellectual genesis of Maven predates the April 2017 memorandum. Robert Work, who had served as Deputy Secretary under Secretary Ash Carter and continued briefly under Secretary Jim Mattis, had been the principal advocate within the Department of the so-called Third Offset Strategy, an attempt to identify areas of qualitative advantage analogous to the nuclear weapons of the First Offset and the stealth and precision-guided munitions of the Second Offset. By 2016, Work and his inner circle had concluded that the most probable locus of a Third Offset was human-machine collaboration and the military application of commercial advances in deep learning. Will Roper, then director of the Strategic Capabilities Office, had been pursuing a narrower effort to apply machine vision to drone footage. The AWCFT represented the broader, institutionalized version of that line of thinking, explicitly framed against what Work viewed as an emerging Chinese lead in military artificial intelligence.

The memorandum assigned the AWCFT three formal tasks: identifying and developing or modifying algorithms to accomplish specified intelligence functions, identifying the computational infrastructure required to field such algorithms, and integrating the resulting capability into existing operational systems. The initial concrete mission was object detection, classification, and tracking in full-motion video collected by ScanEagle, Predator, Reaper, Gray Eagle, and Global Hawk platforms during the counter-Islamic State campaign. Analysts in the Distributed Common Ground System were overwhelmed by the volume of collected FMV, much of which was never reviewed. The computer vision models procured under AWCFT were intended to pre-screen the footage and flag frames containing vehicles, personnel, structures, or other objects of interest, thereby allowing analysts to operate on a filtered feed rather than a raw one.

Lieutenant General Jack Shanahan served as director of the project from April 2017 until December 2018, when he was confirmed as the inaugural director of the Joint Artificial Intelligence Center (JAIC). Marine Corps Colonel Drew Cukor, who had chaired the AWCFT working group during the program’s formative period, provided much of the day-to-day direction and is credited by multiple accounts with articulating the vision that would eventually become the Maven Smart System. Early testing of vendor object-recognition tools against drone footage from Naval Special Warfare operations in Somalia produced disappointing results — models trained on civilian datasets performed poorly against the hazy, off-axis, poorly labeled imagery typical of operational FMV, and different vendors’ systems disagreed on basic taxonomy, classifying the same vehicle variously as a tank, a Soviet design, or a T-72. The early technical lesson was that usable military computer vision would require purpose-built training datasets. By the following year, AWCFT had labeled more than four million images of military objects for that purpose.

The Industrial Base

The AWCFT was deliberately architected to draw on commercial artificial intelligence capability rather than on the established defense contracting base. The Defense Innovation Unit (DIU), then called DIUx, assisted in identifying commercial vendors and in structuring the non-traditional contracting vehicles through which they could be brought into the program. Google, through an arrangement with Northrop Grumman, supplied TensorFlow-based computer vision models and engineering support. Palantir Technologies provided the data integration layer that would later become the foundation of the Maven Smart System. Amazon Web Services supplied commercial cloud infrastructure. ECS Federal served as the prime AI Integration contractor, a role it has retained continuously since 2017 and has recently branded as AI Interoperability Integrator, or AI3. Booz Allen Hamilton received a prime award worth $751.5 million in 2018, subcontracting elements of the work to a broader ecosystem. Maxar Technologies (formerly DigitalGlobe) supplied commercial satellite imagery and associated algorithms. L3Harris, Microsoft, Sierra Nevada Corporation, and more than twenty other firms contributed at various points. Anduril Industries joined the program in 2018, deploying its sensor-fusion platform and edge hardware; in December 2024 Anduril and Palantir announced a consortium intended to interconnect Anduril’s Lattice Mesh with the Maven Smart System and Palantir’s AI Platform.

This industrial arrangement was a departure from the pattern of the Cold War and post-Cold War defense sector, in which a small number of legacy primes supplied most major weapons and command systems through heavily negotiated cost-plus contracts. Maven’s industrial base was commercial, polyglot, and iterative. Capability delivery was measured in weeks or months rather than years, and the program adopted a continuous-delivery posture modeled on commercial software engineering. The Center for Security and Emerging Technology at Georgetown has documented that during its deployment in Europe the program iterated through sixty-two distinct capability evolutions in ten months.

