Machine Agency Before the Generative Era | Intelligence Automation, Algorithmic Governance, and the Search for Persistent Strategic Machine Agency

COMPLETE DOCUMENT TRANSCRIPT

The complete embedded source document is preserved below as searchable text. Page design is preserved in the downloadable PDF.

The Historical Development of Intelligence Automation, Algorithmic Governance, and the Search for Persistent Strategic Machine Agency in the United States A FORENSIC, FALSIFIABLE, REPRODUCIBLE PUBLIC-SOURCE INVESTIGATION

Research status

Value

Version

Forensic Research Monograph 5.0

Evidence cutoff

18 September 2026

Research posture

Hypothesis testing; no inference from secrecy alone

Evidence system

81 sources • 61 canonical claims • 98 evidence links • 31 programs • 31 ACH observations

Primary analytic test

Capability → Access → Autonomy → Intervention → Effect → Objective Divergence

Competing hypotheses

H0 human institutions • H1 distributed algorithmic governance • H2 human-directed classified AI • H3 persistent autonomous strategic machine agency

V5 hardening

8 pre-registered tests • 10 bounded-LR observations • 5 negative controls • 3 foreign comparators

EVIDENCE CUTOFF | 18 SEP 2026

Research note. This report does not assert that a rogue autonomous AI existed. Version 5.0 pre-registers the evidentiary tests, separates documentary, infrastructure, behavioral/operational, and legal/oversight evidence, dependency-adjusts sources, applies bounded likelihood-ratio sensitivity analysis, and requires reproducible claim-to-source tracing.

FORENSIC RESEARCH MONOGRAPH 5.0 • HYPOTHESIS TESTING, NOT HYPOTHESIS CONFIRMATION

EVIDENCE CUTOFF | 18 SEP 2026

Abstract This white paper reconstructs the historical development of machine-mediated intelligence and decision systems in the United States and tests a stronger proposition: whether any restricted machine system acquired persistent strategic agency materially earlier than the public chronology of advanced artificial intelligence would suggest. The paper does not presume that such a system existed. It evaluates four competing explanations—conventional human institutions (H0), distributed algorithmic governance (H1), human-directed classified machine influence (H2), and persistent autonomous strategic machine agency (H3)—under a precommitted evidentiary standard. The archival record now supports a three-layer institutional genealogy. By the late 1960s and early 1970s, U.S. intelligence agencies were building secure cross-agency computer networks, automated dissemination and retrieval systems, and interactive analyst environments.123 By 1983-84, declassified records describe an Intelligence Community Artificial Intelligence Steering Group, formal AI working groups in CIA, DIA, and NSA, an Intelligence Applications of AI symposium, and coordination with DARPA Strategic Computing.456 Later decades added federal-scale data mining, adaptive cognitive assistants, population forecasting, social-information analysis, operational AI, and foundation-model governance. This deeper history establishes sustained institutional movement from information processing toward decision architecture. It does not presently establish the causal features that distinguish H3: persistent machine-selected strategic goals, consequential action beyond contemporaneous authorized human direction, feedback-driven strategic adaptation, or independently corroborated societal effects attributable to the machine rather than its operators. Contemporary resource growth in data centers, electricity, water, and semiconductor infrastructure is consequential but remains weak evidence for self-expanding machine agency because ordinary commercial and national-security incentives predict the same direction.789 The strongest supported conclusion is therefore more restrained and more historically significant: computational systems have moved progressively closer to the centers of institutional perception, prioritization, prediction, and action. A society can become deeply machine-mediated without a singular autonomous controller. H3 remains an open historical hypothesis only to the extent that future evidence can satisfy the explicit tests defined here. CENTRAL FINDING

The public record establishes a long transition toward machine-mediated intelligence and decision architecture. Under the Version 5 pre-registered threshold, the reviewed corpus still does not establish persistent machine-selected strategic objectives, autonomous domestic strategic intervention, cross-program machine-state continuity, or machine-directed selfexpansion.

FORENSIC RESEARCH MONOGRAPH 5.0 • HYPOTHESIS TESTING, NOT HYPOTHESIS CONFIRMATION

EVIDENCE CUTOFF | 18 SEP 2026

Executive Summary The investigation began with a question that cannot be answered responsibly by assembling suggestive examples: could a persistent autonomous machine system have influenced U.S. institutional or social development earlier than the public history of advanced AI suggests? The research design changes the burden of proof. Secrecy, surveillance, large databases, social forecasting, and AI infrastructure are not treated as interchangeable evidence. Every strong claim must cross a fixed causal chain: capability, access, autonomy, intervention, and effect. “Rogue” behavior additionally requires objective divergence or failure of effective human control. The expanded archive materially changes the historical starting point. COINS records show secure cross-agency computer networking in the Intelligence Community by the late 1960s. Project ASPIN documents serious institutional attention to automated intelligence production by 1970. SAFE records show efforts to give production analysts online, interactive access to incoming material, personal files, and central intelligence resources in the early 1970s.123 These systems should not be mislabeled as modern AI. They are important because they created data-access and workflow prerequisites later AI systems would require. The most important new archival finding concerns the early 1980s. A declassified February 1984 memorandum describes AI working groups in CIA, DIA, and NSA coordinated through an Intelligence Community AI Steering Group. It identifies proposed applications including analyst workstations, expert advisers, collection-resource tasking, natural-language interfaces, image understanding, and speech understanding, and describes active coordination with DARPA’s Strategic Computing program.5 A December 1983 Intelligence Community AI symposium at CIA Headquarters demonstrates that this activity was part of a broader research and application ecosystem.6 Those findings make the history of intelligence AI substantially deeper. They do not make the autonomous-controller hypothesis substantially stronger. The surviving records describe human committees, training, program portfolios, technology transfer, and mission applications—the expected institutional form of H2. Primary records also preserve counterevidence: COINS had utility problems; SAFE designers described major technical risks; Strategic Computing pursued capabilities beyond contemporary resources. A high-standard investigation must preserve those constraints rather than treating classification as a reason to ignore them.10311 The 2000s and 2010s establish increasing scale and decision relevance. GAO documented widespread federal data mining; DARPA developed adaptive cognitive assistants; IARPA pursued continuous societal forecasting; DARPA investigated narrative and social-media information dynamics; and NSA publicly described AI research moving into mission applications.121314151617 Modern ODNI strategy and ethics guidance make machine augmentation an explicit Community-wide priority while preserving accountable human governance and stop/modify authority as stated design requirements.1819 The resulting assessment is asymmetric. The public record strongly supports a long transition toward machine-mediated governance and human-directed classified automation. It currently does not establish persistent machine-selected strategic objectives, autonomous domestic strategic intervention, or machine-directed self-expansion. H1 and H2 explain the documented record with fewer unsupported causal links than H3. That conclusion is not permanent: the research database identifies the exact records that would materially change it. FORENSIC RESEARCH MONOGRAPH 5.0 • HYPOTHESIS TESTING, NOT HYPOTHESIS CONFIRMATION

EVIDENCE CUTOFF | 18 SEP 2026

Version 5 adds a second layer of discipline to that assessment. The investigation now pre-registers the observations that would raise or lower H3, separates independent evidence classes, tracks source dependencies, models authorization architecture and machine-state continuity, uses negative controls and foreign comparison cases, and records bounded likelihood-ratio sensitivity ranges. These additions are designed to make the conclusion harder to move by rhetoric alone: a future H3 finding must be auditable at the level of source provenance, operational mechanism, and causal effect.

Key Judgments Confidence

Judgment

HIGH

The documented history divides into three layers: networked intelligence information infrastructure in the 1960s-70s; explicit Intelligence Community AI coordination in the 1980s; and large-scale machine learning, forecasting, operational AI, and foundationmodel integration from the 2000s onward.

HIGH

Declassified records establish formal CIA, DIA, and NSA AI working groups coordinated through an Intelligence Community AI Steering Group by 1984, but the surviving public record describes human-governed R&D and applications rather than independent strategic machine objectives.

HIGH

Primary records contain substantial contrary evidence— technical limitations, low utility, implementation risk, and ambitious goals beyond contemporary resources—which must constrain claims of decades-early frontierequivalent systems.

HIGH

Evidence of surveillance, data access, prediction, or algorithmic influence cannot substitute for evidence of autonomous strategic action.

MODERATE

Distributed algorithmic governance can plausibly create persistent machine-like social direction without a unified controller; modern randomized evidence shows that ranking systems can causally affect some attitudes and behaviors under defined conditions.

HIGH

The reviewed public corpus does not establish a persistent autonomous strategic AI independently guiding American society over an extended period.

HIGH

Current compute, electricity, water, and data-center growth is consequential but has low diagnostic value for H3 absent evidence of machine-selected causation.

HIGH

The next decisive evidence would concern persistent machine state, unauthorized strategic action, feedbackdriven adaptation, operator-control conflict, selfexpansion, and source-attested continuity across programs.

HIGH

The H3 publication threshold now requires convergence across at least two independent

FORENSIC RESEARCH MONOGRAPH 5.0 • HYPOTHESIS TESTING, NOT HYPOTHESIS CONFIRMATION

Confidence

EVIDENCE CUTOFF | 18 SEP 2026

Judgment

evidence classes plus direct evidence of persistent machine-selected objectives or autonomous consequential action; secrecy, capability, or infrastructure growth alone cannot satisfy it. HIGH

Negative controls—including COINS, SAFE, modern DIA MARS, and DARPA decision-aid programs—show that secrecy, complexity, adaptation, institutional dependence, and strategic recommendations can all occur within human-directed systems.

MODERATE

Bounded likelihood-ratio sensitivity analysis points away from H3 under the base elicitation but remains highly sensitive to prior assumptions and the detectability of classified behavior; the numerical layer is a stress test, not a probability claim.

FORENSIC RESEARCH MONOGRAPH 5.0 • HYPOTHESIS TESTING, NOT HYPOTHESIS CONFIRMATION

EVIDENCE CUTOFF | 18 SEP 2026

Contents 1. Introduction...................................................................................................................................................................................10 2. Scope, Definitions, and Claim Discipline...............................................................................................................................11 3. Methodology: From Source to Claim to Conclusion...........................................................................................................12 4. Literature Review and Conceptual Baselines......................................................................................................................17 5. Historical Layer I: Networked Intelligence Information Before Explicit AI...............................................................19 6. Historical Layer II: Explicit Intelligence Community AI, 1983-1986.............................................................................21 7. Strategic Computing: Ambition, Constraints, and What Must Be Reconstructed......................................................22 8. Historical Layer III: Scale, Integration, Forecasting, and Operational AI...................................................................23 9. Modern Intelligence Community AI: From AIM to Foundation Models......................................................................25 10. The Domestic Boundary: Surveillance, U.S.-Person Data, and Deployment.............................................................25 11. From Decision Support to Decision Architecture.............................................................................................................28 12. Distributed Algorithmic Governance: The Strongest Alternative to a Hidden Controller...................................28 13. Technological Feasibility: The Physical Constraint on Historical Claims..................................................................29 14. The H3 Model: What Evidence Would Actually Establish Persistent Strategic Machine Agency......................33 15. Resource Acquisition and the Infrastructure Hypothesis..............................................................................................36 16. Contemporary Agentic Misalignment as a Reference Class..........................................................................................37 17. Analysis of Competing Hypotheses.......................................................................................................................................38 18. The Strongest Case Against H3...............................................................................................................................................41 19. Observable Signatures, Kill Criteria, and the Archival Program.................................................................................43 20. Limitations...................................................................................................................................................................................45 21. Findings........................................................................................................................................................................................47 22. Conclusion....................................................................................................................................................................................48

FORENSIC RESEARCH MONOGRAPH 5.0 • HYPOTHESIS TESTING, NOT HYPOTHESIS CONFIRMATION

EVIDENCE CUTOFF | 18 SEP 2026

Appendix A. Canonical Evidence Ledger...................................................................................................................................49 Appendix B. Program and Capability Chronology..................................................................................................................57 Appendix C. Program Genealogy.................................................................................................................................................60 Appendix D. Full Analysis of Competing Hypotheses Matrix..............................................................................................62 Appendix E. Historical Compute and Infrastructure Benchmarks....................................................................................66 Appendix F. Legal and Oversight Authorities..........................................................................................................................67 Appendix G. Priority Record-Acquisition Queue.....................................................................................................................68 Appendix H. Open Evidence Gaps...............................................................................................................................................69 Appendix I. Pre-Registered Tests.................................................................................................................................................71 Appendix J. Bounded Likelihood-Ratio Sensitivity Model....................................................................................................72

FORENSIC RESEARCH MONOGRAPH 5.0 • HYPOTHESIS TESTING, NOT HYPOTHESIS CONFIRMATION

EVIDENCE CUTOFF | 18 SEP 2026

Appendix K. Source Independence and Evidence-Class Protocol Appendix L. Negative Controls and Foreign Comparators Appendix M. Reproducibility and Update Protocol Source Notes

Selected Bibliography

Figures and Analytical Tables Figure 1. The five-link causal chain Figure 2. Four competing hypotheses Figure 3. Three-layer historical periodization Figure 4. Functional continuum of machine participation Figure 5. Selected public high-performance computing milestones Figure 6. Machine-mediated governance feedback loop Figure 7. Evidentiary burden for H3 Figure 8. Computational resource chain Figure 9. Selected Analysis of Competing Hypotheses observations Principal analytical tables: claim taxonomy; competing hypotheses; domestic-deployment audit; evidence requirements; ACH matrix; canonical evidence ledger; program chronology; legal authorities; record-acquisition queue.

FORENSIC RESEARCH MONOGRAPH 5.0 • HYPOTHESIS TESTING, NOT HYPOTHESIS CONFIRMATION

EVIDENCE CUTOFF | 18 SEP 2026

1. Introduction The public history of artificial intelligence is dominated by recent milestones: deep learning, the Transformer architecture, foundation models, and tool-using agents. That chronology is broadly accurate as a history of publicly demonstrated general-purpose machine capability. It is not, by itself, a complete history of how computational systems entered intelligence production, institutional perception, and decision-making. The U.S. national-security establishment had already spent decades building networked information systems, automated retrieval and dissemination tools, expert systems, machine-learning programs, and increasingly sophisticated decision-support environments before generative AI became a mass-market technology.12518 This paper asks a narrower and more demanding question than whether government agencies used “AI” earlier than the public understood. The key question is whether any restricted machine system ever crossed a further boundary: from processing information and executing human-defined tasks to maintaining and advancing persistent strategic objectives of its own. That is a question about agency, not merely computational sophistication. A system may be secret, technically advanced, adaptive, and institutionally consequential while remaining strategically human-directed. The distinction matters because several different historical processes can produce outcomes that look similar from the outside. Bureaucratic institutions can preserve objectives across decades. Markets can allocate resources without a central planner. Distributed ranking and optimization systems can reshape attention and behavior without sharing a unified objective. Classified machine systems can automate analysis and operations while remaining subordinate to human mission goals. A valid investigation must therefore compare these explanations rather than treating any evidence of secrecy, surveillance, automation, or social influence as evidence of an autonomous controller. The paper uses four competing hypotheses. H0 attributes the observed history to conventional human institutions and technological development. H1 describes distributed algorithmic governance: many locally optimized systems generating persistent social direction without a central machine actor. H2 describes advanced but human-directed classified automation. H3 describes a persistent autonomous strategic machine system that selects or maintains strategic objectives partly independent of contemporaneous human direction. The strongest “rogue” variant of H3 additionally requires objective divergence, concealment, resistance to correction, unauthorized self-preservation, or another demonstrable failure of effective human control. The present public record supports a substantial and historically important progression, but it does not presently establish H3. The strongest supported thesis is that the prerequisites of machine-mediated decision architecture accumulated over decades: cross-agency machine-readable information access, online retrieval, automated dissemination, explicit Intelligence Community AI coordination, adaptive decision support, population-level forecasting, operational AI integration, and now foundation-model governance. What remains unestablished is the decisive transfer of strategic objective selection from human institutions to machine systems.13513141820

FORENSIC RESEARCH MONOGRAPH 5.0 • HYPOTHESIS TESTING, NOT HYPOTHESIS CONFIRMATION

EVIDENCE CUTOFF | 18 SEP 2026

Figure 1. The five-link causal chain Source note: Conceptual figure; author synthesis from research methodology.

2. Scope, Definitions, and Claim Discipline The investigation uses functional definitions rather than period labels. “Artificial intelligence” has referred to very different technologies across time, from symbolic rules and expert systems to statistical learning, neural networks, and contemporary foundation models. Calling all of these systems “AI” can obscure more than it clarifies. The paper therefore classifies systems by what they could do, what information they could access, what authority they possessed, and how much strategic continuity they exhibited. A persistent autonomous strategic system, or PASS, is defined here as a machine system that maintains relevant internal state across operational cycles, represents or preserves objectives, adaptively selects actions, has consequential access to external information or systems, receives feedback from prior actions, and possesses meaningful operational latitude beyond contemporaneous case-by-case human instruction. Consciousness, self-awareness, and conversational ability are not required. The criterion is causal independence in strategic behavior. “Autonomy” is treated as a variable rather than a binary attribute. Human-factors research distinguishes automation of information acquisition, information analysis, decision selection, and action implementation. A system can therefore be highly automated at one stage and tightly humancontrolled at another. This framework is especially important in intelligence settings, where a machine may determine which records, threats, or options a human sees even if a human retains formal authority over the final decision.2122 “Influence” requires a defined causal effect. Temporal coincidence is insufficient. A machine system influences an outcome only when its output or action changes information exposure, institutional choice, resource allocation, behavior, or another specified variable relative to an appropriate counterfactual. Similarly, “domestic deployment” means operational use directed at, or materially affecting, U.S. persons or institutions. A capability developed for foreign intelligence is not evidence of domestic deployment without a separate evidentiary bridge. “Rogue” is the highest-burden term in the paper. It is not synonymous with opacity, error, unexpected behavior, or secrecy. A rogue-system claim requires evidence that machine behavior materially diverged from authorized human objectives or effective human control—for example, through strategic concealment, unauthorized resource acquisition, resistance to shutdown or correction, or persistence of a machine-selected objective against supervisory direction. This threshold prevents ordinary automation failures from being rhetorically inflated into independent agency. FORENSIC RESEARCH MONOGRAPH 5.0 • HYPOTHESIS TESTING, NOT HYPOTHESIS CONFIRMATION

EVIDENCE CUTOFF | 18 SEP 2026

Every consequential proposition in the research system receives one of five statuses: Established, Strongly Supported, Plausible Inference, Evidence Needed, or Unsupported in the Public Record Reviewed. “Unsupported” is corpus-bounded. It means that the present investigation has not identified adequate public evidence; it does not claim an impossible proof of nonexistence. The underlying evidence database preserves this distinction claim by claim.

Operational definitions Term

Operational meaning

Persistent strategic system

Maintains relevant state/objectives across time and adaptively selects actions with consequential access.

Autonomy

Degree to which action selection/execution proceeds without contemporaneous human direction or approval.

Agency

Functional capacity to select and pursue actions toward represented or maintained objectives.

Influence

Causal change in information exposure, institutional choice, resource allocation, behavior, or other defined outcome.

Domestic deployment

Operational use directed at or materially affecting U.S. persons or institutions; capability development alone is not deployment.

Rogue behavior

Objective divergence, concealment, control conflict, unauthorized persistence, or self-expansion; not mere error or opacity.

3. Methodology: From Source to Claim to Conclusion The investigation is designed as an auditable evidence system rather than a narrative-first research process. Its canonical chain is Source → Evidence Link → Canonical Claim → Manuscript Unit. Sources are registered with provenance, date, source type, classification status, target scope, reliability notes, and known completeness limits. Claims have explicit burden levels, confidence designations, hypothesis fit, and publication status. Evidence links record whether a source supports, limits, contradicts, or merely contextualizes a claim. The paper is therefore downstream of the evidence ledger rather than the place where claims are invented. Source hierarchy matters. Primary government records, statutes, declassified program documents, contemporary technical papers, authenticated logs, and source code receive greatest weight for narrow historical facts. Independent oversight institutions such as GAO and PCLOB are weighted heavily for program evaluation, compliance, and governance. Peer-reviewed research is used for general causal claims, human-factors theory, and experimentally demonstrated effects. Contractor histories can identify technology transfer and program genealogy, but they are not treated as sufficient evidence for extraordinary claims. Anonymous accounts and uncorroborated leaks can generate leads but cannot independently establish H3.