The Google Rupture

The most visible crisis of the program’s first phase was the controversy over Google’s participation. In March 2018, Gizmodo reported that Google engineers had been assisting the Department of Defense on computer vision models for drone footage. Within weeks, an internal petition signed by more than three thousand Google employees — a figure that grew to roughly four thousand — demanded that the company withdraw and commit never to build warfare technology. Approximately a dozen engineers resigned in protest. A parallel open letter from academic artificial intelligence researchers called on Google and Alphabet to exit the contract and to pledge against developing autonomous weapons. In June 2018, Google announced that it would not renew its Maven contract when the existing work concluded in 2019, and in parallel published a set of AI principles that precluded work on weapons or on surveillance technologies violating internationally accepted norms. Google subsequently withdrew from the Joint Enterprise Defense Infrastructure (JEDI) cloud competition on related grounds.

The Google episode was treated at the time as a rupture in the relationship between Silicon Valley and the Department. It was also a natural experiment in whether employee pressure could force a commercial prime out of a major defense contract. It could. What has been less frequently noted is that the departure of Google from Maven had no measurable effect on the program’s trajectory. The work was reallocated to other contractors, the computer vision models continued to improve, and the program’s budget increased in every subsequent fiscal year. Shanahan would later describe the Google controversy as a failure of public affairs rather than a failure of policy — the argument being that the Department had failed to shape the narrative around what Maven was actually doing, allowing opponents to fill the informational vacuum. The ethical question that Google’s engineers had raised — whether computer vision optimized for drone footage would eventually be integrated into lethal targeting workflows — was not answered on its merits. It was, in practice, answered by subsequent events.

Institutional Drift

The AWCFT’s institutional history is a succession of reorganizations, each of which has complicated the question of who is actually responsible for the program. The establishment of the Joint Artificial Intelligence Center in June 2018 was the first consolidation, and Maven’s parent office was formally JAIC from that point forward, though day-to-day operations continued under the AWCFT label. In February 2022, JAIC was absorbed into the newly created Chief Digital and Artificial Intelligence Office (CDAO), alongside the Defense Digital Service and the Office of the Chief Data Officer. In April 2022, the Biden administration proposed a further restructuring under which portions of Maven would transfer to the National Geospatial-Intelligence Agency, portions would remain under CDAO, and overarching oversight would move to the Office of the Under Secretary of Defense for Intelligence and Security (OUSD(I&S)). The GEOINT-related lines of effort — reported by NGA to constitute roughly eighty percent of the original program — moved to NGA over the course of 2023, and NGA has since described Maven as its flagship GEOINT AI program and as a formal program of record within the agency.

The non-GEOINT lines of effort remained under CDAO, which simultaneously inherited responsibility for the so-called AI/ML Scaffolding intended to generalize Maven’s software engineering approach to the rest of the Department. Through 2025 this tripartite arrangement was largely opaque to outside observers. Officials from all three organizations were reticent about the boundaries of their respective responsibilities, citing operational security considerations, and budget detail was either compartmented into the intelligence community budget (for NGA) or realigned under new program elements (for CDAO) that obscured year-over-year comparisons. A former senior defense official quoted by DefenseScoop observed that the ambiguity was itself a governance problem, because no single official could be held accountable for the program as a whole.

On 9 March 2026, Deputy Secretary of Defense Steve Feinberg signed a memorandum that substantially collapsed this tripartite structure. The memorandum directed that the Maven Smart System transition to a formal program of record before the close of fiscal year 2026, that system administration and oversight responsibilities move from NGA to a newly constituted CDAO Maven Smart System Program Office within thirty days, that the Under Secretary of Defense for Research and Engineering assume authorizing official responsibilities for the associated commercial cloud enterprise infrastructure, and that all MSS contracts be transitioned to the existing Army Enterprise Agreement contract vehicle. In effect, procurement and contracting responsibilities would move to the United States Army, programmatic authority would consolidate at CDAO, and the GEOINT mission set that NGA had inherited in 2023 would be layered beneath a unified CDAO program office. The Army Combined Arms Command would integrate Maven into its training curriculum. Whether the NGA’s own Maven GEOINT lines of effort would be folded into the same program office was, as of April 2026, an open question that the Department had declined to resolve publicly.

Maven Smart System

The Maven Smart System, or MSS, is the user-facing application that has become the program’s principal deliverable. MSS evolved from the original AWCFT computer vision pipeline by progressive accretion. The earliest versions exposed a simple flagging interface in which yellow bounding boxes indicated objects that the algorithms had identified in FMV or satellite imagery, with blue markings for no-strike locations such as hospitals and schools. Successive iterations added multi-sensor fusion, so that electro-optical, infrared, synthetic-aperture radar, signals intelligence, geolocation metadata from communications intercepts, and commercially available automatic identification system (AIS) data for maritime traffic could be overlaid on a common map display. The system added target tracking, weapon-pairing recommendations through a module Palantir describes as the AI Asset Tasking Recommender, and tactical data link hooks through which a commander’s decision to engage a target could be transmitted machine-to-machine to a firing platform.

MSS currently ingests data from more than one hundred fifty sources according to Palantir’s public demonstrations. The Chief Digital and Artificial Intelligence Officer, Cameron Stanley, has publicly described the system as the consolidation point for what had previously been eight or nine separate targeting and operational workflows. As of March 2026 the system reportedly had more than twenty thousand active military users, a fourfold increase since March 2024. Palantir has stated that MSS is at production level across Indo-Pacific Command, European Command, Central Command, Northern Command and NORAD, Space Command, Transportation Command, Africa Command, and the Joint Staff, with additional deployments at Cyber Command, Strategic Command, and Southern Command. NATO acquired a variant, designated MSS NATO, under a six-month procurement action concluded in March 2025; the NATO Joint Warfare Centre had by August 2025 incorporated it into alliance exercises such as STEADFAST DETERRENCE and STEADFAST DUEL, describing it as the alliance’s first artificial-intelligence-enabled command-and-control system. The United Kingdom announced a parallel partnership with Palantir in September 2025 worth up to £750 million over five years.

The analytical layer of MSS, in at least some configurations, has been driven by large language models rather than purpose-built computer vision models alone. The NGA director, Vice Admiral Frank Whitworth, stated in late 2025 that by June 2026 Maven would begin transmitting intelligence products that were, in his phrasing, entirely machine-generated to combatant commanders, with Booz Allen Hamilton serving as the contractor for the LLM integration phase. Anthropic’s Claude model was integrated into MSS as an analytical and targeting-prioritization layer during 2025 and early 2026. The Claude integration would become a significant controversy in its own right during Operation Epic Fury.

Scarlet Dragon and the XVIII Airborne Corps

The operational proving ground for the transition of Maven from intelligence support to targeting was the series of Scarlet Dragon exercises conducted by the United States Army’s XVIII Airborne Corps beginning in the autumn of 2020. The first Scarlet Dragon iteration paired soldiers from the XVIII Airborne with Marines from II Marine Expeditionary Force at Fort Bragg (renamed Fort Liberty in 2023 and renamed Fort Bragg again in 2025). The exercise demonstrated an AI-flagged detection of a tank — variously described by participants as an inflatable target or a decommissioned hull — followed by human adjudication and an M142 HIMARS strike. The process consumed approximately 743 minutes, or slightly more than twelve hours, end to end. It was nonetheless described at the time as the first artificial-intelligence-enabled artillery strike in the history of the Army.