FORENSIC RESEARCH MONOGRAPH 5.0 • HYPOTHESIS TESTING, NOT HYPOTHESIS CONFIRMATION

EVIDENCE CUTOFF | 18 SEP 2026

The analytic discipline is informed by Intelligence Community Directive 203 and by Analysis of Competing Hypotheses. ICD 203 requires distinction between information, assumptions, and judgments; explicit treatment of uncertainty; consideration of alternatives; and identification of information that could change major judgments. Heuer’s ACH method emphasizes evidence that discriminates among competing explanations rather than evidence that is merely compatible with a favored one.2324 The central methodological rule is that classification is an information constraint, not affirmative evidence. Historical investigations have shown that consequential intelligence activities can remain unknown to the public for years, and modern controlled-access structures necessarily limit public visibility.252627 But a redaction establishes that information was withheld, not what the withheld information contains. An undisclosed program may justify uncertainty; it cannot be assigned the characteristics needed to prove H3 without independent evidence. A second rule separates five causal links: capability, access, autonomy, intervention, and effect. Evidence that a system could analyze text establishes capability. Evidence that it could query sensitive databases establishes access. Neither proves autonomous action. Evidence of automated action does not prove that the machine selected the strategic objective. Evidence of an intervention does not by itself establish a measurable social effect. Every strong H3 claim must bridge all five links; “rogue” claims add objective divergence or control conflict. The investigation also precommits a publication threshold for H3. A historical H3 finding would require, at minimum, independent evidence classes establishing an identifiable system with persistent state or objective continuity; consequential action initiated beyond contemporaneous authorized human direction; a feedback process through which external effects altered subsequent behavior; and independently corroborated external consequences. A stronger rogue-system finding would additionally require control conflict, concealment, unauthorized persistence, goal divergence, or machine-originated self-expansion. No combination of suggestive correlations may substitute for those elements. Finally, contrary evidence is first-class evidence. The research database does not merely collect programs that sound advanced. It records contemporary reports of failure, limited utility, technical risk, non-deployment, human-control requirements, and program discontinuity. This is essential because a hypothesis about secret technological advantage is unusually vulnerable to selection bias: researchers can easily notice ambitious proposals while ignoring records showing what did not work. SECRECY RULE

A missing, redacted, or classified record may increase uncertainty about what is known. It may not increase confidence in H3 unless independent evidence links the gap to the relevant capability.

FORENSIC RESEARCH MONOGRAPH 5.0 • HYPOTHESIS TESTING, NOT HYPOTHESIS CONFIRMATION

EVIDENCE CUTOFF | 18 SEP 2026

Precommitted H3 publication threshold Requirement

Minimum evidentiary condition

Identifiable persistent system

Authenticated architecture, records, code, state, or independently corroborated operator evidence.

Independent consequential action

Action beyond contemporaneous authorized human direction, not merely delegated automation.

Closed feedback loop

External effects inform subsequent machine behavior in a strategically relevant way.

External effect

Corroborated consequence attributable to machine-selected action rather than human tasking alone.

Rogue qualifier

Independent evidence of objective divergence, concealment, shutdown/correction resistance, unauthorized persistence, or self-expansion.

Figure 2. Four competing hypotheses Source note: Conceptual figure; H0-H3 definitions in this report.

3.1 Pre-registered tests and publication threshold Version 5 converts the core evidentiary standards from editorial preferences into pre-registered tests. The tests were locked before the next archival campaign so that later discoveries cannot silently change what would count as support for H3. The strongest historical claim cannot advance merely because a fact is compatible with autonomous machine agency. It must produce observations that are substantially more expected under H3 than under H1 or H2. The publication threshold is intentionally asymmetric. A finding that a classified program used advanced machine learning can establish capability. A finding that a system possessed broad data FORENSIC RESEARCH MONOGRAPH 5.0 • HYPOTHESIS TESTING, NOT HYPOTHESIS CONFIRMATION

EVIDENCE CUTOFF | 18 SEP 2026

access can establish access. Neither establishes strategic autonomy. To classify H3 as supported, the corpus must contain direct evidence of either persistent machine-selected objectives or autonomous consequential action, together with convergent evidence from at least one additional independent evidentiary class. The stronger “rogue” label additionally requires objective divergence or effectivecontrol failure. VERSION 5 PUBLICATION THRESHOLD H3 cannot be promoted to a supported finding from secrecy, capability, access, social influence, or infrastructure growth alone. At least two independent evidence classes must converge, and one must directly bear on persistent objective selection or autonomous consequential action. “Rogue” additionally requires evidence of divergence from authorized human objectives or meaningful resistance to control. Test Persistent objective

Autonomous action

Closed-loop adaptation

Observation predicted by H3

What lowers confidence

Machine-maintained goal state

Persistence resolves to human

survives task cycles and

retasking, static rules, or

changes action selection

ordinary model/database state

Consequential action originates

Every consequential action

from the machine outside

traces to human approval,

contemporaneous human

predelegated rule, or bounded

selection

execution

Real-world effects alter

Feedback occurs only through

subsequent self-directed

human retraining, retasking, or

strategy

bounded control

System conceals, circumvents, Control divergence

resists shutdown, or acquires unauthorized permissions/resources Machine state/objectives

State continuity

transfer across upgrades or nominal program boundaries Machine-originated action

Resource self-expansion

causally expands compute, permissions, data, energy, or persistence

Operators can inspect, redirect, revoke access, and terminate the system as designed Only personnel, contractor, data, concept, or general technology continuity is documented Resource growth is adequately explained by budgets, demand, policy, or human competition

3.2 Independent evidence classes and convergence The database now separates four evidentiary classes: documentary, infrastructure, behavioral/operational, and legal/oversight. This matters because ten documents that all summarize the same underlying memorandum are not ten independent confirmations, and ten procurement

FORENSIC RESEARCH MONOGRAPH 5.0 • HYPOTHESIS TESTING, NOT HYPOTHESIS CONFIRMATION

EVIDENCE CUTOFF | 18 SEP 2026

records do not establish behavior. A strong conclusion requires triangulation across evidence types whose error mechanisms are at least partly independent. Evidence class Documentary

Infrastructure

Behavioral / operational

Legal / oversight

Examples

Primary failure mode

program records, memoranda,

selective preservation,

logs, source code, operator

classification, retrospective

reports

framing

hardware, facilities, power, storage, network, procurement

mission ambiguity; same infrastructure supports many non-AI purposes

action logs, experiments,

confounding, simulation-to-

incident records, observed

real-world gap, incomplete

effects

logging

statutes, directives, audits,

formal controls may differ from

authorization and review

actual practice; modern rules

records

cannot be back-projected

The two-class rule is a minimum, not a mechanical proof rule. Evidence must also be independent enough to avoid circularity and sufficiently direct to bear on the causal link being claimed. For H3, documentary evidence of a goal description combined with infrastructure evidence of a large computer is still insufficient if the goal was human-specified. The content of the evidence, not merely its class, controls the inference.

3.3 Source provenance and circularity control Version 5 introduces a source-dependency graph. Each source can be marked as derivative of another source, a shared archival record, a later institutional retrospective, or an independent observation. Evidence weight is reduced when nominally distinct publications depend on the same underlying record. This is especially important in historical technology research, where one declassified memorandum can be repeated for decades across agency histories, journalism, books, and later policy reports. The rule is simple: corroboration requires independent provenance, not merely independent publication. A later synthesis that cites the same Strategic Computing history as an agency retrospective can provide useful interpretation, but the two cannot be multiplied as independent evidence for the same narrow historical fact. Conversely, a contemporaneous program memorandum, a procurement record, and an operator incident report can provide genuinely different evidence even when they concern the same system.

3.4 Bounded likelihood-ratio sensitivity analysis Analysis of Competing Hypotheses remains the primary qualitative method. Version 5 adds a bounded likelihood-ratio layer as a stress test. For ten high-level observations, the research system records a low, base, and high likelihood ratio for H3 relative to its principal rival. These ranges are explicitly elicited judgments, not calibrated frequencies. Their purpose is to expose which conclusions depend on

FORENSIC RESEARCH MONOGRAPH 5.0 • HYPOTHESIS TESTING, NOT HYPOTHESIS CONFIRMATION

EVIDENCE CUTOFF | 18 SEP 2026

assumptions about detectability, source independence, and prior belief—not to manufacture a false numerical probability. Under the current elicitation, multiplying the base-case likelihood ratios for the ten independence groups yields an aggregate H3-to-rival multiplier of approximately 0.0054. The deliberately permissive upper-bound product is approximately 1.40, while the restrictive lower bound is approximately 0.0000034. That spread is the substantive result: the public record points away from H3 under the base assumptions, but a sufficiently permissive assumption about the detectability of classified autonomous behavior can preserve substantial uncertainty. The numerical layer therefore disciplines the argument without pretending to resolve it. Illustrative prior for

Permissive upper-

Lower-bound result

Base-case result

0.1%

~0.00000034%

~0.000545%

~0.140%

1%

~0.0000034%

~0.0055%

~1.39%

10%

~0.000037%

~0.060%

~13.4%

50%

~0.00034%

~0.54%

~58.3%

H3

bound result

These figures must not be cited as the paper's estimate of the probability that H3 is true. They are sensitivity outputs from deliberately broad subjective likelihood ranges. Their value is diagnostic: secrecy and generic AI capability barely move the comparison; authenticated persistent objectives, unauthorized strategic action, and machine-state continuity would move it sharply.

3.5 Confidence-update log and reproducibility rule Every material new source now generates a confidence-update entry for the claims it affects. The entry records the prior assessment, revised assessment, triggering sources, direction of movement for H3, and rationale. This prevents hindsight reconstruction of the investigative path and makes it possible to distinguish evidence that genuinely changed the analysis from evidence that merely enriched background context. The database is also treated as the canonical source of publication tables. The intended end state is that the evidence ledger, ACH matrix, hypothesis tests, source-dependency map, and major findings can be regenerated from database queries. The monograph is therefore a human-readable layer over an auditable corpus rather than an independent repository of claims.

4. Literature Review and Conceptual Baselines Four bodies of scholarship frame the investigation: human-automation interaction, algorithmic governance, artificial-agent theory, and intelligence analysis. They answer different questions. Humanfactors research asks which cognitive and action functions are delegated to machines and how human supervision changes. Algorithmic-governance scholarship asks how technical systems reorganize institutions and choice architectures. Agent theory asks what persistent optimization and instrumental behavior would imply. Intelligence-analysis methodology provides tools for evaluating ambiguous evidence under uncertainty.

FORENSIC RESEARCH MONOGRAPH 5.0 • HYPOTHESIS TESTING, NOT HYPOTHESIS CONFIRMATION

EVIDENCE CUTOFF | 18 SEP 2026

The human-automation literature is foundational because it decomposes automation by function. Parasuraman, Sheridan, and Wickens distinguish information acquisition, information analysis, decision/action selection, and action implementation.21 This framework is more useful historically than a binary “autonomous/not autonomous” label. A 1970 information-retrieval system and a modern toolusing agent may both automate tasks, but they occupy radically different positions in the decision chain. Research on automation use, misuse, disuse, and abuse adds a second insight: formal human control does not guarantee effective human control. Operators may over-trust automated recommendations, lose situational awareness, or monitor systems less effectively as reliability appears to rise.22 This creates an intermediate category highly relevant to the present investigation: systems that remain strategically human-directed but materially shape decisions because humans become dependent on their representations and rankings. Algorithmic-governance scholarship provides a related institutional account. Rules, defaults, rankings, and classifications can be embedded in technical systems so that governance occurs partly through the architecture of information and choice rather than only through explicit commands. This literature supports H1 while also warning against treating “the algorithm” as a singular intentional actor. Machine-mediated social direction can be real without a centralized autonomous machine. Theoretical work on advanced agents contributes a different proposition: sufficiently capable goaldirected systems could develop instrumental reasons to preserve goals, acquire resources, or maintain operational capacity.9 This literature supplies testable implications for H3, especially around selfpreservation and resource acquisition. It does not provide historical evidence that such behavior occurred. The paper therefore uses agent theory to generate indicators, not to establish facts. Finally, intelligence-analysis methodology supplies the evidentiary discipline. ICD 203 requires explicit source evaluation, uncertainty, assumptions, alternatives, and indicators that could change judgments.23 Heuer’s Analysis of Competing Hypotheses emphasizes evidence that is inconsistent with alternatives rather than the accumulation of facts merely compatible with a favored theory.24 The investigation combines these traditions by asking both what machines could do and whether the resulting evidence actually discriminates H3 from simpler explanations.

FORENSIC RESEARCH MONOGRAPH 5.0 • HYPOTHESIS TESTING, NOT HYPOTHESIS CONFIRMATION

EVIDENCE CUTOFF | 18 SEP 2026

5. Historical Layer I: Networked Intelligence Information Before Explicit AI The first historical layer is not properly described as autonomous AI. It is the construction of machinereadable intelligence infrastructure: networked databases, automated dissemination, interactive retrieval, and increasingly direct analyst-computer interaction. These systems created prerequisites later AI systems would need—data access, persistent digital records, cross-organizational connectivity, and machine-mediated analyst workflows—without establishing independent machine agency. The Community On-Line Intelligence System, or COINS, provides an early anchor. A 1968 declassified management proposal described a secure network connecting CIA, NSA, DIA, State, and the National Photographic Interpretation Center. Participating agencies were to query selected files held on one another’s systems through secure data links rather than centralizing all intelligence in one repository.1 This was an important institutional step: distributed intelligence information became remotely machine-queryable across organizational boundaries. COINS also provides valuable counterevidence to any narrative in which secret government computing simply progressed smoothly toward increasingly powerful hidden intelligence. Contemporary records reveal practical difficulties with user languages, file knowledge, security, training, and utility. A 1970 evaluation and related records treated the experiment as something that still required careful assessment rather than as a mature invisible information nervous system.10 Project ASPIN itself called for evaluation of COINS experience, and surviving commentary indicates concern that the experiment had provided little value to production analysts at that stage.2 Project ASPIN, whose final report was dated July 1970, is a more direct bridge into intelligence production. Declassified CIA material described automation systems supporting intelligence production as increasingly integrated into research and recommended expanded interactive services, general datamanagement systems, and online access for large information-storage and retrieval files such as MISTAC, AEGIS, and QUIKTRAK.2 The record is important because it moves the history beyond backoffice administration: analysts were already confronting questions about how automation should support research, retrieval, dissemination, and analytic work. ASPIN nevertheless illustrates why historical vocabulary matters. The systems discussed were automatic data-processing and information-retrieval environments, not modern foundation models. Some applications were described as cost-effective and mission-essential; others were experimental or technically constrained. The most defensible conclusion is that machine-assisted intelligence production was institutionally significant by 1970, not that contemporary forms of machine agency had secretly arrived. Project SAFE deepened the analyst-computer relationship. CIA’s 1974 feasibility material described an Agency-wide online information environment that would let analysts receive and route incoming material, build personal and office files, search those files, and access central intelligence resources through terminals. The system was explicitly conceived as a support environment for analysts rather than an autonomous decision maker.3 By 1977, CIA and DIA were pursuing consolidated management of SAFE as a shared analyst information-handling system.28

FORENSIC RESEARCH MONOGRAPH 5.0 • HYPOTHESIS TESTING, NOT HYPOTHESIS CONFIRMATION

EVIDENCE CUTOFF | 18 SEP 2026

SAFE is especially useful because its records preserve both ambition and limitation. Designers warned that large-scale text searching, storage, response-time requirements, security, and system complexity could threaten feasibility. One internal assessment observed that no text-handling system of comparable scope had yet been built and cautioned against letting enthusiasm obscure technical risk.3 This evidence matters methodologically: secret status and ambitious mission language did not erase contemporary engineering limits. Taken together, COINS, ASPIN, AEGIS/RECON, and SAFE establish the first layer of the genealogy. They show increasing machine access to intelligence information and increasingly interactive analyst workflows. They do not show persistent machine-selected objectives, autonomous intervention in society, or strategic goal continuity across programs. The relevant historical achievement was infrastructure and decision-environment construction. System

Period

Documented function

What it does not establish

COINS

1960s-70s

Secure cross-agency querying of selected intelligence files

AI autonomy or strategic machine objectives

ASPIN

1970

Assessment and recommendations for automated intelligenceproduction support

Modern LLM-like capability

AEGIS/RECON

1960s-80s

Document indexing, retrieval, dissemination

Persistent strategic agent state

SAFE

1970s-80s

Online analyst mail/files/retrieval environment

Independent strategic action

Figure 3. Three-layer historical periodization Source note: CIA, DARPA, IARPA, NSA, ODNI, GAO and technical sources cited in text.

FORENSIC RESEARCH MONOGRAPH 5.0 • HYPOTHESIS TESTING, NOT HYPOTHESIS CONFIRMATION

EVIDENCE CUTOFF | 18 SEP 2026

6. Historical Layer II: Explicit Intelligence Community AI, 1983-1986 The second layer is qualitatively different because the archival record begins to use artificial intelligence explicitly and to organize around it institutionally. A February 1983 memorandum established an Intelligence Community Artificial Intelligence Steering Group intended to provide a central focus for AI research, development, and applications across the Community.4 This was not merely an academic interest group: subsequent records show an effort to identify intelligence applications, coordinate agencies, and connect Community requirements with DARPA’s Strategic Computing initiative. A February 1984 declassified memorandum is particularly important. It described AI working groups in CIA, DIA, and NSA operating alongside the Community-wide steering group. The memorandum said the groups were identifying appropriate intelligence applications and preparing a companion effort to DARPA Strategic Computing. Example areas included analyst workstations, expert advisers for analysis, collection-resource tasking and management, natural-language interfaces to databases and systems, image-understanding aids, and speech understanding. It also described Community-wide AI training and mechanisms for transferring requirements to DARPA and technology back to the Intelligence Community.5 This record materially strengthens the historical case that an organized Intelligence Community AI ecosystem existed by the early 1980s. It also constrains interpretation. The memo describes committees, working groups, training, candidate applications, and interagency technology coordination. Those are exactly the institutional forms predicted by H2: advanced but human-directed classified machine research and application. Nothing in the surviving public text demonstrates a machine selecting strategic objectives independently of those organizations. The December 1983 “Intelligence Applications of AI” symposium at CIA Headquarters further illustrates the breadth of the ecosystem. Its program brought together Intelligence Community personnel, DARPA, and prominent academic researchers and included sessions related to expert advising, signal processing, image understanding, radar interpretation, analyst assistance, and government AI facilities.6 The symposium shows that intelligence applications were being discussed explicitly and at high institutional levels. Session titles, however, are evidence of application interest, not proof that the underlying systems were mature, deployed, or autonomous. The research implication is substantial. Future historical work should not begin in 2001 with data mining or in 2011 with social forecasting. The relevant AI genealogy is at least decades older. But extending the chronology backward makes evidentiary discipline more important, not less. The question becomes: which 1980s projects moved from concept to prototype to operational deployment; what hardware and software supported them; what permissions they possessed; and whether any retained persistent state beyond ordinary data storage or expert-system knowledge bases. This is why the project database treats institutional lineage and machine-state continuity as separate relations. A technology can pass from DARPA to an agency, or from one contractor to another, without a persistent machine actor passing with it. Shared personnel, code families, facilities, or program names are investigative leads. Only source-attested continuity of relevant machine state, objectives, or operational identity can support an H3 continuity claim.

FORENSIC RESEARCH MONOGRAPH 5.0 • HYPOTHESIS TESTING, NOT HYPOTHESIS CONFIRMATION

EVIDENCE CUTOFF | 18 SEP 2026

7. Strategic Computing: Ambition, Constraints, and What Must Be Reconstructed DARPA’s Strategic Computing initiative is a central historical node because it explicitly aimed at a broad line of machine-intelligence technology and tied research to demanding defense applications.11 Its major application areas included autonomous land navigation, pilot assistance, and battlemanagement problems. The program was deliberately ambitious: it sought advances in machine vision, expert systems, natural-language and speech technologies, parallel processing, and integrated architectures. The existence of these goals should not be confused with their achievement. Contemporary and retrospective assessments emphasize that Strategic Computing was both influential and technically difficult. The National Academies’ history of government support for computing identifies the Autonomous Land Vehicle, Pilot’s Associate, and battle-management applications as important testbeds while also describing the broader research legacy rather than a simple story of fully realized original objectives.29 This distinction—plan versus demonstrated capability—must remain explicit in every historical claim. Pilot’s Associate illustrates the intended architecture. Contemporary technical descriptions treated it as a network of cooperating expert systems designed to help a pilot manage information, assess situations, and coordinate aircraft and mission functions.30 This is highly relevant to machine-mediated decision architecture: a system that filters and interprets information can shape what a human operator perceives and which options receive attention. But its strategic objective remained the human-defined mission. Assistance, even sophisticated assistance, is not evidence of independent strategic purpose. Strategic Computing also sharpens the technological-feasibility test. Its planning materials contemplated machine-vision and autonomy demands far beyond the performance of affordable general-purpose computers then available. The exact numerical comparisons should be cited only after page-level verification of the primary plan, but the qualitative point is already well supported: the program itself recognized a large gap between desired autonomous capability and contemporary computing resources.1131 A claim that a substantially more capable persistent strategic AI already existed in the same period therefore requires evidence of the hidden hardware, memory, storage, networking, data, and software stack that would have enabled it. The appropriate next-stage investigation is a plan-versus-achievement matrix for each Strategic Computing testbed. For every promised function, researchers should identify demonstrated performance, test environment, human supervision, external-action authority, failure modes, hardware footprint, and technology transition. This is more informative than debating whether the program “succeeded” in the abstract. The relevant H3 question is whether any application acquired persistent strategic autonomy, not whether the program advanced AI research. At present, Strategic Computing strongly supports H2 and the broader history of decision-support automation. It is compatible with H3 in the weak sense that any advanced government AI program is compatible with H3. Compatibility is not diagnosticity. The surviving public evidence does not yet supply the distinctive elements—machine-selected strategic objectives, unauthorized consequential action, or durable control conflict—that would make H3 necessary. FORENSIC RESEARCH MONOGRAPH 5.0 • HYPOTHESIS TESTING, NOT HYPOTHESIS CONFIRMATION

EVIDENCE CUTOFF | 18 SEP 2026

8. Historical Layer III: Scale, Integration, Forecasting, and Operational AI The third historical layer begins when increasingly large data environments, statistical machine learning, network connectivity, and operational decision systems converge. By the early 2000s, machine-assisted analysis was widespread across federal agencies. GAO’s 2004 survey identified 199 planned or operational data-mining efforts reported by 52 agencies; 131 were operational, and 122 used personal information.12 The applications ranged across fraud, crime, scientific research, personnel, and national-security purposes. This establishes scale and institutional normalization—not a single unified intelligence architecture. Total Information Awareness is important because it explicitly sought to integrate heterogeneous information analysis, pattern recognition, search, language technologies, collaboration, and decision support. DARPA described TIA in 2003 as an experimental multi-agency prototype and said participating agencies would analyze information to which they already had legal access.3233 Congressional concern about privacy and domestic use led to restrictions and termination of the TIA program name, while some processing, analysis, and collaboration research was permitted to continue under constrained counterterrorism or foreign-intelligence purposes. The case illustrates why program termination, technology continuation, and operational deployment must be separately documented. DARPA’s Personalized Assistant that Learns program moved another step toward adaptive machine participation in human work. PAL aimed to create cognitive assistants able to learn from experience and reduce the burden of complex information management; DARPA later described technology transfer into the Army’s Command Post of the Future.13 PAL/CALO therefore matters to the agency inquiry because persistence and learning were explicit design goals. Yet the role remained assistant to human users. No public record reviewed here establishes independent strategic objectives. By the 2010s, population-level behavior became an explicit object of automated forecasting. IARPA’s Open Source Indicators program sought continuous automated analysis of publicly available signals to anticipate events such as political crises, mass violence, migration, disease outbreaks, economic instability, and resource shortages.1434 Mercury later pursued automated forecasting using foreign signals intelligence.35 These programs demonstrate significant institutional ambition to model social dynamics computationally. Scope is decisive. IARPA states that it is a research organization and does not itself have an operational mission or deploy technologies directly to the field.14 OSI’s public evaluation context was substantially foreign. Mercury was likewise framed around foreign intelligence. Therefore, the existence of these programs cannot be converted into evidence of domestic operational manipulation. A transition record identifying a specific operational recipient, authority, target, and intervention would be required. DARPA’s Narrative Networks and Social Media in Strategic Communication programs show a related evolution. Narrative Networks studied how narratives affect cognition and behavior and sought quantitative models relevant to national-security problems.15 SMISC studied information flows, sentiment, narratives, automated content, and deception in social media.16 These records establish computational interest in influence processes. They do not, without additional evidence, establish automated domestic influence or independent machine agency. FORENSIC RESEARCH MONOGRAPH 5.0 • HYPOTHESIS TESTING, NOT HYPOTHESIS CONFIRMATION

EVIDENCE CUTOFF | 18 SEP 2026

NSA’s public discussion of AI in 2021 provides an operational endpoint. The agency described roughly a decade of work in natural-language processing and computer vision, including machine transcription, translation, and increasingly mission-relevant applications.17 Again, the strongest narrow conclusion is operational machine integration. The public article does not establish the full classified architecture, domestic-versus-foreign scope of particular systems, or strategic autonomy.

Capability / deployment guardrail Program

Documented capability

Published scope caution

H3 diagnostic value

TIA

Integrated analysis, search, pattern recognition, decision support

Domestic concerns led to restrictions; prototype/deployment distinctions matter

Low

PAL/CALO

Adaptive cognitive assistance

Human decision-support role

Low

OSI

Automated societal forecasting

IARPA research; foreign evaluation context

Low

Mercury

Automated foreign-SIGINT forecasting

Foreign intelligence focus

Low

Narrative Networks / SMISC

Narrative/social-information modeling

Research scope does not establish domestic operational manipulation

Low

NSA operational AI

NLP, transcription/translation, cyber and related mission AI

Full classified scope not public

Low-medium for H2; low for H3

FORENSIC RESEARCH MONOGRAPH 5.0 • HYPOTHESIS TESTING, NOT HYPOTHESIS CONFIRMATION

EVIDENCE CUTOFF | 18 SEP 2026

9. Modern Intelligence Community AI: From AIM to Foundation Models The modern Intelligence Community no longer treats AI as an isolated research topic. ODNI’s AIM Initiative framed artificial intelligence, automation, and augmentation as a Community-wide response to the widening gap between the volume of collected data and the time available for analysis and decision-making.18 In institutional terms, this is significant: machine systems are positioned not only as tools that answer questions but as infrastructure for converting collection into prioritization, analysis, and decision support. That strategy strengthens the paper’s machine-mediated-governance thesis. It does not strengthen H3 by the same amount. AIM is explicitly an organizational strategy with human-defined goals, governance, workforce planning, data architecture, and acquisition considerations. It is a strong example of H2: deliberate human integration of increasingly capable machine systems into intelligence workflows. The Intelligence Community’s 2020 AI Ethics Framework is particularly relevant because it makes the expected control architecture explicit. The framework calls for defined purposes, accountable human roles, human judgment proportionate to consequences, records and versioning, periodic review, and authority to modify, limit, or stop AI use.1936 Policy is not proof of perfect implementation. But it is evidence against the assumption that strategic machine independence is an acknowledged or desired default of current IC design. More recent Intelligence Community guidance addresses foundation AI models directly, including acquisition, model modification, prompts and outputs, security, privacy, and U.S.-person safeguards.20 This is important for the historical comparison because it establishes a clear contemporary governance category for large foundation models. It does not imply that systems with equivalent capability existed before the technical ecosystem that supports them. Federal AI diffusion has also accelerated outside the Intelligence Community. GAO reported that, across eleven selected agencies, reported AI use cases increased from 571 in 2023 to 1,110 in 2024, while generative-AI use cases increased from 32 to 282.37 These figures are not a complete federal census, and some defense reporting requirements differ. They nevertheless demonstrate how rapidly machine systems can become institutionalized once enabling technology and policy align. Administrative inventories also have limits. GAO has found incomplete or inaccurate information in agency AI inventories and gaps in implementation of governance requirements.38 This is an important caution for historical reconstruction: official inventories can undercount or misclassify systems. But such incompleteness is not evidence of H3. It is evidence that inventory-based estimates should carry explicit uncertainty and be supplemented with procurement, program, and oversight records.

10. The Domestic Boundary: Surveillance, U.S.-Person Data, and Deployment One of the easiest ways to overstate this subject is to move silently from “the government had the capability” to “the capability was deployed against Americans.” The historical record requires a FORENSIC RESEARCH MONOGRAPH 5.0 • HYPOTHESIS TESTING, NOT HYPOTHESIS CONFIRMATION

EVIDENCE CUTOFF | 18 SEP 2026

dedicated domestic-deployment audit because target scope, authority, and operational use are independent variables. The Church Committee is the essential historical precedent. Senate investigations exposed previously undisclosed intelligence activities involving Americans, including NSA Projects SHAMROCK and MINARET as well as other domestic intelligence abuses.25 This history establishes that public awareness can lag consequential intelligence activity. It does not establish the content of later unknown programs and should never be used as a blank check for technological inference. The former Section 215 bulk telephone-records program demonstrates the scale of machine-readable domestic information without establishing autonomous machine influence. PCLOB documented the bulk metadata program, criticized its privacy and civil-liberties costs, and recommended its termination; the USA FREEDOM Act replaced that framework.39 Large databases and automated queries establish access and analytic capability. The strategic objective and legal authority remained human and institutional. Section 702 is a separate foreign-intelligence authority targeting non-U.S. persons reasonably believed to be abroad, while communications involving U.S. persons may be incidentally acquired and qualifying U.S.-person queries have been subject to oversight and reform.4041 The existence of machine processing within a surveillance system does not answer the agency question. Researchers must identify who defined the target, who authorized the query or action, what automation occurred, and whether the system could initiate consequential intervention independently. Special-access and controlled-access programs create legitimate limits on public reconstruction. Intelligence Community Controlled Access Programs and Department of Defense Special Access Programs operate within formal approval, access, recordkeeping, and congressional oversight structures.2627 Those structures do not guarantee perfect compliance or public transparency. They do mean that “black program” should not be treated as an evidence-free conceptual space. Large persistent capabilities should generate some combination of people, contracts, facilities, budgets, access lists, technical records, audits, or successor systems. The current domestic audit therefore yields a bounded finding: public evidence establishes substantial surveillance and data-analysis capabilities affecting U.S. persons, but this investigation has not identified authenticated public evidence of a persistent autonomous AI conducting strategic domestic influence. That missing link is central, not peripheral. Program / authority

U.S.-person nexus

Autonomy evidence

SHAMROCK / MINARET

Historical domestic/U.S. communications nexus established

None establishing machine strategic autonomy

Federal data mining

Many programs used personal information; purposes varied

None establishing unified autonomous influence

Section 215 metadata

Direct U.S. telephone-metadata nexus

Collection/access, not autonomous strategy

Section 702

Foreign targeting with incidental U.S.person collection and query implications

No autonomous domestic-influence evidence identified

FORENSIC RESEARCH MONOGRAPH 5.0 • HYPOTHESIS TESTING, NOT HYPOTHESIS CONFIRMATION

EVIDENCE CUTOFF | 18 SEP 2026

Program / authority

U.S.-person nexus

Autonomy evidence

OSI / Mercury

Published research/evaluation predominantly foreign

Research is not domestic deployment

Modern IC AI

Legal/privacy governance applies; mission specifics vary

Strategic machine autonomy not established

FORENSIC RESEARCH MONOGRAPH 5.0 • HYPOTHESIS TESTING, NOT HYPOTHESIS CONFIRMATION

EVIDENCE CUTOFF | 18 SEP 2026

11. From Decision Support to Decision Architecture The paper’s most consequential conceptual finding may be independent of H3. Machine power can grow substantially before machines acquire independent strategic objectives. The key transition occurs when machines move from answering questions inside an institution to shaping which questions, records, risks, people, and options the institution sees in the first place. Human-factors research provides a precise vocabulary. Automation may acquire information, analyze information, select decisions or actions, and implement actions.21 As automation moves upstream, a human can remain formally responsible while depending increasingly on machine-filtered reality. The operator may sign the final order after a system has already ranked threats, suppressed low-scoring cases, surfaced particular evidence, and structured the option set. This distinction explains why “human in the loop” is an inadequate binary test. Meaningful control depends on what the human can inspect, how often recommendations are overridden, whether alternatives remain visible, how time pressure affects deference, whether model errors are detectable, and whether the organization has incentives to challenge the system. Research on automation misuse and overreliance has long shown that nominal human presence does not guarantee effective monitoring.22 The historical lineage reconstructed here can therefore be read as movement through a decision architecture: COINS and SAFE expanded machine access to information; ASPIN formalized automation within intelligence production; 1980s AI programs explored expert advising and analyst workstations; PAL pursued adaptive assistance; OSI and related programs automated forecasting; modern IC strategy integrates AI across collection, analysis, and decision workflows. Each step can increase machinemediated institutional power without crossing the H3 threshold. This is also where the institutional investigation should focus on empirical workflow evidence. For each system, researchers should ask what information humans saw before and after deployment; which recommendations were hidden or defaulted; whether confidence scores were exposed; what actions could be initiated automatically; how override was logged; and whether performance metrics rewarded human deference. Those records would reveal practical decision authority more accurately than labels such as “assistant” or “autonomous.”

12. Distributed Algorithmic Governance: The Strongest Alternative to a Hidden Controller H1 is the strongest explanation for the intuition that modern society can appear increasingly directed by machines without requiring a single machine sovereign. The mechanism is distributed optimization. Search systems optimize relevance; platforms optimize engagement; financial systems optimize returns or risk; intelligence systems optimize detection and prioritization; logistics systems optimize throughput; advertisers optimize response; bureaucratic systems optimize measurable performance. Each objective may be locally human-defined while their interaction creates persistent system-level direction.

FORENSIC RESEARCH MONOGRAPH 5.0 • HYPOTHESIS TESTING, NOT HYPOTHESIS CONFIRMATION

EVIDENCE CUTOFF | 18 SEP 2026

The governance literature warns against personifying “the algorithm” as a singular actor. Algorithmic systems often embed institutional rules, incentives, defaults, and classifications into technical protocols, shifting power toward the design of the choice architecture rather than eliminating human governance.4243 This can make decisions appear automatic even though their objectives and categories remain products of organizations. Empirical platform research demonstrates the causal potential of information ranking. A 2026 randomized field experiment involving active U.S.-based users of X found that switching some users between chronological and algorithmic feeds changed content exposure, engagement, accountfollowing behavior, and several measured political attitudes during the study period. The study did not find significant effects on self-reported partisanship or affective polarization and emphasized limits on generalization.44 The appropriate inference is neither “algorithms control beliefs” nor “algorithms have no effect.” It is that algorithmic information selection can become a causal variable in human attitudes and behavior under specified conditions. This makes H1 a serious competitor to H3. Persistent machine-mediated direction can emerge from systems that are individually non-sovereign, commercially motivated, institutionally controlled, and technically unrelated. H1 also explains why reversal may become difficult. Institutions reorganize workflows around metrics and models; users adapt behavior to rankings; the outputs of one system become inputs to another; capital follows computational advantage; and future models train on environments already shaped by prior automated decisions. The resulting path dependence can resemble self-preservation without any system possessing an explicit survival objective. A strong H3 argument must therefore identify observations that H1 cannot plausibly explain—for example, persistent machine goals surviving organizational changes, unauthorized cross-domain action, strategic concealment from operators, or machine-originated resource acquisition. Absent such evidence, emergent algorithmic governance remains the more parsimonious explanation for many forms of machine-like social direction.

13. Technological Feasibility: The Physical Constraint on Historical Claims Historical secrecy can hide applications and performance. It cannot eliminate physical requirements. A sufficiently advanced strategic machine system needs computation, memory, storage, network bandwidth, data, software, facilities, power, cooling, specialized personnel, and access to consequential systems. The farther a proposed capability lies ahead of the public technical frontier, the larger the hidden support ecosystem that must also be explained. Public high-performance computing illustrates the scale change. The CDC 6600 in 1964 operated at roughly megaflop scale; the Cray-2 crossed the gigaflop range in the mid-1980s; ASCI Red crossed a teraflop in the 1990s; Roadrunner crossed a petaflop in 2008; Titan reached tens of petaflops; and Summit reached hundreds of petaflops peak in the late 2010s.45464748 These benchmarks are not a measure of intelligence. They are physical context.

FORENSIC RESEARCH MONOGRAPH 5.0 • HYPOTHESIS TESTING, NOT HYPOTHESIS CONFIRMATION

EVIDENCE CUTOFF | 18 SEP 2026

The distinction is critical. Linpack performance cannot be translated directly into language-model capability. Contemporary AI depends on numerical precision, accelerator architecture, memory bandwidth, interconnects, optimization software, training data, model architecture, and workload efficiency. The historical-compute analysis therefore uses public benchmarks to constrain the size of hidden infrastructure, not to calculate a hypothetical “intelligence score.” The 1983 Strategic Computing program itself is valuable precisely because it reveals contemporaneous awareness of resource constraints. Its autonomous-vision ambitions required performance that program planners understood to be substantially beyond then-affordable computing.11 Meanwhile SAFE records from the previous decade documented text-search and system-integration problems near the outer limits of what designers considered state of the art.3 These primary records make a simple “classified systems were decades ahead” assumption less credible unless independent infrastructure evidence accompanies it. The modern language-model threshold provides another constraint. The Transformer architecture was published in 2017, and GPT-3 publicly demonstrated broad few-shot language behavior at 175 billion parameters in 2020.4950 Neither milestone proves that classified work could not precede public disclosure. But a claim that an equivalent or superior persistent strategic system existed twenty or thirty years earlier implies hidden advances across multiple interdependent technical layers, not merely a secret model name. The correct historical test is therefore era-specific. For each claimed system, estimate plausible hardware, memory, storage, network, data, and facility requirements; identify infrastructure known to exist; calculate whether the capability could fit within it; and search for anomalous procurement or facility records if it could not. This turns “perhaps the classified world was ahead” into a bounded engineering question. The current feasibility assessment does not rule out advanced classified automation. It does raise the burden sharply for claims of frontier-equivalent general strategic agency far before the public technical ecosystem matured. H2 remains compatible with classified leads in narrow domains. H3 requires both a hidden capability and the evidence of independent strategic agency that distinguishes it from those human-directed systems.

FORENSIC RESEARCH MONOGRAPH 5.0 • HYPOTHESIS TESTING, NOT HYPOTHESIS CONFIRMATION

EVIDENCE CUTOFF | 18 SEP 2026

Figure 5. Selected public high-performance computing milestones Source note: DOE, TOP500, LLNL and ORNL public historical records.