Successive Scarlet Dragon iterations compressed the targeting cycle dramatically. By 2024, the XVIII Airborne Corps was reporting end-to-end targeting cycles of less than one minute in exercise conditions and was claiming targeting throughput comparable to the approximately two-thousand-person targeting cell used during Operation Iraqi Freedom, but with a cell of roughly twenty personnel. The same senior targeting officer who quantified this efficiency estimated that with MSS an individual analyst could process approximately eighty potential targets per hour, against roughly thirty per hour under legacy workflows. The corps also developed a formal taxonomy of the kill chain into six stages — identify, locate, filter to lawful valid targets, prioritize, assign to firing units, and engage — and reported that MSS could automate or substantially accelerate four of the six, with the filtering and prioritization stages producing outputs that human analysts would adjudicate rather than originate.

The Scarlet Dragon process was overseen during its early iterations by the corps commander, Lieutenant General Michael Kurilla, who in April 2022 assumed command of United States Central Command. Kurilla carried the data-centric command approach and the MSS toolset with him to Tampa, and CENTCOM emerged as the most aggressive early adopter of MSS at the combatant command level. The CENTCOM chief technology officer, Schuyler Moore, stated in early 2024 that Maven had supported more than eighty-five precision airstrikes in Iraq, Syria, and Yemen and had been used to locate rocket launchers and surface vessels in the Red Sea region. The Houthi shipping campaign in the Red Sea during 2023 and 2024 provided the first sustained demonstration of MSS in a hot naval environment.

Operational Employment in Ukraine

The most significant operational laboratory for Maven outside of CENTCOM has been the war in Ukraine. The XVIII Airborne Corps, under Lieutenant General Christopher Donahue, took command of the Security Assistance Group-Ukraine in 2022 and progressively expanded its role from logistics and training into direct operational targeting support for Ukrainian forces. Maven computer vision models and an MSS variant adapted to the intelligence constraints of a partner-force engagement were reported to have supported Ukrainian targeting operations, including strikes on Russian command and logistics nodes. The NGA publicly disclosed that Maven had been used in Ukraine to compress find-fix-finish cycles to durations reported as under ten minutes. Independent analysts have attributed a range of specific Ukrainian targeting successes to Maven support, including the reported targeting of a senior Russian military visit to a forward headquarters in 2022 and Ukrainian long-range strikes against Russian refining and energy infrastructure in 2023 and 2024.

The New York Times reported in 2024 that Ukrainian forces were operating with a constrained version of Maven that did not integrate the most sensitive American intelligence inputs, and that the system’s performance had been mixed. Lessons from the Ukrainian deployment — including the fragility of overhead collection in contested electromagnetic environments and the imperative of distributed, resilient sensor architectures of the kind associated with commercial low-Earth-orbit constellations — have been reported to have reshaped Pentagon thinking on both satellite architecture and on the degree to which targeting pipelines should be designed to degrade gracefully under adversary action.

The Palantir Consolidation

The contractual history of Maven between 2020 and 2026 traces a progressive consolidation around Palantir as the prime systems integrator for the Maven Smart System, even as the broader Maven program ecosystem has retained a multi-vendor character for training data, algorithms, imagery, and edge hardware. Palantir received a $91.2 million contract running from September 2020 to November 2022 for its foundational work on MSS. In May 2024, the United States Army awarded Palantir USG a five-year indefinite-delivery, indefinite-quantity contract worth $480 million to build the MSS prototype, with completion scheduled by 28 May 2029. In September 2024, the Army added a $99.8 million task order to extend MSS access to the Army, Navy, Air Force, Space Force, and Marine Corps. In May 2025, the Department raised the MSS contract ceiling by $795 million to approximately $1.3 billion through 2029, citing adoption by combatant commands that had, according to industry analysts, exceeded expectations. The NGA simultaneously awarded a separate $28 million contract to expand MSS access for its analysts. In July 2025, the Army signed a $10 billion enterprise framework agreement consolidating roughly seventy-five existing Palantir contracts, a move that positioned the Army as the central procurement authority for subsequent Maven work. Industry reporting in early 2026 described the cumulative public-sector commitment to the platform as on the order of $13 billion.