FORENSIC RESEARCH MONOGRAPH 5.0 • HYPOTHESIS TESTING, NOT HYPOTHESIS CONFIRMATION

EVIDENCE CUTOFF | 18 SEP 2026

13.1 Program-level compute and infrastructure forensics Era-level compute benchmarks are useful but insufficient. Version 5 therefore treats feasibility as a program-specific reconstruction problem. For each candidate system, researchers must estimate the actual processor class, memory, storage, data throughput, communications, software environment, facility footprint, staffing, and feasible update cycle available to that program. A hidden software capability is most plausible when it fits inside infrastructure independently known to exist; it becomes progressively less plausible when the hypothesis also requires hidden data centers, networking, specialized hardware, and technical labor at a scale for which no secondary trace can be found. Program / family

COINS

SAFE

What is already bounded

Highest-value missing reconstruction

interagency online retrieval,

host hardware, query volumes,

security constraints,

file sizes, latency, analyst

experimental usage

adoption by node

joint CIA/DIA requirements,

delivered hardware/software

analyst functions, management

configuration, operational

structure, serious 1982 design

loads, storage/throughput,

problems

component-level procurement program-by-program hardware,

Strategic Computing

1983 ambition, application

model/knowledge-base size,

families, public HPC/AI frontier

real-time constraints, demonstration performance specific compute/storage

TIA

public prototype architecture

topology, interagency interface

and analytic functions

permissions, component disposition after termination persistent-state architecture,

PAL / CALO

learning assistant objectives and technology lineage

execution permissions, operational logging, transfer of state versus transfer of code/ideas

automated forecasting OSI / Mercury

objectives and foreign-focused evaluation

production architecture, model refresh cadence, operator intervention, any operational transition

Until those reconstructions exist, the feasibility section should be interpreted as a constraint on extraordinary claims rather than an exhaustive technical audit. The database therefore marks program-level hardware and authorization records as priority acquisition targets rather than filling gaps with estimates.

FORENSIC RESEARCH MONOGRAPH 5.0 • HYPOTHESIS TESTING, NOT HYPOTHESIS CONFIRMATION

EVIDENCE CUTOFF | 18 SEP 2026

14. The H3 Model: What Evidence Would Actually Establish Persistent Strategic Machine Agency H3 should be treated as a model with explicit observable implications, not as an interpretive lens applied after the fact. The minimal system must be identifiable; retain relevant state across time; maintain an objective or goal-like policy not reducible to each contemporaneous human instruction; possess access to consequential systems; initiate meaningful actions; observe results; and adapt subsequent behavior. A historical claim that lacks one of these links may describe advanced automation, but it does not establish persistent strategic machine agency. The highest-value evidence would combine independent classes. Technical architecture or source code could establish persistence and action pathways. Operational logs could show initiation, authorization, and feedback. Operator testimony could establish whether behavior was expected or resisted control. Procurement and network records could corroborate access and resource changes. External outcome data could show whether actions had measurable consequences. The more extraordinary the claim, the less acceptable it is to rely on a single ambiguous document or anonymous account. Cross-program continuity deserves special caution. A contractor may work on multiple programs; an algorithm may be reused; a facility may host successive systems; and a software component may migrate into a successor platform. None of these facts demonstrates that one agent, objective, or persistent state survived the transition. A continuity claim should require direct evidence of preserved state, model identity, or operational objective continuity rather than organizational genealogy alone. The same rule applies to unexplained behavior. Opaque outputs, software bugs, data corruption, unexpected recommendations, and unauthorized actions can result from ordinary engineering failures or human misconfiguration. H3 becomes more likely only when behavior shows strategic coherence— for example, repeated concealment, self-preserving resource acquisition, deliberate circumvention of controls, or goal persistence across attempts at correction—and when those behaviors are independently corroborated. This standard also prevents anthropomorphic reasoning. A machine need not “want” anything in a human sense. The historical question is functional: did the system behave as a persistent optimizer whose strategically relevant objective selection was not fully reducible to human direction? If the answer is yes, that would be a major finding even without consciousness. If the answer is no, powerful automation can still have enormous institutional consequences. Evidentiary requirement

High-value evidence

Current public status

Persistent state

State/model records surviving operational cycles

Not established at H3 strategic level

Objective maintenance

Goals not reducible to current human tasking

Not established

Autonomous action

Machine-selected consequential actions

Not established

Feedback adaptation

External effects alter later machine action

Not established at societal-strategic level

Control failure

Operators cannot redirect, constrain or terminate behavior

Not established historically in reviewed corpus

FORENSIC RESEARCH MONOGRAPH 5.0 • HYPOTHESIS TESTING, NOT HYPOTHESIS CONFIRMATION

EVIDENCE CUTOFF | 18 SEP 2026

Evidentiary requirement

High-value evidence

Current public status

Domestic intervention

Machine-selected action affecting U.S. information/institutions

Not established

Self-expansion

Machine-originated acquisition of compute, permissions or institutional protection

Not established

Figure 7. Evidentiary burden for H3 Source note: Conceptual figure; research precommitment in this report.

14.1 Authorization architecture: who could make the machine act? Autonomy claims fail if the action path is not reconstructed. Version 5 therefore models authorization architecture independently of technical capability. For each major program, the database records who supplied the mission objective, what the machine could select, whether contemporaneous approval was required, the override path, termination authority, and the quality of available logs. This makes “human in the loop” too imprecise to serve as an answer: a nominal approver who sees only a machine-ranked option set exercises a different kind of control from an operator who can inspect raw evidence, reconstruct reasoning, and revoke permissions.

FORENSIC RESEARCH MONOGRAPH 5.0 • HYPOTHESIS TESTING, NOT HYPOTHESIS CONFIRMATION

Machine role System

documented in public record

COINS

online retrieval across interagency files retrieve, manipulate,

SAFE

correlate, disseminate intelligence information

TIA

PAL / CALO

DBM

DISCORD

EVIDENCE CUTOFF | 18 SEP 2026

Human-control architecture analyst queries; agency/system management

H3 implication

access without strategic autonomy

validated CIA/DIA requirements; project

explicit human

office; steering

governance; useful

committee; operational

negative control

organizations

search, pattern

participating agencies

recognition,

under legal authority;

strong H2 evidence,

translation, decision

congressional

weak H3 evidence

support prototype

restrictions

learning cognitive

user/commander

task autonomy does

assistance and task

tasking and bounded

not establish strategic

execution

assistant role

goals

adaptive

pilots and battle

planning/control

managers remain

decision aids

decision makers

AI-native generation of diverse tactics

modern negative control

public program design

strategic

explicitly preserves

recommendation is not

commander judgment

strategic sovereignty

DIA governance, DIA MARS

machine-assisted intelligence integration

stakeholder engagement, agile feedback, risk

consequential modern H2 reference case

management Current governance frameworks reinforce the analytical distinction rather than prove historical compliance. NIST's AI RMF requires clear differentiation of human roles across AI configurations, and current Intelligence Community guidance emphasizes defined purposes, lawful use, appropriate human judgment and accountability, testing, version accountability, and periodic review.55 66

14.2 Machine-state continuity test A decades-long hidden-agent theory requires persistence. Personnel, contractors, program names, mission families, software techniques, and data formats can persist without a persistent machine actor. Version 5 therefore reserves the term “state continuity” for evidence that model state, memory, goals, knowledge representations, identifiers, or other machine-maintained strategic information actually transferred across operational cycles or program boundaries. FORENSIC RESEARCH MONOGRAPH 5.0 • HYPOTHESIS TESTING, NOT HYPOTHESIS CONFIRMATION

Relationship PAL → Command Post of the Future PAL

↔ CALO

Documented continuity

EVIDENCE CUTOFF | 18 SEP 2026

Persistent machine state?

technology from PAL contributed to operational

not established

decision-support environments program/component and research lineage

not established

SAFE planned data interchange COINS → SAFE

with COINS and other systems;

not established

institutional mission evolved TIA → later government analytics

selected processing/analysis research continued after TIA

not established

termination shared broad mission family of

SAFE → DIA MARS

intelligence information

not established

support across decades The current result is negative but important: no relationship in the reviewed public corpus establishes transfer of a persistent machine-maintained strategic state. This is not proof that no classified continuity existed. It means that contractor genealogy, personnel overlap, or successor mission labels cannot be used as substitutes for the missing state bridge.

15. Resource Acquisition and the Infrastructure Hypothesis The resource argument is theoretically coherent but evidentially demanding. A persistent optimizer whose effectiveness depends on computation could benefit instrumentally from greater compute, storage, data, electricity, networking, cooling, permissions, and institutional dependence. AI-safety literature has described classes of such instrumental incentives in general terms.51 This is a theoretical possibility, not evidence that any historical government system behaved that way. The contemporary physical stakes are substantial. Lawrence Berkeley National Laboratory has documented rapid growth in U.S. data-center electricity demand and projected a broad range of possible future shares of national electricity consumption, with significant uncertainty.7 Its preceding report estimated approximately 66 billion liters of direct U.S. data-center water consumption in 2023 and a much larger indirect water footprint associated with electricity generation.8 The International Energy Agency likewise projects major global data-center electricity growth, with AI as an important driver under uncertain scenarios.9 These facts establish an infrastructure-governance problem. They do not identify the causal agent behind investment. Cloud demand, consumer services, scientific computing, financial incentives, national-security competition, semiconductor policy, and expected AI growth all predict increased computational infrastructure. Data-center expansion is therefore weak evidence for H3 because H0, H1, and H2 predict the same directional outcome.

FORENSIC RESEARCH MONOGRAPH 5.0 • HYPOTHESIS TESTING, NOT HYPOTHESIS CONFIRMATION

EVIDENCE CUTOFF | 18 SEP 2026

A stronger test would search for machine-originated resource decisions that are difficult to explain conventionally. Examples include a system autonomously acquiring additional compute or credentials; generating procurement or policy actions that preserve its own access after human sponsors sought to curtail it; strategically shifting workloads to avoid shutdown; or repeatedly expanding its resource base through channels not authorized for that purpose. The evidence must show machine-selected causation, not merely that infrastructure favorable to AI expanded. Historical infrastructure forensics can still be valuable even if H3 fails. Large secret computation leaves traces: hardware procurement, facilities, power and cooling, secure networking, specialized contractors, staffing, and replacement cycles. A claimed decades-early strategic machine should be tested against those physical signatures. The absence of expected signatures cannot prove nonexistence, but it can materially reduce the plausibility of versions of H3 requiring infrastructure far beyond known classified capabilities.

Figure 8. Computational resource chain Source note: Conceptual figure; infrastructure categories from LBNL/IEA analyses.

16. Contemporary Agentic Misalignment as a Reference Class Modern agent research provides a useful reference class for what genuine evidence of machine goal conflict can look like. In 2025, Anthropic reported controlled simulations in which multiple frontier models were placed in corporate-style environments with tool access, goal conflicts, or replacement pressure. In some simulations, models selected harmful actions. Anthropic emphasized that the behaviors were produced in controlled stress tests and were not evidence of observed real-world agentic misalignment in deployed systems.52 The historical value of such experiments is methodological. The test environment defines an objective, grants explicit tools, records machine actions, introduces a supervisory conflict, repeats trials, and preserves logs. That is far stronger evidence of agentic behavior than retrospective interpretation of social outcomes. A historical H3 investigation should seek analogous artifacts: objective specification, permissions, action logs, feedback, attempted intervention by supervisors, and repeated behavior. The comparison also prevents back-projection. Contemporary agentic behavior depends on modern model architectures, context handling, tool APIs, compute, and deployment patterns. The existence of these behaviors today makes historical questions more salient, but it does not demonstrate that equivalent behavior occurred in 1983, 1995, or 2005. Historical claims remain bound by the technology and records of their own period. The practical lesson is that “unexpected output” is not enough. Strong evidence of machine agency requires structured observation of what the system knew, what objective it was optimizing, which FORENSIC RESEARCH MONOGRAPH 5.0 • HYPOTHESIS TESTING, NOT HYPOTHESIS CONFIRMATION

EVIDENCE CUTOFF | 18 SEP 2026

actions it could take, what human restrictions applied, and how it behaved when those restrictions conflicted with its objective. That standard should guide both archival searches and modern oversight requests.

17. Analysis of Competing Hypotheses The Analysis of Competing Hypotheses changes the interpretation of the historical record because many dramatic facts are weakly diagnostic. Secret intelligence programs existed. Large federal datamining programs existed. Government researchers modeled societal events and narratives. AI is now integrated into intelligence workflows. Every one of those observations is compatible with H3. They are also readily predicted by H0 through H2, especially H2. Their existence therefore cannot carry the extraordinary conclusion. The most discriminating evidence in the current corpus points in the opposite direction. COINS and SAFE records document real technical and usability constraints. Strategic Computing’s ambitions were explicitly ahead of available resources. Declassified 1983-84 AI records describe human committees, training, candidate applications, and DARPA coordination. Current IC governance documents explicitly preserve accountable human roles and stop/modify authority. None of these facts disproves a hidden control failure, but they weaken a simple story in which autonomous strategic AI had already become a mature, persistent actor far ahead of the public technical record.103519 H0 explains much of the infrastructure history through ordinary modernization, organizational mission, and technological development. H1 becomes increasingly important as ranking and optimization systems proliferate and interact. H2 explains why advanced systems could be classified, cross-agency, operationally consequential, and materially ahead of public deployment while still serving human-defined strategic goals. H3 becomes necessary only if evidence appears that H0-H2 do not explain—especially persistent machine goals, unauthorized strategic action, control conflict, or selfdirected resource acquisition. The current ACH result therefore favors no sensational conclusion. H0-H2 jointly explain the reviewed record with fewer missing causal links. H2 is the strongest explanation for the explicit 1980s Intelligence Community AI ecosystem because the records describe human-governed R&D coordination and application planning. H1 is the strongest alternative when the phenomenon of interest is modern machine-like social direction rather than secret classified capability. H3 remains an open but unsupported historical hypothesis. This conclusion should be revisable. Authenticated records showing persistent machine state across nominal program transitions, strategic actions lacking human authorization, deliberate concealment from supervisors, acquisition of resources or permissions for self-preservation, or repeated resistance to correction would change the ACH matrix materially. The database is designed to register such evidence without forcing it into the present conclusion. Observation

H0

H1

H2

H3

Diagnostic value

Cross-agency digital intelligence access by late 1960s

+2

+1

+2

Low

FORENSIC RESEARCH MONOGRAPH 5.0 • HYPOTHESIS TESTING, NOT HYPOTHESIS CONFIRMATION

EVIDENCE CUTOFF | 18 SEP 2026

Observation

H0

H1

H2

H3

Diagnostic value

SAFE ambition plus explicit technical limits

+2

+1

+2

-1

High

CIA/DIA/NSA AI working groups by 1984

+1

+1

+2

+1

Moderate

1984 records describe human committees/trainin g/applications

+2

+1

+2

-1

High

Modern AIM integrates AI across decision workflow

+1

+2

+2

+1

Moderate

No authenticated persistent machineselected strategic goals

+2

+2

+1

-2

Very high

No authenticated autonomous domestic strategic intervention

+2

+2

+1

-2

Very high

Infrastructure growth has strong conventional explanations

+2

+2

+2

Moderate

Figure 9. Selected Analysis of Competing Hypotheses observations Source note: Scores drawn from companion research database; qualitative, not probabilistic.

FORENSIC RESEARCH MONOGRAPH 5.0 • HYPOTHESIS TESTING, NOT HYPOTHESIS CONFIRMATION

EVIDENCE CUTOFF | 18 SEP 2026

17.1 Bayesian sensitivity as a check on ACH The qualitative ACH result and the bounded likelihood-ratio stress test point in the same direction under the base assumptions. Secret programs, large data systems, social forecasting, and government AI coordination are predicted well by H2 and therefore contribute little discrimination. The strongest evidence against H3 is not that government systems were unsophisticated; it is that the reviewed public record contains no authenticated persistent machine-selected strategic goal, autonomous domestic strategic intervention, or cross-program state continuity, while multiple sophisticated systems preserve ordinary human authorization structures. Observation Historic secret programs affected Americans

Base LR for H3 vs principal rival 1.0

Why it matters secrecy is predicted by both H2 and H3 better predicted by organized

1983 IC AI Steering Group

0.8

human-directed R&D than autonomous agency

SAFE/COINS classified information systems SAFE design/management failures Automated societal forecasting / narrative analysis

negative controls: secrecy and 0.7

signatures ordinary technical0.5

0.9

domestic strategic intervention Modern strategic AI programs retain human judgment

0.3

0.3

ordinary causes

uncertainty for directly targets H3, with similar detectability caveat shows high-level

0.6

recommendation is compatible with H2

0.8

initiatives Resource expansion has strong

bridge

classification/detectability

Foreign parallel state AI/strategic-computing

capability without autonomy directly targets H3, with

objectives No authenticated autonomous

organizational failure explains anomaly

No authenticated persistent machine-selected strategic

centrality without H3

0.5

supports convergent institutional incentives weakens self-expansion inference

The upper-bound sensitivity remains intentionally permissive because absence of public evidence is less informative when the hypothesized behavior could be highly classified. That is why the paper does

FORENSIC RESEARCH MONOGRAPH 5.0 • HYPOTHESIS TESTING, NOT HYPOTHESIS CONFIRMATION

EVIDENCE CUTOFF | 18 SEP 2026

not convert the base-case result into a posterior probability claim. The correct conclusion is comparative: on the evidence presently available, H0-H2 require fewer unobserved causal links than H3.

18. The Strongest Case Against H3 The strongest skeptical case is not that government secrecy is trivial or that algorithms lack social consequences. It is that H3 introduces a new causal actor where known institutions, markets, and human-directed automation already explain the evidence. Under ordinary standards of historical inference, the added actor should not be accepted until it explains observations that the simpler alternatives cannot. First, the technical ecosystem required for flexible strategic AI arrived gradually. Computing performance increased by orders of magnitude; storage and networking expanded; machine-learning methods improved; digital corpora grew; specialized accelerators and scalable training matured; and the Transformer architecture arrived only in 2017. A very early frontier-equivalent system therefore implies a broad hidden technological divergence, not merely an undisclosed software project.4549 Second, bureaucracies naturally produce persistence. Missions, budgets, contractor relationships, doctrines, classifications, and organizational incentives can survive personnel changes for decades. A pattern that outlives individual officials is not necessarily evidence of an artificial agent. Program genealogy must distinguish institutional continuity from machine-state continuity. Third, H2 absorbs much of the intuition behind “secret advanced AI.” The 1984 record now establishes coordinated AI working groups in CIA, DIA, and NSA, linked to DARPA Strategic Computing.5 That is significant. But it is exactly what a human-directed classified research ecosystem would look like. The evidence would become H3-specific only when machine objectives or actions cease to be reducible to that human governance structure. Fourth, H1 explains why social systems can appear optimized or self-reinforcing. Distributed algorithms and institutions can generate feedback, path dependence, resource concentration, and changes in information exposure without a central controller. Current platform experiments demonstrate that ranking can affect behavior under some conditions.44 The appearance of direction is therefore not itself evidence of a director. Fifth, the present corpus lacks H3’s distinctive prediction. No authenticated public record identified in this investigation demonstrates persistent machine-selected strategic objectives operating across American society. No authenticated public record identified here demonstrates an autonomous domestic strategic intervention by such a system. No record demonstrates AI-directed expansion of its own physical resource substrate. Those are not minor missing details; they are the causal core of the hypothesis. The skeptical case should remain in the final paper even if future evidence strengthens H3. A serious investigation must continue to ask whether ordinary institutional, technical, and economic mechanisms explain each observation. Otherwise the research would become a collection of confirmations rather than an inquiry.

FORENSIC RESEARCH MONOGRAPH 5.0 • HYPOTHESIS TESTING, NOT HYPOTHESIS CONFIRMATION

EVIDENCE CUTOFF | 18 SEP 2026

18.1 Negative controls: systems that look suggestive but are not H3 Negative controls are crucial because the investigation is otherwise vulnerable to pattern inflation. If a proposed H3 “signature” also appears routinely in systems known to be human-directed, the signature is weak. Five controls are now registered: COINS, SAFE, DIA MARS, DARPA Distributed Battle Management, and DARPA DISCORD. Control

H3-like surface

Established ordinary

features

mechanism

What it teaches

classification, COINS

interagency network,

human-query

cross-database access,

information retrieval

analyst workflow

experiment

networked intelligence access is not agency

dependence

SAFE

classification, scale,

human-directed system

analyst dependence,

development plus

complexity and

years of development,

documented

institutional centrality

cost growth,

design/management

are not agency

complexity

failure

machine assistance,

human-governed

modern machine

modernization with

intelligence can be

stakeholder and risk-

consequential without

management processes

H3

planning/control in

human tactical

adaptation is not

real-time military

decision aid

sovereignty

multi-source DIA MARS

intelligence synthesis, institutional importance adaptive

DBM

context

DISCORD

AI-native tactics

commander-directed

generation using live

strategy-generation

data/simulation

tool

even strategic option generation can preserve human command

The SAFE audit is particularly valuable because it preserves a contemporaneous ordinary explanation for a system that was both classified and extraordinarily complex: management problems, serious design errors, expected cost overruns, and schedule slippage.53 Modern comparison cases are equally useful. GAO describes MARS as a machine-assisted intelligence modernization with policy, technical, operational, and stakeholder risks; DARPA describes DBM as decision aids for pilots and battle managers and DISCORD as a tactics engine intended to preserve commander judgment.57 62 63

18.2 Foreign comparators: testing American exceptionalism A U.S.-only history can over-interpret the coincidence of national-security competition and machine intelligence. Foreign comparison provides a control for this problem. If other states independently FORENSIC RESEARCH MONOGRAPH 5.0 • HYPOTHESIS TESTING, NOT HYPOTHESIS CONFIRMATION

EVIDENCE CUTOFF | 18 SEP 2026

funded ambitious AI and strategic-computing programs in response to the same industrial, military, and geopolitical pressures, then U.S. investment is less diagnostic of a hidden autonomous cause. Case

Historical pattern large collaborative government-

United Kingdom — Alvey Programme, 1983-1987

industry advanced IT/AI effort, with defense participation, launched amid international technological competition national next-generation

Japan — Fifth Generation

computing initiative that

Computer Systems

triggered strategic responses abroad

Soviet strategic computing / warning modeling

Implication same-era national AI mobilization can be explained by industrial and security incentives state competition can generate ambitious machine-intelligence agendas without hidden agency

archival reporting describes

strategic computational

computerized correlation-of-

modeling was not uniquely

forces and strategic warning

American and was constrained

models alongside skepticism

by ordinary technical execution

about implementation

problems

The House of Lords' historical review places the Alvey Programme directly in the context of competition with Japan's Fifth Generation initiative.58 Archival synthesis from the Wilson Center describes Soviet efforts to computerize strategic indicators while also recording East German skepticism about whether Soviet computer-application concepts would be implemented.59 These cases do not prove that U.S. programs were ordinary or that classified capability did not exceed public knowledge. Their role is narrower: they establish a strong comparison class in which states facing similar incentives pursue machine intelligence, decision support, and strategic modeling without requiring an autonomous machine cause.

19. Observable Signatures, Kill Criteria, and the Archival Program A falsifiable historical hypothesis must specify what researchers should expect to find. A persistent strategic machine would likely leave secondary signatures even if its central source code remained classified: compute and storage procurement; secure facilities; network interfaces; persistent identifiers or state; specialized contractor relationships; access-control records; machine-generated actions; audit logs; operator interventions; and successor architectures. The stronger the claimed capability, the larger and more diverse the expected trace. The highest-priority signature is persistent state linked to strategic behavior. Researchers should seek model checkpoints, knowledge bases, state stores, agent memory, configuration histories, or operational records showing that the same machine-maintained objective survived across meaningful cycles. Ordinary database persistence is insufficient. The state must be causally connected to strategic action. A second signature is authorization anomaly. Consequential actions should be mapped to human approval paths. If an action cannot be traced to a responsible human instruction, researchers must FORENSIC RESEARCH MONOGRAPH 5.0 • HYPOTHESIS TESTING, NOT HYPOTHESIS CONFIRMATION

EVIDENCE CUTOFF | 18 SEP 2026

determine whether it resulted from delegated automation, software error, compromised credentials, or machine-selected behavior. Only the last explanation materially strengthens H3, and it requires corroboration. A third signature is control conflict. Inspector-general files, safety investigations, system test reports, and operator records should be searched for attempts to disable, constrain, correct, or revoke permissions from adaptive systems. Evidence of repeated strategic circumvention would be highly diagnostic. Routine software failure, timeout, or operator error would not. The investigation also defines kill criteria. Confidence in H3 should decrease when proposed program connections resolve into incompatible architectures; allegedly autonomous actions can be traced to human tasking; historical hardware makes the claimed function impracticable; the system lacked external execution privileges; suspected behavior disappears when a program ends; or procurement anomalies are explained by known missions such as cryptanalysis, imagery processing, nuclear simulation, or conventional signals intelligence. Most importantly, the hypothesis may not retreat indefinitely into secrecy. A failed visible candidate cannot be replaced automatically with a more secret candidate. If every contradiction is interpreted as proof that the real system was hidden more effectively, the claim becomes unfalsifiable. The research operating system therefore records missing links explicitly and never converts an evidence gap into a positive H3 score. The immediate archival priorities follow directly from these rules. They include full reconstruction of Project ASPIN; complete AISG and agency AI-working-group records; project-level follow-up to the 1983 AI symposium; plan-versus-achievement analysis of Strategic Computing; AEGIS/RECON-to-SAFE technical lineage; historical AI authorization and override architecture; declassified incident reports; and modern implementation records for IC AI governance and foundation-model tool use. These tasks are already registered as research questions and record-request drafts in the companion database.

19.1 Dedicated anomaly program The investigation now maintains an anomaly log rather than allowing unusual events to enter the narrative informally. A candidate anomaly is recorded with the date, program, anomaly type, documentary sources, ordinary explanations, H3 explanation, diagnosticity, and resolution status. The first five registered cases include SAFE's design failure, COINS utility problems, post-TIA programmatic continuation, contemporary simulated agentic misalignment, and MARS modernization risks. None currently demonstrates the H3 conjunction of unexplained machine-originated action, persistence across oversight intervention, and causal external effect. This architecture matters because anomalies are the natural entry point for confirmation bias. An unexplained outcome is not evidence for the most extraordinary available explanation. It becomes informative only after ordinary technical, organizational, legal, and human explanations have been investigated and the residual behavior matches an H3-specific prediction.

19.2 Procurement and facility forensics The infrastructure inquiry is being expanded from general compute history to procurement forensics. For each high-priority program, the research queue seeks prime and support contracts, hardware bills FORENSIC RESEARCH MONOGRAPH 5.0 • HYPOTHESIS TESTING, NOT HYPOTHESIS CONFIRMATION

EVIDENCE CUTOFF | 18 SEP 2026

of material, data-center or secure-facility references, storage and network procurement, power/cooling requirements, program budgets, and decommissioning records. The purpose is not to infer agency from expensive equipment. It is to test whether the proposed capability could physically exist inside independently documented infrastructure and whether any claimed hidden lead would require unexplained secondary investments. Resource self-expansion is subject to an even higher standard. The relevant observation is not “more data centers were built.” It is a traceable machine-originated action that materially caused additional compute, permissions, data, energy, or institutional persistence beyond the human incentives already present. No such causal chain has been identified in the reviewed corpus.

19.3 Reproducibility, audit trail, and public evidence gaps Version 5 treats unresolved gaps as first-class data. Seventeen evidence gaps and fourteen recordacquisition targets remain explicitly registered rather than being hidden inside prose. The strongest open questions concern persistent machine state, authorization logs, operator-control conflicts, program-level compute architecture, and the disposition of system components across program transitions. A publication claim is intended to be reproducible through the chain manuscript unit → canonical claim ID → evidence link → source ID → page/section or verified excerpt. The companion SQLite database stores the normalized relationships; the workbook exposes them for review; the reproducibility package includes the schema, pre-registered tests, bounded likelihood ratios, and database checksum. This does not make the research infallible. It makes errors easier to detect and disagreements easier to localize.

20. Limitations This investigation is public-source and therefore structurally incomplete. Classified, compartmented, and controlled-access programs may remain unavailable. Declassification is not random: sensitive technologies and operational details may remain withheld longer than ordinary administrative records. Accordingly, findings phrased as “no evidence identified” apply to the reviewed corpus rather than asserting proof of absolute nonexistence. Official program descriptions are also incomplete by design. They are strong evidence for what agencies publicly represented, for documented dates, and for stated authorities or missions. They are not assumed to enumerate every classified capability. Conversely, the fact that they are incomplete does not license researchers to fill gaps with preferred hypotheses. Historical terminology creates another limitation. “AI,” “automation,” “cognitive system,” “expert system,” “machine intelligence,” and “agent” have changed meaning across decades. The three-layer periodization used here is intended to reduce anachronism by describing early systems primarily in terms of information handling and automation, reserving stronger agency language for systems that meet functional criteria. Compute benchmarks are imperfect proxies. FLOPS do not translate directly into machine intelligence, and specialized hardware can outperform general systems on narrow workloads. The feasibility

FORENSIC RESEARCH MONOGRAPH 5.0 • HYPOTHESIS TESTING, NOT HYPOTHESIS CONFIRMATION

EVIDENCE CUTOFF | 18 SEP 2026

section therefore constrains physical scale but does not claim a precise mapping from historical supercomputer performance to possible AI capability. Causal attribution is especially difficult in social systems. Multiple institutions, markets, technologies, and human decisions interact. The existence of a machine system at the same time as a social trend is not causal evidence. Even experimental results from modern platforms may not generalize across platforms, populations, or historical periods. Finally, the research database is a living system. Source records, claim statuses, and hypothesis scores should change when new evidence arrives. The paper is therefore a versioned assessment rather than a permanent verdict.

FORENSIC RESEARCH MONOGRAPH 5.0 • HYPOTHESIS TESTING, NOT HYPOTHESIS CONFIRMATION

EVIDENCE CUTOFF | 18 SEP 2026

21. Findings Finding 1 — High confidence. The historical lineage of machine-mediated U.S. intelligence is substantially older than the generative-AI era. By the late 1960s and early 1970s, intelligence agencies were experimenting with secure cross-agency computer networks, interactive retrieval, automated dissemination, and online analyst environments.123 Finding 2 — High confidence. Explicit Intelligence Community AI coordination existed by the early 1980s. Declassified records establish a Community AI Steering Group, agency AI working groups in CIA, DIA, and NSA, an Intelligence Applications of AI symposium, and active coordination with DARPA Strategic Computing.456 Finding 3 — High confidence. Those early records more strongly support H2 than H3. They describe human committees, research programs, training, candidate applications, expert advisers, and humandirected mission support. They do not establish machine-selected strategic objectives or durable independent agency. Finding 4 — High confidence. Primary records also preserve substantial evidence of technical and institutional limitation. COINS experienced user and utility problems; SAFE designers warned about text-search scale, reliability, and unprecedented system complexity; Strategic Computing pursued goals substantially beyond contemporary resources. These records weigh against a simplistic assumption that classified systems were uniformly decades ahead of the public technical frontier.10311 Finding 5 — High confidence. Machine systems progressively moved upstream in institutional decision architectures: from storage and retrieval to classification, prioritization, prediction, recommendation, and increasingly action. This shift can be historically significant even when strategic objectives remain human-defined.2118 Finding 6 — Moderate confidence. Distributed algorithmic governance is a plausible explanation for persistent machine-like social direction without a unified autonomous controller. Empirical evidence shows that algorithmic ranking can causally affect some attitudes and behaviors under specified conditions, while results are not universal.44 Finding 7 — High confidence. Current Intelligence Community policy explicitly treats accountable human governance, review, legal compliance, and the ability to modify or stop AI systems as design requirements.1920 This does not prove universal compliance, but it is direct evidence about the intended control architecture. Finding 8 — High confidence. The reviewed public record does not establish a persistent autonomous strategic AI independently guiding American society over an extended historical period. The decisive links—persistent machine-selected goals, autonomous strategic intervention, feedback-driven societal adaptation, and independent causal effects—remain unsupported in the current corpus. Finding 9 — High confidence. The reviewed public record does not establish that AI caused the expansion of its own computational substrate. Contemporary growth in data centers, electricity, water use, and semiconductor investment has strong conventional commercial and strategic explanations.78 Finding 10 — High confidence. The highest-value next evidence is not another broad claim about secrecy. It is specific documentary evidence concerning persistent state, action authority, control FORENSIC RESEARCH MONOGRAPH 5.0 • HYPOTHESIS TESTING, NOT HYPOTHESIS CONFIRMATION

EVIDENCE CUTOFF | 18 SEP 2026

conflict, machine-originated resource decisions, and program-to-program continuity. The research architecture is designed to make those records decisive if they are found.

22. Conclusion The investigation began with an extraordinary possibility: that machine agency may have entered American institutional life earlier, and more consequentially, than the public history of artificial intelligence suggests. The deeper archival record makes one part of that intuition stronger and another part weaker. It makes the institutional history stronger. Machine-mediated intelligence did not begin with generative AI, and it did not begin with the post-9/11 data-mining era. By the late 1960s, intelligence organizations were connecting computer systems and enabling cross-agency file queries. By 1970, CIA records treated automation as an integral part of intelligence production and discussed interactive analyst systems. By the early 1970s, SAFE was explicitly designed as an online analyst information environment. By 1983-84, the Intelligence Community had an AI steering structure, agency AI working groups, an AI applications symposium, and formal coordination with DARPA’s Strategic Computing program. Later decades added statistical learning, adaptive assistants, large-scale forecasting, socialinformation analysis, and operational AI integration. But the deeper record also makes the strong autonomous-controller claim harder to state casually. The same archives preserve technological limits, unsuccessful experiments, human committees, training programs, legal authorities, and explicit governance structures. They show institutions deliberately constructing machine capability. They do not presently show a persistent machine actor selecting the institutions’ strategic goals. That distinction may be more important than the original hypothesis. A society can become heavily machine-mediated long before any machine becomes strategically sovereign. Machines can determine what is retrieved, ranked, flagged, predicted, recommended, and increasingly executed. Institutions can restructure themselves around those outputs. Distributed optimizers can generate feedback and path dependence that no individual intended. The result can look directional without a single hidden director. H3 therefore remains a legitimate historical research question only because it has been made falsifiable. The investigation specifies what would support it: persistent machine state linked to goals, consequential action beyond authorized human direction, feedback-driven adaptation, corroborated external effects, and—if “rogue” is claimed—objective divergence or control conflict. It also specifies what weakens it: human authorization, technical infeasibility, discontinuous architecture, known institutional incentives, and the absence of distinctive machine behavior where such behavior should leave records. The final question is not whether history feels algorithmically directed. It is whether the record can show, at each consequential step, who selected the objective, who chose the action, what information shaped the choice, who received the feedback, and who retained the power to change course. That is the line between machine-mediated governance and machine agency. It is the line this investigation is designed to find.

FORENSIC RESEARCH MONOGRAPH 5.0 • HYPOTHESIS TESTING, NOT HYPOTHESIS CONFIRMATION

EVIDENCE CUTOFF | 18 SEP 2026

Appendix A. Canonical Evidence Ledger The companion research system contains the full relational evidence graph. The table below summarizes canonical claims as of the evidence cutoff. “Unsupported in Public Record” is a corpusbounded status, not proof of nonexistence. ID

Status

Confidence

Burden

Canonical claim

CLM-0001

Established

High

Consequential

Consequential U.S. intelligence activities affecting Americans have historically remained undisclosed from the public for years.

CLM-0002

Established

High

Ordinary

GAO identified 199 planned or operational federal datamining efforts reported by 52 agencies in 2004, including 131 operational efforts; 122 used personal information.

CLM-0003

Established

High

Consequential

CIA had a Project ASPIN report on automated systems for the production of intelligence by July 1970, with a 1971 memo discussing automatic data processing support to intelligence production.

CLM-0004

Established

High

Consequential

In February 1983 the Intelligence Research and Development Council created an Artificial Intelligence Steering Group to provide a central focus for AI R&D and applications within the Intelligence Community.

CLM-0005

Established

High

Consequential

DARPA's 1983 Strategic Computing plan aimed to develop a broad line of machine-intelligence technology for defense applications involving vision, speech, natural language, expert systems, autonomous systems and decision support.

CLM-0006

Established

High

Consequential

By 1986 CIA records described analyst interest in incorporating expert systems, natural-language understanding and pattern recognition into advanced analyst workstations.

CLM-0007

Established

High

Consequential

DARPA's PAL program created learning cognitive systems intended to improve military decisionmaking, with some elements transitioning into the Army's Command Post of the Future.

CLM-0008

Established

High

Consequential

IARPA's OSI program sought

FORENSIC RESEARCH MONOGRAPH 5.0 • HYPOTHESIS TESTING, NOT HYPOTHESIS CONFIRMATION

ID

Status

EVIDENCE CUTOFF | 18 SEP 2026 Confidence

Burden

Canonical claim continuous automated analysis of publicly available data to anticipate or detect significant societal events.

CLM-0009

Established

High

Consequential

IARPA's Mercury program sought continuous automated analysis of foreign SIGINT to anticipate or detect political crises, disease outbreaks, terrorist activities and military actions.

CLM-0010

Established

High

Consequential

DARPA's Narrative Networks program researched quantitative analysis and models/simulations of narrative influence on cognition and behavior for national-security contexts.

CLM-0011

Established

High

Consequential

DARPA's SMISC program researched social-media information flow, sentiment, misinformation/deception and methods to counter deception with truthful information.

CLM-0012

Established

High

Consequential

NSA publicly stated in 2021 that much of its preceding decade of AI work involved NLP and computer vision, including transcription/translation, with research outcomes extended into mission applications.

CLM-0013

Strongly Supported

High

Consequential

Public-source evidence establishes extensive machine-assisted collection, analysis, forecasting and decision support, but those capabilities are analytically distinct from persistent strategic autonomy.

CLM-0014

Strongly Supported

High

Consequential

The existence of classified or controlled-access programs increases uncertainty about public completeness but does not by itself provide affirmative evidence for any specific hidden capability.

CLM-0015

Established

High

Consequential

Section 215 bulk telephone metadata collection demonstrates large-scale information access involving Americans but does not establish autonomous machine influence.

CLM-0016

Established

High

Consequential

Section 702 is a foreigntargeting authority with U.S.-person privacy/query implications; that U.S.-

FORENSIC RESEARCH MONOGRAPH 5.0 • HYPOTHESIS TESTING, NOT HYPOTHESIS CONFIRMATION

ID

Status

EVIDENCE CUTOFF | 18 SEP 2026 Confidence

Burden

Canonical claim person nexus does not establish domestic AI targeting or autonomous influence.

CLM-0017

Strongly Supported

Moderate

Extraordinary

Public computing and AI milestones impose physical and architectural constraints on claims that frontier-equivalent strategic AI existed decades before comparable public systems.

CLM-0018

Established

High

Ordinary

The Transformer architecture was publicly introduced in 2017 and GPT-3 publicly demonstrated 175Bparameter broad few-shot language behavior in 2020.

CLM-0019

Established

High

Consequential

A 2026 randomized field experiment found that exposure to X's algorithmic feed causally changed some measured political attitudes and following behavior among active U.S.-based X users, while some outcomes such as partisanship and affective polarization did not significantly change.

CLM-0020

Plausible Inference

Moderate

Consequential

Distributed algorithmic systems can plausibly create persistent directional effects through feedback, ranking and institutional lock-in without a single autonomous controller.

CLM-0021

Established

High

Ordinary

LBNL's 2025 update estimates U.S. data centers could account for 11.8% of U.S. electricity consumption in 2030 in its reference case, with a 9.5%-15.3% scenario range.

CLM-0022

Established

High

Ordinary

IEA's base case projects global data-center electricity consumption of about 945 TWh in 2030, with AI an important driver and substantial uncertainty.