The commercial logic of this consolidation is unsurprising. Palantir’s Foundry platform, the underlying data integration layer that became MSS, has substantial switching costs. Once a combatant command has ingested its data into Palantir’s ontology, trained its personnel on the interface, and connected its sensor feeds and targeting workflows to the system, substitution would be prohibitively disruptive. A former senior defense official quoted in April 2026 observed that if Palantir were to succeed in positioning MSS as what he called the platform of platforms for CJADC2, the effect would be to lock the Department into a decade-long dependency on a single commercial vendor for its central command-and-control infrastructure. This is a structural risk that has no direct analogue in the traditional defense industrial base, where platforms such as aircraft and ships are typically procured from multiple primes and where switching costs, while considerable, are at least measurable in units and dollars rather than in institutional knowledge embedded in proprietary software stacks.

Operation Epic Fury and the Claude Controversy

Operation Epic Fury, the joint United States-Israeli air and missile campaign against Iran that commenced on 28 February 2026, has been characterized by Department of War officials and by Palantir executives as the first major conflict in which an artificial intelligence targeting system served as the primary planning and execution substrate. Within the first twenty-four hours of the campaign, United States forces reportedly struck approximately one thousand discrete targets across Iran. Over the first three weeks, CENTCOM reported having struck between 5,500 and 6,000 targets. These figures have been attributed by both Admiral Brad Cooper, the CENTCOM commander, and by the CDAO Cameron Stanley to the operational employment of MSS, which according to public descriptions had collapsed what had previously been an eight-or-nine-system targeting workflow into a single interface.

The analytical layer of MSS during Operation Epic Fury included Claude, the large language model developed by Anthropic. According to reporting by the Washington Post and by Wired, Claude generated prioritized target recommendations, produced location coordinates, and supported post-strike battle-damage assessment. The integration was simultaneously the subject of a significant political dispute. On 27 February 2026 — one day before Operation Epic Fury commenced — Secretary of War Pete Hegseth designated Anthropic a supply-chain risk to national security and directed federal agencies to phase out use of Claude within six months. The designation followed Anthropic’s refusal to remove usage-policy language restricting fully autonomous weapons applications and mass domestic surveillance applications of its model. CENTCOM continued to employ Claude within MSS during the opening strikes of the campaign on the basis that the phase-out timeline had not yet run.

The contradiction at the heart of this episode is significant. A commercial AI provider that declined to enable fully autonomous lethal applications of its technology was nonetheless providing the analytical substrate for the largest AI-assisted targeting operation in the history of warfare, while being simultaneously barred from further federal business for declining to go further. On 28 February 2026, the opening hours of the campaign included a reported strike on the Shajareh Tayyebeh girls’ school in Minab, in Iran’s Hormozgan province, that killed at least 175 people, most of them children. Three members of the House of Representatives sent a letter to the Secretary of War asking what role MSS had played in the identification and validation of that target and whether the target had been verified by human review. A response was requested by 20 March 2026. Secretary Hegseth had, on 2 March 2026, publicly stated that Operation Epic Fury would be conducted without what he described as unnecessary rules of engagement. He had previously, in February 2025, dismissed the senior uniformed lawyers of the military services, whose responsibilities had included compliance oversight for the laws of armed conflict. An AI strategy document published by the Department of War on 9 January 2026 had framed commercial AI safety guardrails as ideological constraints inconsistent with warfighter needs, a framing under which the distinction between a guardrail against endorsing unlawful strikes and a guardrail against overconfident probability estimates collapsed into a single category of political obstruction.