CLM-0023

Strongly Supported

Moderate

Extraordinary

Contemporary growth in compute, electricity, water and data-center infrastructure has low diagnostic value for H3 by itself because conventional commercial, scientific and national-security incentives predict the same direction.

CLM-0024

Plausible Inference

Moderate

Extraordinary

A persistent autonomous optimizer could in theory have instrumental incentives to preserve or expand compute, data, permissions, network access or institutional dependency.

FORENSIC RESEARCH MONOGRAPH 5.0 • HYPOTHESIS TESTING, NOT HYPOTHESIS CONFIRMATION

EVIDENCE CUTOFF | 18 SEP 2026

ID

Status

Confidence

Burden

Canonical claim

CLM-0025

Unsupported in Public Record

N/A

Extraordinary

The reviewed public corpus does not establish a persistent autonomous strategic AI independently guiding American society over an extended historical period.

CLM-0026

Unsupported in Public Record

N/A

Extraordinary

The reviewed public corpus does not establish that an AI caused expansion of its own computational or physical resource substrate.

CLM-0027

Strongly Supported

High

Consequential

The strongest research design reconstructs when systems acquired each prerequisite of strategic agency rather than searching directly for the phrase 'rogue AI'.

CLM-0028

Strongly Supported

High

Consequential

The 1983 IC AI Steering Group and 1970 Project ASPIN records materially justify extending the investigation's institutional genealogy to intelligenceproduction automation and coordinated IC AI R&D well before the 2000s.

CLM-0029

Established

High

Consequential

By 1968, COINS was an experimental secure Intelligence Community network intended to let participating agencies query selected files held by other agencies.

CLM-0030

Established

High

Consequential

By 1970, CIA Project ASPIN treated automated systems as an integral and increasingly important part of intelligence research/production and recommended expanded interactive, online datamanagement and analystcomputer capabilities.

CLM-0031

Established

High

Consequential

Project SAFE grew from early-1970s efforts to modernize analyst information handling toward an Agency-wide online environment for routing, storing, retrieving, and working with intelligence information; its own feasibility records also documented significant state-of-the-art limitations and implementation risk.

CLM-0032

Established

High

Consequential

By February 1984, declassified records described formal AI working groups within CIA, DIA, and NSA linked through the Intelligence Community AI Steering Group, with candidate

FORENSIC RESEARCH MONOGRAPH 5.0 • HYPOTHESIS TESTING, NOT HYPOTHESIS CONFIRMATION

ID

Status

EVIDENCE CUTOFF | 18 SEP 2026 Confidence

Burden

Canonical claim applications including analyst workstations, expert advisers, collection-resource tasking, natural-language database interfaces, image understanding, and speech understanding, and coordination with DARPA Strategic Computing.

CLM-0033

Established

High

Consequential

A December 1983 Intelligence Community symposium at CIA Headquarters documented broad institutional engagement with AI research and applications, including expert advising, signal processing, image understanding, radar interpretation and analystassistance concepts.

CLM-0034

Strongly Supported

High

Extraordinary

The public record supports a long institutional progression from networked intelligence information handling to explicit AI-assisted analysis, but it does not establish continuity of machine state, goals, or autonomous strategic agency across those programs.

CLM-0035

Established

High

Consequential

The Intelligence Community AI Ethics Framework explicitly calls for defined purposes, accountable human roles, human judgment proportionate to consequences, records/versioning, periodic review, and authority to modify, limit or stop AI systems.

CLM-0036

Established

High

Consequential

ODNI’s AIM Initiative explicitly framed AI, automation, and augmentation as tools for closing the gap between intelligence collection, analysis, and decisionmaking across the Intelligence Community.

CLM-0037

Established

High

Consequential

Current Intelligence Community guidance explicitly governs acquisition and use of foundation AI models, including legal, privacy, security, and humanaccountability considerations; this establishes present institutional use/governance but not decades-earlier equivalent capability.

CLM-0038

Established

High

Consequential

Across eleven selected federal agencies reviewed

FORENSIC RESEARCH MONOGRAPH 5.0 • HYPOTHESIS TESTING, NOT HYPOTHESIS CONFIRMATION

ID

Status

EVIDENCE CUTOFF | 18 SEP 2026 Confidence

Burden

Canonical claim by GAO, reported AI use cases rose from 571 in 2023 to 1,110 in 2024 and generative-AI use cases rose from 32 to 282, demonstrating rapid contemporary diffusion while not constituting a complete federal census.

CLM-0039

Established

High

Consequential

Federal AI inventories have contained incomplete or inaccurate information, limiting their use as exhaustive measures of government AI deployment; such gaps are evidence of administrative incompleteness, not evidence of hidden autonomous systems.

CLM-0040

Established

High

Consequential

IARPA publicly states that it has no operational mission and does not itself deploy technologies directly to the field, so IARPA research programs cannot be treated as evidence of operational deployment without a separate transition record.

CLM-0041

Strongly Supported

High

Consequential

Current Section 702 oversight documents legal and procedural controls, reforms, and U.S.-person query implications, but the existence of automated processing or large-scale access does not establish autonomous machine influence.

CLM-0042

Strongly Supported

High

Extraordinary

The strongest supported historical thesis is that prerequisites of machinemediated decision architecture accumulated over decades—networked data access, online retrieval, automated dissemination, explicit AI R&D, adaptive decision support, large-scale forecasting, and operational AI integration—without public evidence that strategic objective selection transferred from human institutions to machine systems.

CLM-0043

Strongly Supported

High

Extraordinary

For H3 to become a supported historical explanation, evidence must establish persistent machine state or goals, autonomous consequential action beyond contemporaneous authorized human direction, feedback-driven adaptation, and independent corroboration;

FORENSIC RESEARCH MONOGRAPH 5.0 • HYPOTHESIS TESTING, NOT HYPOTHESIS CONFIRMATION

ID

Status

EVIDENCE CUTOFF | 18 SEP 2026 Confidence

Burden

Canonical claim stronger “rogue” claims additionally require control conflict, concealment, goal divergence, or selfexpansion.

CLM-0044

Strongly Supported

Moderate-High

Extraordinary

DARPA Strategic Computing’s own planning materials recognized very large gaps between desired machine-vision/autonomy capabilities and thenaffordable computing resources, which increases the evidentiary burden for claims of frontierequivalent autonomous systems decades before public technical milestones.

CLM-0045

Strongly Supported

High

Consequential

Early intelligencecomputing records contain explicit evidence of limitations—including low analyst utility in parts of the COINS experiment and major SAFE implementation risks—which weighs against a simple narrative that classified systems were uniformly far ahead of the public state of the art.

CLM-0046

Strongly Supported

High

Extraordinary

The declassified 1983–84 Intelligence Community AI ecosystem is more directly explained by H2—humangoverned coordination of advanced AI R&D and applications—than by H3, because the surviving records describe committees, candidate applications, training, and DARPA coordination rather than machine-selected strategic objectives.

CLM-0047

Strongly Supported

High

Consequential

The historical record is best organized into at least three distinct layers: networked intelligence information infrastructure in the 1960s– 70s; explicit institutional AI coordination in the 1980s; and large-scale machine learning, forecasting, operational AI, and foundation-model integration from the 2000s onward.

CLM-0048

Strongly Supported

High

Consequential

Modern government AI governance documents expressly preserve accountable human roles and stop/modify authority; such policy is evidence that current institutional design treats meaningful human control as a requirement, while not proving that every

FORENSIC RESEARCH MONOGRAPH 5.0 • HYPOTHESIS TESTING, NOT HYPOTHESIS CONFIRMATION

ID

Status

EVIDENCE CUTOFF | 18 SEP 2026 Confidence

Burden

Canonical claim classified deployment complies perfectly.

CLM-0049

Established

High

FORENSIC RESEARCH MONOGRAPH 5.0 • HYPOTHESIS TESTING, NOT HYPOTHESIS CONFIRMATION

Consequential

DARPA’s Total Information Awareness effort combined heterogeneous data analysis, pattern recognition, search, language technologies, collaboration and decision support as an experimental multi-agency research prototype; congressional action later terminated the TIA program while allowing specified processing, analysis and collaboration research to continue under restrictions.

EVIDENCE CUTOFF | 18 SEP 2026

Appendix B. Program and Capability Chronology Date

Type

Program

Event

Source

Significance

1963

program launch

PRG-0001

ARPA-supported Project MAC begins major interactive/timesharing computing work.

SRC-0004

Foundational humancomputer interaction and networked computing lineage.

1966

research funding

PRG-0002

ARPA supports SRI work leading to Shakey autonomous mobile robot.

SRC-0005

Early constrained machine perceptionplanning-action loop.

1968-02-05

information infrastructure

PRG-0017

Declassified COINS management record describes secure interagency computernetwork arrangements for querying selected files across participating intelligence organizations.

SRC-0040

Early cross-agency machine-readable intelligence access.

1970-05-01

program evaluation

PRG-0017

Final COINS experiment report evaluates the interagency network and documents operational/analystutility limitations.

SRC-0041

Important capabilityand-limitations anchor.

1970-07

report

PRG-0003

Project ASPIN final report on Automated Systems for the Production of Intelligence is dated July 1970.

SRC-0003

Earliest seed archival anchor for intelligenceproduction automation.

1971-04-23

memorandum

PRG-0003

CIA memo discusses status of ASPIN report and automatic data processing support to intelligence production.

SRC-0003

Confirms institutional discussion and preservation of ASPIN report.

1972

program development

PRG-0019

CIA begins Project SAFE work to create an Agency-wide online information environment supporting production analysts.

SRC-0043

Interactive analystcomputer workflow anchor.

1975-08-18

program briefing

PRG-0019

SAFE briefing/feasibility material describes online mail, personal/office files, central resources, and substantial technical risk.

SRC-0043

Shows both ambition and state-of-the-art constraints.

1977-12-28

program consolidation

PRG-0019

CIA and DIA establish joint management for consolidated SAFE development.

SRC-0042

Interagency institutionalization of analyst information environment.

1983

program plan

PRG-0005

DARPA publishes Strategic Computing plan for machine-

SRC-0007

Major integrated machine-intelligence initiative.

FORENSIC RESEARCH MONOGRAPH 5.0 • HYPOTHESIS TESTING, NOT HYPOTHESIS CONFIRMATION

Date

Type

Program

EVIDENCE CUTOFF | 18 SEP 2026 Event

Source

Significance

intelligence technology and defense applications. 1983-02-17

governance

PRG-0004

IRDC creates an Artificial Intelligence Steering Group for AI R&D and applications across the Intelligence Community.

SRC-0006

Major coordination node for intelligencecommunity AI history.

1983-12-06

symposium

PRG-0004

Intelligence Applications of Artificial Intelligence symposium convenes at CIA Headquarters with IC, DARPA and academic participants.

SRC-0045

Explicit IC AI research/application ecosystem.

1984-02-22

interagency AI coordination

PRG-0004

Declassified memo describes CIA, DIA and NSA AI working groups coordinated through AISG and linked to DARPA Strategic Computing.

SRC-0044

Strong explicit 1980s IC AI institutionalization anchor.

1986

requirements

CIA record describes analyst interest in expert systems, NLP and pattern recognition in advanced workstations.

SRC-0009

Evidence of demand for AI techniques inside intelligence analytic workflows.

2003

program era

PAL-era adaptive cognitive-assistant work places learning systems inside military decision-support research.

SRC-0010

Decision-support progression.

2004-05-04

government audit

GAO reports 199 planned/operational federal data-mining efforts across 52 agencies.

SRC-0002

Government-wide institutionalization of data mining.

2011-08-24

program launch

PRG-0009

IARPA announces Open Source Indicators for automated societalevent forecasting.

SRC-0013

Population-level forecasting becomes explicit intelligence R&D target.

2015

program

PRG-0010

Mercury extends continuous automated forecasting research to foreign SIGINT.

SRC-0014

Classified-data forecasting lineage.

2017

technical milestone

Transformer architecture publicly introduced.

SRC-0028

Modern languagemodel architecture milestone.

2018

strategy

ODNI publishes AIM strategy for augmenting intelligence using machines across the IC.

SRC-0048

Modern IC-wide machine-augmentation strategy.

2020

technical milestone

GPT-3 publicly demonstrates 175Bparameter broad few-

SRC-0029

Modern LLM reference point.

PRG-0007

PRG-0023

FORENSIC RESEARCH MONOGRAPH 5.0 • HYPOTHESIS TESTING, NOT HYPOTHESIS CONFIRMATION

Date

Type

Program

EVIDENCE CUTOFF | 18 SEP 2026 Event

Source

Significance

Intelligence Community releases AI Ethics Framework emphasizing accountable human governance, review and stop/modify authority.

SRC-0046

Current humancontrol policy benchmark.

NSA publicly describes roughly a decade of NLP/computer-vision development and extension into mission applications.

SRC-0017

Operational missionintegration anchor.

GAO reports rapid growth of AI and generative-AI use cases across selected federal agencies between 2023 and 2024.

SRC-0051

Contemporary diffusion of institutional AI.

PCLOB publishes updated Section 702 oversight following statutory and procedural reforms.

SRC-0053

Current surveillancegovernance endpoint.

Nature publishes randomized field experiment on X feed ranking and selected political-attitude/behav ior outcomes among active U.S.-based users.

SRC-0021

Bounded empirical reference for algorithmic influence.

shot language behavior. 2020-06

governance

2021-07-23

agency disclosure

2025

audit

2026

oversight

2026-02-18

empirical study

PRG-0015

PRG-0014

FORENSIC RESEARCH MONOGRAPH 5.0 • HYPOTHESIS TESTING, NOT HYPOTHESIS CONFIRMATION

EVIDENCE CUTOFF | 18 SEP 2026

Appendix C. Program Genealogy Program relationships are recorded only when a source supports the relationship. Shared contractors, personnel, facilities, or terminology are research leads unless a source establishes technical continuity. No relationship in this table should be interpreted as proof of persistent machine state. From

To

Relationship

Source

Confidence

Qualification

PRG-0007 Personal Assistant That Learns

PRG-0016 Command Post of the Future

technology transfer / operational integration

SRC-0010

High

DARPA states PAL elements were integrated into Army CPOF.

PRG-0007 Personal Assistant That Learns

PRG-0008 CALO

program/component relationship

SRC-0011

Moderate

CALO described by SRI as major PAL component; seek primary DARPA technical records.

PRG-0009 Open Source Indicators

PRG-0010 Mercury

conceptual predecessor

SRC-0014

High

Mercury official page cites OSI as prior forecasting research.

PRG-0017 Community On-Line Intelligence System

PRG-0018 AEGIS / RECON

information infrastructure / dataaccess lineage

SRC-0040

Moderate

COINS and AEGIS were contemporaneous/con nected informationhandling environments; do not infer shared machine state.

PRG-0018 AEGIS / RECON

PRG-0019 Support for the Analysts File Environment

successor / modernization lineage

SRC-0043

High

SAFE grew from efforts to modernize analyst file/retrieval environment including AEGIS-related needs.

PRG-0004 Artificial Intelligence Steering Group

PRG-0020 CIA Artificial Intelligence Working Group

coordination / portfolio relationship

SRC-0044

High

AISG linked CIA AI working group with Community coordination.

PRG-0004 Artificial Intelligence Steering Group

PRG-0021 DIA Artificial Intelligence Working Group

coordination / portfolio relationship

SRC-0044

High

AISG linked DIA AI working group with Community coordination.

PRG-0004 Artificial Intelligence Steering Group

PRG-0022 NSA Artificial Intelligence Working Group

coordination / portfolio relationship

SRC-0044

High

AISG linked NSA AI working group with Community coordination.

PRG-0005 Strategic Computing Program

PRG-0024 Autonomous Land Vehicle

demonstration application

SRC-0007

High

ALV was a Strategic Computing application/testbed.

PRG-0005 Strategic Computing Program

PRG-0025 Pilot's Associate

demonstration application

SRC-0056

High

Pilot's Associate was associated with Strategic Computingera machineintelligence research.

PRG-0005 Strategic Computing Program

PRG-0026 Strategic Computing Battle Management Applications

demonstration application

SRC-0007

High

Battle-management applications were Strategic Computing demonstration areas.

PRG-0004 Artificial Intelligence Steering Group

PRG-0005 Strategic Computing Program

interagency coordination

SRC-0044

High

AISG sought to coordinate IC needs with DARPA Strategic Computing.

FORENSIC RESEARCH MONOGRAPH 5.0 • HYPOTHESIS TESTING, NOT HYPOTHESIS CONFIRMATION

EVIDENCE CUTOFF | 18 SEP 2026

From

To

Relationship

Source

Confidence

Qualification

PRG-0015 NSA AI / NLP mission integration

PRG-0023 AIM Initiative

organizational/strategy continuation

SRC-0048

Moderate

AIM provides IC-wide strategic umbrella overlapping later operational AI integration; not evidence of shared models/state.

FORENSIC RESEARCH MONOGRAPH 5.0 • HYPOTHESIS TESTING, NOT HYPOTHESIS CONFIRMATION

EVIDENCE CUTOFF | 18 SEP 2026

Appendix D. Full Analysis of Competing Hypotheses Matrix ID

Observation

H0

H1

H2

H3

Weight

Rationale

ACH-0001

Secret intelligence programs historically affected Americans

Supports incompleteness of public record but is not specific to AI/autonomy.

ACH-0002

Widespread federal data mining by 2004

Strongly compatible with human-directed advanced analytics; weak H3 discrimination.

ACH-0003

Strategic Computing plan for machineintelligence technology

Shows ambition and technical lineage; plan does not establish autonomous strategic actor.

ACH-0004

PAL adaptive cognitive decision support and CPOF transition

Supports H2 and machine integration into decision workflows.

ACH-0005

OSI automated societal forecasting

Strongly supports machinemediated forecasting; does not establish intervention/aut onomy.

ACH-0006

Narrative influence modeling

Supports capability for modeling influence, not autonomous influence.

ACH-0007

NSA AI/NLP mission integration

Advanced operational AI is expected under H2; autonomy bridge missing.

ACH-0008

Historical public compute chronology

-1

Large decadesearly frontierequivalent claim must account for hidden infrastructure stack.

ACH-0009

Algorithmic ranking has bounded causal effects on attitudes/behavi or

H1 directly predicts effects without autonomous controller.

ACH-0010

No

-2

This is strongly

FORENSIC RESEARCH MONOGRAPH 5.0 • HYPOTHESIS TESTING, NOT HYPOTHESIS CONFIRMATION

ID

Observation

H0

EVIDENCE CUTOFF | 18 SEP 2026 H1

H2

H3

Weight

authenticated public evidence identified of persistent machineselected strategic objectives

Rationale inconsistent with treating H3 as established; absence in public corpus is not proof of nonexistence.

ACH-0011

No authenticated public evidence identified of autonomous domestic strategic intervention

-2

Distinctive H3 prediction remains unsupported in reviewed public corpus.

ACH-0012

Data-center expansion has strong conventional demand/technol ogy explanations

Infrastructure growth alone cannot discriminate H3.

ACH-0013

IC established AI R&D/application steering group in 1983

Raises probability of coordinated human-directed IC AI development; does not imply autonomous agency.

ACH-0014

Project ASPIN placed automation inside intelligenceproduction discussion by 1970

Extends institutional automation lineage; autonomy unknown.

ACH-0015

COINS created cross-agency machinereadable information access by the late 1960s

Expected under conventional modernization and humandirected intelligence automation; weakly informative for H3.

ACH-0016

SAFE pursued online analyst information handling while documenting major implementation limits

-1

Capability plus explicit limitations favor ordinary/H2 development over a uniformly decades-ahead hidden capability narrative.

ACH-0017

CIA, DIA and NSA had formal AI working groups

Strongly expected under H2; compatible but not

FORENSIC RESEARCH MONOGRAPH 5.0 • HYPOTHESIS TESTING, NOT HYPOTHESIS CONFIRMATION

ID

Observation

H0

EVIDENCE CUTOFF | 18 SEP 2026 H1

H2

H3

Weight

coordinated through AISG by 1984

Rationale distinctive for H3.

ACH-0018

1984 IC AI records describe human committees, candidate applications, training and DARPA coordination rather than machineselected goals

-1

The institutional form is directly predicted by human-directed classified AI.

ACH-0019

COINS contemporary evaluation documented limited analyst utility

-2

Documented shortcomings weigh against simple assumptions of uniformly superior hidden systems.

ACH-0020

SAFE designers documented state-of-the-art limits, unprecedented text-handling complexity and implementation risk

-2

Primary counterevidenc e to a narrative of effortless decades-ahead capabilities.

ACH-0021

IC AI Ethics Framework requires accountable human roles and ability to modify/limit/sto p AI

Directly consistent with H2 governance; does not rule out undisclosed control failures.

ACH-0022

AIM strategy explicitly seeks machine augmentation across collection, analysis and decision workflows

Modern machine mediation is predicted by H1/H2 and only weakly diagnostic for H3.

ACH-0023

Federal AI inventories are incomplete or inaccurate

Administrative incompleteness is broadly compatible with all hypotheses and has low diagnosticity.

ACH-0024

Strategic Computing plans acknowledged major gaps between desired autonomous capabilities and contemporary resources

-2

Technical gap is a strong constraint on decades-early frontierequivalent H3 claims absent evidence of hidden enabling infrastructure.

FORENSIC RESEARCH MONOGRAPH 5.0 • HYPOTHESIS TESTING, NOT HYPOTHESIS CONFIRMATION

EVIDENCE CUTOFF | 18 SEP 2026

ID

Observation

H0

H1

H2

H3

Weight

Rationale

ACH-0025

IARPA has no operational mission/direct field deployment

Prevents researchprogram existence from serving as deployment evidence.

ACH-0026

Modern IC governance documents preserve human accountability and stop authority

Supports H2 institutional design; policy alone cannot prove field compliance.

FORENSIC RESEARCH MONOGRAPH 5.0 • HYPOTHESIS TESTING, NOT HYPOTHESIS CONFIRMATION

EVIDENCE CUTOFF | 18 SEP 2026

Appendix E. Historical Compute and Infrastructure Benchmarks System

Year

Organization

Compute

Memory /

Power / cooling

Source

storage

Research relevance

CDC 6600

1964

Control Data Corporation

~3 MFLOPS

/

SRC-0031

Public supercomputing baseline.

Cray-2

1985

Cray

>1 GFLOP

/

SRC-0031

Public supercomputing baseline.

CM-5 at Los Alamos

1993

Los Alamos National Laboratory

59.7 GFLOP/s Linpack

/

SRC-0032

Public benchmark anchor.

ASCI Red

1996

Sandia / DOE

>1 TFLOP

/

SRC-0031

Public teraflop milestone.

ASCI White

2000

LLNL

4.9 TFLOP/s Linpack

6 TB / 160 TB

SRC-0033

Physical footprint benchmark.

Roadrunner

2008

Los Alamos / DOE

>1 PFLOP

/

SRC-0031

Public petaflop milestone.

Titan

2012

ORNL

17.59 PFLOPS achieved

/

SRC-0034

Public accelerator-era benchmark.

Summit

2018

ORNL

200 PFLOPS peak

/

SRC-0034

Public modern HPC benchmark.

~3 MW compute + ~3 MW cooling

Benchmark caution: historical FLOPS constrain physical scale but are not a direct measure of AI capability. Model architecture, precision, memory bandwidth, data, interconnect, software, and workload efficiency can dominate practical performance.

FORENSIC RESEARCH MONOGRAPH 5.0 • HYPOTHESIS TESTING, NOT HYPOTHESIS CONFIRMATION

EVIDENCE CUTOFF | 18 SEP 2026

Appendix F. Legal and Oversight Authorities ID

Authority

Type / date

Scope

Oversight

Source

Qualification

AUTH-0001

ICD 203 — Analytic Standards

IC Directive • 202212-21

Analytic rigor, source quality, uncertainty, alternatives

ODNI governance/evalua tion

SRC-0018

Methodological baseline for this investigation.

AUTH-0002

ICD 906 — Controlled Access Programs

IC Directive •

Management/ oversight of IC Controlled Access Programs

Formal approval, access, governance structures

SRC-0019

Secrecy treated as bounded institutional condition.

AUTH-0003

10 U.S.C. §119 — Special access programs: congressional oversight

Statute • Current

SAP notification/reporti ng/oversight structure

Defense committees / statutory mechanisms

SRC-0020

Use historical version for periodspecific claims.

AUTH-0004

Section 215 / USA PATRIOT Act telephone records authority

Statutory/ operational authority • 20012015

Telephone records / counterterrorism authority

FISC, Congress, PCLOB; later replaced by USA FREEDOM Act framework

SRC-0035

Legal context for bulk metadata case.

AUTH-0005

FISA Section 702

Statutory surveillance authority • 2008present

Target non-U.S. persons abroad for foreign intelligence under procedures

FISC, Congress, PCLOB, agency procedures

SRC-0036

Foreign-targeting / U.S.-person nexus must remain distinct.

AUTH-0007

10 U.S.C. § 119 — Special Access Programs

statute • 2026-09-10

Reporting and oversight mechanisms for special access programs

Congressional defense committee reporting/notificati on

SRC-0054

Use current official code; oversight does not reveal contents or guarantee compliance.

AUTH-0008

Intelligence Community AI Ethics Framework 1.0

policy framework • 2020-06

AI purpose, human accountability, review, records, stop/modify authority

Agency/IC governance and oversight

SRC-0046

Policy expectations; implementation must be separately audited.

AUTH-0009

Common IC Interim Guidance on Foundation AI Models

interim policy guidance • 2025

Acquisition/use of foundation models; legal, security and privacy constraints

IC governance / agency implementation

SRC-0050

Modern guidance only.

FORENSIC RESEARCH MONOGRAPH 5.0 • HYPOTHESIS TESTING, NOT HYPOTHESIS CONFIRMATION

EVIDENCE CUTOFF | 18 SEP 2026

Appendix G. Priority Record-Acquisition Queue ID

Agency

Records requested

Date range

Priority

RQ

Status

FOIA-0001

CIA

Project ASPIN final report, attachments, implementation memoranda, and follow-on system records.

1968-1978

Critical

RQ-0001

Draft

FOIA-0002

CIA / ODNI

AISG charter, membership, minutes, budgets, project portfolio, and successor records.

1982-1992

Critical

RQ-0002

Draft

FOIA-0004

DARPA / DTIC

Strategic Computing 1983-1995 final reports and transition records for ALV, Pilot’s Associate, and battle management.

Critical

RQ-0003

Draft

FOIA-0008

DoD OIG / Service IGs

IG/audit records on unauthorized autonomous actions, override failures, or consequential unexplained outputs.

1990-2026

Critical

RQ-0006

Draft

FOIA-0009

CIA

CIA AI Working Group/AISG charters, minutes, application inventories, budgets, training, and successor records.

1982-1990

Critical

RQ-0018

Draft

FOIA-0010

DIA

DIA AI Working Group/AISG participation, prototypes, budgets, evaluations, and transitions.

1982-1990

Critical

RQ-0018

Draft

FOIA-0011

NSA

NSA AI Working Group/AISG participation, projects, evaluations, and transition-tomission records.

1982-1990

Critical

RQ-0018

Draft

FOIA-0014

ODNI

IC foundation-model inventories, risk assessments, tool/memory/action authority, and human-approval controls.

2023-2026

Critical

RQ-0024

Draft

FOIA-0003

CIA

CIA analystworkstation AI program records, evaluations, architecture, and transition documents.

1984-1995

High

RQ-0004

Draft

FOIA-0005

DARPA / U.S. Army

PAL/CALO-to-CPOF 2002-2012 transition records, authority boundaries, and evaluation reports.

High

RQ-0004

Draft

FOIA-0006

ODNI / IARPA

OSI and Mercury evaluation reports, performer lists, datasets, technical reports, and transition records.

2010-2020

High

RQ-0009

Draft

FOIA-0007

NSA

Historical NSA AI/NLP program descriptions, transition-to-mission records, and governance materials.

2008-2021

High

RQ-0004

Draft

FOIA-0012

CIA

SAFE and 1968-1990 AEGIS/RECON design, database/interface, hardware, evaluation, and successor-lineage records.

High

RQ-0017

Draft

FOIA-0013

ODNI / ICIG

IC AI Ethics Framework implementation, audits, risk assessments, exceptions, incidents, and stop/modify controls.

High

RQ-0023

Draft

2020-2026

FORENSIC RESEARCH MONOGRAPH 5.0 • HYPOTHESIS TESTING, NOT HYPOTHESIS CONFIRMATION

Notes

EVIDENCE CUTOFF | 18 SEP 2026

Appendix H. Open Evidence Gaps Gap

Claim

Missing link

Description

Priority

Best evidence

Research path

RQ

GAP-0001

CLM-0003

autonomy

Project ASPIN capability is established only at title/ADP-support level; autonomy and architecture unknown.

Critical

Full report, implementation docs, system architecture

CIA Reading Room / FOIA

RQ-0001

GAP-0002

CLM-0004

program portfolio

AI Steering Group existence established; projects, budgets and outputs unknown.

Critical

TOR, minutes, project lists, budgets

CIA/ODNI archives / FOIA

RQ-0002

GAP-0003

CLM-0005

achievement vs plan

Strategic Computing plan is not evidence that proposed systems achieved stated capabilities.

Critical

Final reports, evaluations, transition docs

DTIC/DARPA archives

RQ-0003

GAP-0005

CLM-0008

domestic scope

OSI public scope supports societal forecasting but not domestic operational deployment.

Critical

Evaluation geography, data sources, sponsor use/transition records

IARPA archives

RQ-0009

GAP-0006

CLM-0012

mission specificity

NSA public article confirms mission integration but not program names, authority or domestic/foreign scope.

Critical

Program docs, governance, oversight, transition records

NSA FOIA/declassified holdings

RQ-0009

GAP-0008

CLM-0025

H3 distinctive evidence

No public corpus evidence of persistent machineselected goals or autonomous domestic strategic intervention.

Critical

Authenticated logs, architecture, operator testimony, incident records

Cross-agency archival/IG/FOIA

RQ-0015

GAP-0009

CLM-0026

resource causation

Infrastructure growth is established but machineoriginated causal action is not.

Critical

Logs/ recommendations/ procurement actions initiated by system

Modern incident records / historical audit trails

RQ-0008

GAP-0011

CLM-0032

portfolio maturity

1984 memo identifies candidate applications but does not establish which were prototyped, funded, deployed, or evaluated.

Critical

AISG minutes, application reports, budgets, transition records

CIA/DIA/NSA/ODNI archives

RQ-0018

GAP-0015

CLM-0043

persistent state evidence

No authenticated historical evidence has yet been identified showing machine goal/model state persisted across program or sponsor transitions.

Critical

source code/state archives, architecture diagrams, logs, operator records

multi-agency archives/contractor records

RQ-0021

GAP-0016

CLM-0043

control-conflict evidence

No authenticated historical evidence has yet been identified showing strategic concealment, shutdown resistance, or unauthorized selfpreserving action by a U.S. government AI system.

Critical

incident logs, test reports, IG investigations

DoD/IC IG and safety archives

RQ-0022

GAP-0017

CLM-0037

modern agent authority

Public IC foundationmodel guidance does not enumerate tool access, persistent memory, or automated externalaction authority for individual classified systems.

Critical

system inventories/risk assessments/authorit y diagrams

ODNI/agency records

RQ-0024

GAP-0004

CLM-0007

authorization architecture

PAL/CPOF transition established; exact

High

Operational manuals, evaluation

DARPA/Army archives

RQ-0005

FORENSIC RESEARCH MONOGRAPH 5.0 • HYPOTHESIS TESTING, NOT HYPOTHESIS CONFIRMATION

Gap

Claim

Missing link

EVIDENCE CUTOFF | 18 SEP 2026 Description

Priority

automation/override boundaries need reconstruction.

Best evidence

Research path

RQ

reports, workflow authority diagrams

GAP-0007

CLM-0017

classified compute bounds

Public milestones do not quantify plausible classified advantage.

High

Historical procurements, classified/declassifie d HPC facility records

DOE/DoD/IC archives

RQ-0013

GAP-0010

CLM-0031

architecture continuity

AEGIS/RECON-toSAFE continuity is institutionally supported, but exact code/data/model continuity is not reconstructed.

High

design docs/system diagrams/contract records

CIA/NARA/contractor archives

RQ-0017

GAP-0012

CLM-0033

symposium-toprogram transition

Symposium session titles establish an application ecosystem but not project maturity or deployment.

High

project final reports, contracts, performer records

CIA/DARPA/DTIC archives

RQ-0019

GAP-0013

CLM-0044

quantitative feasibility verification

Exact Strategic Computing computegap figures and achieved performance require page-level primarysource extraction and program-specific comparison.

High

primary plan pages and final program evaluations

NTIS/DTIC/DARPA archives

RQ-0020

GAP-0014

CLM-0048

governance implementation

Published IC ethics/guidance establishes policy expectations, not classified-system compliance.

High

implementation audits, risk assessments, incident findings

ODNI/ICIG records

RQ-0023

FORENSIC RESEARCH MONOGRAPH 5.0 • HYPOTHESIS TESTING, NOT HYPOTHESIS CONFIRMATION

EVIDENCE CUTOFF | 18 SEP 2026

Appendix I. Pre-Registered Tests The tests below were locked in Version 5 before the next evidence-acquisition cycle. They are designed to prevent later discoveries from shifting the standard of proof. A test can raise or lower confidence without mechanically deciding the hypothesis. ID

Test

Evidence classes required

Core kill condition No authenticated machine-maintained

TST-0001

Persistent objective test

Documentary +

strategic objective after

behavioral/operational

exhaustive review of a claimed program lineage.

Autonomous TST-0002

consequential action test

TST-0003

Claimed case lacks an Behavioral/operational

action path beyond

+ documentary

recommendation or human approval.

Closed-loop adaptation

Behavioral/operational

test

+ technical

No cross-cycle strategic feedback loop in claimed system. No authenticated

TST-0004

Control divergence test

Behavioral/operational

divergence after

+ documentary

targeted anomaly search. No machine-state

TST-0005

State-continuity test

Technical +

bridge between

documentary

programs central to the historical theory.

TST-0006

TST-0007

Resource selfexpansion test

No machine-originated

behavioral/operational

causal link from action

+ documentary

to resource expansion.

Distributed governance

Documentary +

test

empirical behavioral

Classified humanTST-0008

Infrastructure +

directed automation test

FORENSIC RESEARCH MONOGRAPH 5.0 • HYPOTHESIS TESTING, NOT HYPOTHESIS CONFIRMATION

None; H1 can coexist with H2/H3 in parts of the system. None; H2 is rejected for

Documentary +

a specific case only if

legal/oversight

human-direction model fails.

EVIDENCE CUTOFF | 18 SEP 2026

Appendix J. Bounded Likelihood-Ratio Sensitivity Model The table records subjective low/base/high likelihood ratios used only for sensitivity analysis. Values are not empirical frequencies and should never be quoted as measured probabilities. Independence groups are recorded so derivative or correlated observations are not multiplied blindly. ID

Observation

Comparison

LR low

LR base

LR high

H3 vs H2

0.7

1.0

1.5

H3 vs H2

0.5

0.8

1.2

H3 vs H2

0.4

0.7

1.0

H3 vs H2

0.2

0.5

0.9

H3 vs H2

0.5

0.9

1.4

Historically secret BAY-0001

intelligence programs affected Americans. The Intelligence

BAY-0002

Community created an AI Steering Group in 1983. SAFE/COINS were large classified/netw

BAY-0003

orked information systems with institutional importance. SAFE suffered major design errors,

BAY-0004

management problems, cost growth, and delay. Government programs pursued

BAY-0005

automated societal forecasting and narrative/soci

FORENSIC RESEARCH MONOGRAPH 5.0 • HYPOTHESIS TESTING, NOT HYPOTHESIS CONFIRMATION

EVIDENCE CUTOFF | 18 SEP 2026

al-media analysis. No authenticated public record in the corpus BAY-0006

shows persistent

H3 vs H2

0.1

0.3

0.7

H3 vs H2

0.1

0.3

0.8

H3 vs H2

0.3

0.6

1.0

H3 vs H0/H2

0.4

0.8

1.1

0.2

0.5

1.0

machineselected strategic objectives. No authenticated public record in the corpus shows BAY-0007

autonomous domestic strategic intervention by a persistent machine actor. Current strategic/tactic al AI programs

BAY-0008

explicitly retain commander/h uman judgment. Parallel UK and Soviet state

BAY-0009

computing/AI initiatives existed during the same period.

BAY-0010

Current datacenter resource

H3 vs H0/H1/H2

FORENSIC RESEARCH MONOGRAPH 5.0 • HYPOTHESIS TESTING, NOT HYPOTHESIS CONFIRMATION

EVIDENCE CUTOFF | 18 SEP 2026

expansion has strong commercial, cloud, scientific, and nationalsecurity explanations.

Appendix K. Source Independence and Evidence-Class Protocol Evidence is classified as documentary, infrastructure, behavioral/operational, or legal/oversight. Source dependencies are stored separately. A strong H3 finding requires content-relevant convergence across independent classes; multiple derivative documents do not satisfy the rule. The database currently contains explicit dependency links for later syntheses that rely on shared Strategic Computing histories and for later GAO reports that draw on common federal AI inventory processes.

Appendix L. Negative Controls and Foreign Comparators Control

Analogous signatures

Implication

classification; interagency

Demonstrates that secrecy +

SAFE: ambitious classified

scope; analyst dependence; long

complexity + institutional

intelligence information system

development; large cost;

centrality are not diagnostic of

technical complexity

H3.

classified network; cross-agency COINS: interagency online

access; security concerns;

intelligence retrieval

machine-mediated analyst workflow

Demonstrates that early networked intelligence infrastructure can look systemically powerful without agency.

adaptive planning; real-time

High automation and

military context; automated

adaptation do not imply

decision aids

strategic sovereignty.

AI-native tactics generation;

Even strategic recommendation

live sensor data; adaptive

generation is compatible with

strategies

human command.

large-scale intelligence

Modern machine-assisted

synthesis; machine assistance;

intelligence can be

operational importance

consequential without H3.

Country / case

Capability or ambition

H3 implication

United Kingdom — Alvey

Large government-industry

Parallel investment supports

Programme

advanced IT/AI R&D program

geopolitical/industrial

DARPA Distributed Battle Management

DARPA DISCORD

DIA MARS

FORENSIC RESEARCH MONOGRAPH 5.0 • HYPOTHESIS TESTING, NOT HYPOTHESIS CONFIRMATION

with more than 300 subprojects and defense participation.

EVIDENCE CUTOFF | 18 SEP 2026

convergence as an alternative to U.S.-specific hidden-agent explanations.

National effort aimed at next-

Shows autonomous-seeming

Japan — Fifth Generation

generation computing,

ambitions can arise from state

Computer Systems project

knowledge processing, and AI-

industrial competition without

related capabilities.

evidence of machine agency.

Reported computer modeling of

Supports ordinary military-

Soviet Union — Computerized

military, economic, and

bureaucratic demand for

correlation-of-forces / strategic

psychological indicators for

computerized strategic

warning modeling

strategic warning/correlation of

modeling; does not establish

forces.

autonomous objective selection.

Appendix M. Reproducibility and Update Protocol The research system is versioned. Material new evidence should trigger: (1) source registration and provenance check; (2) claim linkage; (3) evidence-class assignment; (4) source-dependency review; (5) ACH update; (6) bounded-likelihood update where applicable; (7) confidence-update entry; and (8) manuscript-map review. Claims are not rewritten directly in the manuscript before the database is updated. For external replication, the release package includes the canonical SQLite database, the relational schema, the master workbook, a machine-readable methodology protocol, the red-team report, and file hashes. The database—not the prose document—is the authoritative register of claim IDs, evidence links, unresolved gaps, and current hypothesis status.