The Autonomy Trajectory

It has been a consistent feature of official framing since 2017 that Maven is not an autonomous weapons system. Successive program directors — Shanahan, Cukor, and their successors — have emphasized that the system’s outputs are decision-support products subject to human adjudication, and that no existing AI component fires on a self-designated target. This framing is accurate at the level of the individual strike, in that no Maven-equipped munition fires without a human actuation. It is less straightforward at the level of the workflow. Bob Work, in interviews following his departure from government, distinguished between what he called autonomy at rest and autonomy on the move. Autonomy at rest is the automation of the analytical and targeting pipeline that delivers a named target with associated weapon pairing and collateral estimate to a human decision-maker who retains trigger authority. Autonomy on the move is an armed platform that selects and engages a target without human intervention. Maven has, over nine years, produced a substantially complete version of the first. It has not produced the second, but the algorithms, sensor architectures, and training data developed under the program have been reported in early 2026 to be undergoing adaptation for precisely that purpose under the Replicator initiative, which aims to produce large quantities of attritable one-way attack drones for a contested Western Pacific scenario.

The distinction between human-on-the-loop and human-in-the-loop becomes operationally meaningful when the tempo of the kill chain approaches the limits of human cognition. If MSS generates one thousand prioritized targets per hour, the time available for meaningful human review of each target is on the order of a few seconds. At that cadence the human role shifts from adjudication to rubber-stamping, and the question of where accountability lies for an erroneous or unlawful strike becomes difficult to answer in any operationally honest way. The program has been presented throughout its history as preserving meaningful human control. The arithmetic of tempo is not obviously consistent with that presentation.

Assessment

Nine years into its existence, the AWCFT has accomplished most of what its original sponsors set out to accomplish, and more. It has introduced machine learning into the core workflows of American military intelligence and targeting. It has produced a software platform that combatant commands find genuinely useful, as evidenced by the adoption curve and the consistent upward revision of contract ceilings. It has demonstrated, in Ukraine and in the Red Sea and in Iran, that the compression of the kill chain from days to minutes is a real and operationally significant capability. It has also reconfigured the relationship between the Department and the commercial technology sector in ways that were probably underappreciated at the program’s inception.

Four structural observations follow from the program’s history. First, the original premise — that commercial AI could be rapidly transferred into military workflows if the contracting and cultural obstacles were removed — has been substantially validated, but the political and ethical obstacles have proven more durable than the technical ones. The Google departure in 2018, the tech-worker open letters of 2024 and 2026, the Anthropic designation of February 2026, and the congressional inquiries following the Minab strike all point to a persistent political friction around commercial AI in lethal applications that program management has not resolved so much as periodically absorbed. Second, the institutional history of Maven — from OUSD(I) to JAIC to CDAO and NGA and back again, with procurement now consolidating at the Army — reflects the Department’s continuing difficulty in locating AI programs within a bureaucratic structure designed around platforms and commodities rather than around software. The Feinberg memorandum of March 2026 is the latest attempt at a stable settlement. It is unlikely to be the last. Third, the consolidation around Palantir as the prime systems integrator for MSS represents a form of vendor lock-in qualitatively different from anything in the traditional defense industrial base. The government has accepted this consolidation because it has produced operationally meaningful capability on timelines that the legacy contracting process cannot match. The long-term consequences of that acceptance remain unclear. Fourth, the framing of Maven as a decision-support system has eroded at each successive stage of the program’s expansion. It is not accurate, in 2026, to describe MSS as a system that merely flags objects for human analysts. It is a targeting system. The framing is significant because it bears on the legal and ethical constraints that apply to the program’s further development.

In the tradition of practitioner-adjacent analysis, it is useful to distinguish between the program’s instrumental successes and the questions its expansion has raised. Maven works. It has compressed targeting cycles, reduced manpower requirements, and enabled operations at scales and tempos that legacy systems could not support. These are real achievements. They are also achievements whose governance architecture has lagged the capability by, at this point, the better part of a decade. The Department has not produced an adequate public account of what it means for an AI system to be in the kill chain at the tempo MSS now operates, nor of how accountability would attach to an erroneous or unlawful strike in which the analytical recommendation was generated by a commercial large language model, weapon pairing was suggested by a software module, and the human decision-maker had, in practice, seconds rather than minutes to adjudicate. These are questions the program will, one way or another, be required to answer.

References