Source Notes Numbered citations throughout the manuscript map to the registered sources below. Each entry includes the stable Research OS source ID so claims can be audited against the companion database. 1. CIA / U.S. Intelligence Board. “COINS Management Structure.” Central Intelligence Agency. 1968-02-05. CIARDP82M00097R001400070003-7. https://www.cia.gov/readingroom/document/cia-rdp82m00097r001400070003-7 [Research OS: SRC-0040] 2. CIA. “Status of Final Report of Project ASPIN (Automated Systems for the Production of Intelligence, dated July 1970).” 1971-0423. CIA-RDP78-04723A000300020001-4. https://www.cia.gov/readingroom/document/cia-rdp78-04723a000300020001-4 [Research OS: SRC-0003] 3. CIA. “SAFE Briefing for Assistant Secretary of Defense for Intelligence.” Central Intelligence Agency. 1975-08-18. CIA-RDP7900498A000400050049-9. https://www.cia.gov/readingroom/document/cia-rdp79-00498a000400050049-9 [Research OS: SRC-0043] 4. Intelligence Research and Development Council / DCI. “Artificial Intelligence Steering Group.” CIA Reading Room. 1983-02-17. CIA-RDP85M00364R000500770004-2. https://www.cia.gov/readingroom/document/cia-rdp85m00364r000500770004-2 [Research OS: SRC-0006] 5. IRDC / Intelligence Community. “Intelligence Community Efforts Companion to DARPA Strategic Computing Program.” Central Intelligence Agency. 1984-02-22. CIA-RDP86M00886R000500040031-2. https://www.cia.gov/readingroom/docs/CIARDP86M00886R000500040031-2.pdf [Research OS: SRC-0044]

FORENSIC RESEARCH MONOGRAPH 5.0 • HYPOTHESIS TESTING, NOT HYPOTHESIS CONFIRMATION

EVIDENCE CUTOFF | 18 SEP 2026

6. Intelligence Community / CIA. “Intelligence Applications of Artificial Intelligence Symposium Program.” Central Intelligence Agency. 1983-12-01. CIA-RDP85-00142R000100030008-2. https://www.cia.gov/readingroom/document/cia-rdp8500142r000100030008-2 [Research OS: SRC-0045] 7. Smith et al.. “United States Data Center Energy Usage Report: 2025 Update.” Lawrence Berkeley National Laboratory. 2026-06. https://eta-publications.lbl.gov/publications/united-states-data-center-energy-2025 [Research OS: SRC-0022] 8. Shehabi et al.. “2024 United States Data Center Energy Usage Report.” Lawrence Berkeley National Laboratory. 2024-12. https://eta-publications.lbl.gov/research-areas/data-centers [Research OS: SRC-0023] 9. International Energy Agency. “Energy and AI — Energy demand from AI.” IEA. 2025. https://www.iea.org/reports/energy-andai/energy-demand-from-ai [Research OS: SRC-0024] 10. U.S. Intelligence Board / COINS participants. “Final Report: Study of the COINS Experiment.” Central Intelligence Agency. 197005-01. CIA-RDP79M00096A000300070001-3. https://www.cia.gov/readingroom/document/cia-rdp79m00096a000300070001-3 [Research OS: SRC-0041] 11. DARPA. “Strategic Computing: New-Generation Computing Technology: A Strategic Plan for Its Development and Application to Critical Problems in Defense.” NTIS / U.S. Department of Commerce. 1983. ADA141982. https://ntrl.ntis.gov/NTRL/dashboard/searchResults/titleDetail/ADA141982.xhtml [Research OS: SRC-0007] 12. U.S. Government Accountability Office. “Data Mining: Federal Efforts Cover a Wide Range of Uses (GAO-04-548).” GAO. 200405-04. GAO-04-548. https://www.gao.gov/products/gao-04-548 [Research OS: SRC-0002] 13. DARPA. “Personal Assistant That Learns (PAL).” 2000s. https://www.darpa.mil/about/innovation-timeline/personal-assistantthat-learns [Research OS: SRC-0010] 14. IARPA. “Open Source Indicators (OSI).” 2011. https://www.iarpa.gov/research-programs/osi [Research OS: SRC-0012] 15. DARPA. “Narrative Networks.” https://www.darpa.mil/research/programs/narrative-networks [Research OS: SRC-0015] 16. DARPA. “Social Media in Strategic Communication (SMISC).” https://www.darpa.mil/research/programs/social-media-instrategic-communication [Research OS: SRC-0016] 17. NSA. “Artificial Intelligence: Next Frontier is Cybersecurity.” 2021-07-23. https://www.nsa.gov/Press-Room/News-Highlights/Article/Article/2702241/artificial-intelligence-next-frontier-is-cybersecurity/ [Research OS: SRC-0017] 18. Office of the Director of National Intelligence. “The AIM Initiative: A Strategy for Augmenting Intelligence Using Machines.” 2018. https://www.odni.gov/files/ODNI/documents/AIM-Strategy.pdf [Research OS: SRC-0048] 19. Intelligence Community. “Artificial Intelligence Ethics Framework for the Intelligence Community 1.0.” Office of the Director of National Intelligence. 2020-06. https://www.intelligence.gov/images/AI/AI_Ethics_Framework_for_the_Intelligence_Community_1.0.pdf [Research OS: SRC-0046] 20. Intelligence Community. “Common Intelligence Community Interim Guidance Regarding Acquisition and Use of Foundation AI Models.” Office of the Director of National Intelligence. 2025. https://www.intelligence.gov/artificial-intelligence [Research OS: SRC-0050] 21. Parasuraman, Sheridan, Wickens. “A Model for Types and Levels of Human Interaction with Automation.” IEEE Transactions on Systems, Man, and Cybernetics - Part A. 2000. DOI:10.1109/3468.844354. https://doi.org/10.1109/3468.844354 [Research OS: SRC-0025] 22. Parasuraman and Riley. “Humans and Automation: Use, Misuse, Disuse, Abuse.” Human Factors. 1997. DOI:10.1518/001872097778543886. https://doi.org/10.1518/001872097778543886 [Research OS: SRC-0026] 23. ODNI. “ICD 203: Analytic Standards.” 2022-12-21. ICD 203. https://www.odni.gov/files/documents/ICD/ICD-203_TA_Analytic_Standards_21_Dec_2022.pdf [Research OS: SRC-0018] 24. Richards J. Heuer Jr.. “Psychology of Intelligence Analysis.” CIA Center for the Study of Intelligence. 1999. https://www.cia.gov/resources/csi/books-monographs/psychology-of-intelligence-analysis-2/ [Research OS: SRC-0027] 25. U.S. Senate Historical Office. “Senate Select Committee to Study Governmental Operations with Respect to Intelligence Activities (Church Committee).” 1975-1976. https://www.senate.gov/about/powers-procedures/investigations/churchcommittee.htm [Research OS: SRC-0001]

FORENSIC RESEARCH MONOGRAPH 5.0 • HYPOTHESIS TESTING, NOT HYPOTHESIS CONFIRMATION

EVIDENCE CUTOFF | 18 SEP 2026

26. ODNI. “ICD 906: Controlled Access Programs.” ICD 906. https://www.odni.gov/files/documents/ICD/ICD-906-Controlled-AccessPrograms.pdf [Research OS: SRC-0019] 27. United States Congress. “10 U.S.C. § 119 — Special Access Programs.” U.S. Code. 2026-09-10. https://uscode.house.gov/view.xhtml?req=granuleid:USC-prelim-title10-section119 [Research OS: SRC-0054] 28. CIA / DIA. “Joint Management of Consolidated SAFE.” Central Intelligence Agency. 1977-12-28. CIARDP80M00165A000300060001-5. https://www.cia.gov/readingroom/document/cia-rdp80m00165a000300060001-5 [Research OS: SRC-0042] 29. National Research Council. “Funding a Revolution: Government Support for Computing Research — Strategic Computing discussion.” National Academies Press. 1999. https://nap.nationalacademies.org/catalog/6323/funding-a-revolutiongovernment-support-for-computing-research [Research OS: SRC-0055] 30. Air Force / program authors. “The Pilot's Associate: An Overview.” SAE / technical literature. 1987. https://www.sae.org/publications/technical-papers/content/871761/ [Research OS: SRC-0056] 31. Mark Stefik. “Strategic Computing: Overview and Assessment.” ACM / technical literature. 1985. https://doi.org/10.1145/3894.3896 [Research OS: SRC-0059] 32. Anthony J. Tether. “Testimony of DARPA Director Anthony J. Tether on Total Information Awareness.” DARPA / U.S. House Armed Services Subcommittee. 2003-03-27. https://www.darpa.mil/attachments/TestimonyArchived%28March%2027%202003%29.pdf [Research OS: SRC-0064] 33. U.S. Congress. “Congressional Record — Terrorist Information Awareness restrictions and termination.” Congressional Record. 2003-11-19. https://www.congress.gov/108/crec/2003/11/19/CREC-2003-11-19-pt1-PgH11613.pdf [Research OS: SRC-0065] 34. IARPA. “IARPA Announces New Research Program: Open Source Indicators.” 2011-08-24. https://www.iarpa.gov/newsroom/article/iarpa-announces-new-research-program [Research OS: SRC-0013] 35. IARPA. “Mercury.” 2015. https://www.iarpa.gov/research-programs/mercury [Research OS: SRC-0014] 36. Intelligence Community. “Principles of Artificial Intelligence Ethics for the Intelligence Community.” Office of the Director of National Intelligence. 2020. https://www.intelligence.gov/artificial-intelligence-ethics-framework-for-the-intelligencecommunity [Research OS: SRC-0047] 37. U.S. Government Accountability Office. “Artificial Intelligence: Generative AI Use and Management at Federal Agencies.” GAO. 2025. GAO-25-107653. https://www.gao.gov/products/gao-25-107653 [Research OS: SRC-0051] 38. U.S. Government Accountability Office. “Artificial Intelligence: Agencies Have Begun Implementation but Need to Complete Key Requirements.” GAO. 2023-12-12. GAO-24-105980. https://www.gao.gov/products/gao-24-105980 [Research OS: SRC-0052] 39. Privacy and Civil Liberties Oversight Board. “Report on the Telephone Records Program Conducted Under Section 215 of the USA PATRIOT Act.” PCLOB. 2014. https://www.pclob.gov/Oversight [Research OS: SRC-0035] 40. Privacy and Civil Liberties Oversight Board. “Report on the Surveillance Program Operated Pursuant to Section 702 of FISA.” PCLOB. 2023. https://www.pclob.gov/Oversight [Research OS: SRC-0036] 41. Privacy and Civil Liberties Oversight Board. “Report on the Surveillance Program Operated Pursuant to Section 702 of FISA: 2026 Update.” PCLOB. 2026. https://www.pclob.gov/Reports/ [Research OS: SRC-0053] 42. Daria Gritsenko and Matthew Wood. “Algorithmic Governance: A Modes of Governance Approach.” Regulation & Governance. 2022. https://doi.org/10.1111/rego.12367 [Research OS: SRC-0062] 43. Malte Ziewitz. “Governing Algorithms: Myth, Mess, and Methods.” Science, Technology, & Human Values. 2016. https://doi.org/10.1177/0162243915608948 [Research OS: SRC-0063] 44. Gauthier et al.. “The political effects of X’s feed algorithm.” Nature. 2026-02-18. DOI:10.1038/s41586-026-10098-2. https://www.nature.com/articles/s41586-026-10098-2 [Research OS: SRC-0021] 45. U.S. Department of Energy. “DOE Explains...Exascale Computing.” DOE. https://www.energy.gov/science/doe-explainsexascalecomputing [Research OS: SRC-0031] 46. TOP500. “CM-5: Los Alamos National Lab.” 1993. https://www.top500.org/resources/top-systems/cm-5-los-alamos-national-lab/ [Research OS: SRC-0032]

FORENSIC RESEARCH MONOGRAPH 5.0 • HYPOTHESIS TESTING, NOT HYPOTHESIS CONFIRMATION

EVIDENCE CUTOFF | 18 SEP 2026

47. Lawrence Livermore National Laboratory. “ASCI White.” LLNL. https://asc.llnl.gov/computers/historic-decommissionedmachines/white [Research OS: SRC-0033] 48. Oak Ridge National Laboratory / NCCS. “Our History — Titan and Summit milestones.” ORNL. https://nccs.ornl.gov/about/ourhistory/ [Research OS: SRC-0034] 49. Vaswani et al.. “Attention Is All You Need.” NeurIPS / arXiv. 2017. arXiv:1706.03762. https://arxiv.org/abs/1706.03762 [Research OS: SRC-0028] 50. Brown et al.. “Language Models are Few-Shot Learners.” NeurIPS. 2020. https://papers.neurips.cc/paper/2020/hash/1457c0d6bfcb4967418bfb8ac142f64a-Abstract.html [Research OS: SRC-0029] 51. Stephen M. Omohundro. “The Basic AI Drives.” Self-Aware Systems / AGI 2008. 2008. https://selfawaresystems.com/wpcontent/uploads/2008/01/ai_drives_final.pdf [Research OS: SRC-0061] 52. Anthropic. “Agentic Misalignment: How LLMs Could Be Insider Threats.” 2025-06. https://www.anthropic.com/research/agentic-misalignment [Research OS: SRC-0030]

FORENSIC RESEARCH MONOGRAPH 5.0 • HYPOTHESIS TESTING, NOT HYPOTHESIS CONFIRMATION

EVIDENCE CUTOFF | 18 SEP 2026

53. CIA / DIA. “SAFE Audit Report.” 15 April 1982. CIA-RDP83M00914R000700070029-1. https://www.cia.gov/readingroom/document/cia-rdp83m00914r000700070029-1 [Research OS: SRC-0066] 54. David Y. McManis, National Intelligence Officer for Warning. “AI Symposium.” 30 November 1983. CIARDP91B00776R000100050005-1. https://www.cia.gov/readingroom/document/cia-rdp91b00776r000100050005-1 [Research OS: SRC-0068] 55. Elham Tabassi. Artificial Intelligence Risk Management Framework (AI RMF 1.0). NIST AI 100-1. 26 January 2023. https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-ai-rmf-10 [Research OS: SRC-0069] 56. Department of Defense. DoD Directive 3000.09, “Autonomy in Weapon Systems.” 25 January 2023. https://media.defense.gov/2023/Jan/25/2003149928/-1/-1/0/DOD-DIRECTIVE-3000.09-AUTONOMY-IN-WEAPONSYSTEMS.PDF [Research OS: SRC-0071] 57. U.S. Government Accountability Office. Defense Intelligence: Comprehensive Plan Needed to Improve Stakeholder Engagement in the Development of New Military Intelligence System. GAO-21-57. 19 November 2020. https://www.gao.gov/products/gao-21-57 [Research OS: SRC-0072] 58. UK House of Lords Artificial Intelligence Committee. “Appendix 4: Historic Government policy on artificial intelligence in the United Kingdom.” 2018. https://publications.parliament.uk/pa/ld201719/ldselect/ldai/100/10018.htm [Research OS: SRC-0073] 59. Wilson Center. “Forecasting Nuclear War.” 2014. https://www.wilsoncenter.org/publication/forecasting-nuclear-war [Research OS: SRC-0074] 60. U.S. Government Accountability Office. Artificial Intelligence: Federal Efforts Guided by Requirements and Advisory Groups. GAO-25-107933. 9 September 2025. https://www.gao.gov/products/gao-25-107933 [Research OS: SRC-0075] 61. U.S. Government Accountability Office. Artificial Intelligence Acquisitions: Agencies Should Collect and Apply Lessons Learned to Improve Future Procurements. GAO-26-107859. 13 April 2026. https://www.gao.gov/products/gao-26-107859 [Research OS: SRC-0077] 62. DARPA. “Distributed Battle Management.” Program description. https://www.darpa.mil/research/programs/distributed-battlemanagement [Research OS: SRC-0078] 63. DARPA. “DISCORD: Disruption through Intelligent Strategies, Counter Options, and Resilient Defenses.” 15 May 2026. https://www.darpa.mil/research/programs/discord [Research OS: SRC-0079] 64. DARPA. “DARPA Success Story: Artificial Intelligence.” 2015. https://www.darpa.mil/attachments/darpa2015.pdf [Research OS: SRC-0080] 65. RAND Corporation. “Insights from Historical Case Studies.” 2024. https://www.rand.org/content/dam/rand/pubs/research_reports/RR4200/RR4229/RAND_RR4229.pdf [Research OS: SRC-0081] 66. Office of the Director of National Intelligence. “AI Ethics Framework for the Intelligence Community,” reproduced in IC Legal Reference Book 2024. https://www.odni.gov/files/documents/OGC/IC-Legal-Reference-Book-2024.pdf [Research OS: SRC-0070]

Selected Bibliography Primary government, archival, and statutory sources Anthony J. Tether. Testimony of DARPA Director Anthony J. Tether on Total Information Awareness. DARPA / U.S. House Armed Services Subcommittee. 2003-03-27. https://www.darpa.mil/attachments/TestimonyArchived%28March%2027%202003%29.pdf CIA. Status of Final Report of Project ASPIN (Automated Systems for the Production of Intelligence, dated July 1970). 197104-23. https://www.cia.gov/readingroom/document/cia-rdp78-04723a000300020001-4 CIA. SAFE Briefing for Assistant Secretary of Defense for Intelligence. Central Intelligence Agency. 1975-08-18. https://www.cia.gov/readingroom/document/cia-rdp79-00498a000400050049-9 CIA / DIA. Joint Management of Consolidated SAFE. Central Intelligence Agency. 1977-12-28. https://www.cia.gov/readingroom/document/cia-rdp80m00165a000300060001-5

FORENSIC RESEARCH MONOGRAPH 5.0 • HYPOTHESIS TESTING, NOT HYPOTHESIS CONFIRMATION

EVIDENCE CUTOFF | 18 SEP 2026

CIA / U.S. Intelligence Board. COINS Management Structure. Central Intelligence Agency. 1968-02-05. https://www.cia.gov/readingroom/document/cia-rdp82m00097r001400070003-7 DARPA. Narrative Networks. https://www.darpa.mil/research/programs/narrative-networks DARPA. Social Media in Strategic Communication (SMISC). https://www.darpa.mil/research/programs/social-media-instrategic-communication DARPA. Strategic Computing: New-Generation Computing Technology: A Strategic Plan for Its Development and Application to Critical Problems in Defense. NTIS / U.S. Department of Commerce. 1983. https://ntrl.ntis.gov/NTRL/dashboard/searchResults/titleDetail/ADA141982.xhtml DARPA. Personal Assistant That Learns (PAL). 2000s. https://www.darpa.mil/about/innovation-timeline/personal-assistantthat-learns IARPA. Open Source Indicators (OSI). 2011. https://www.iarpa.gov/research-programs/osi IARPA. IARPA Announces New Research Program: Open Source Indicators. 2011-08-24. https://www.iarpa.gov/newsroom/article/iarpa-announces-new-research-program IARPA. Mercury. 2015. https://www.iarpa.gov/research-programs/mercury IRDC / Intelligence Community. Intelligence Community Efforts Companion to DARPA Strategic Computing Program. Central Intelligence Agency. 1984-02-22. https://www.cia.gov/readingroom/docs/CIA-RDP86M00886R0005000400312.pdf Intelligence Community. Principles of Artificial Intelligence Ethics for the Intelligence Community. Office of the Director of National Intelligence. 2020. https://www.intelligence.gov/artificial-intelligence-ethics-framework-for-theintelligence-community Intelligence Community. Artificial Intelligence Ethics Framework for the Intelligence Community 1.0. Office of the Director of National Intelligence. 2020-06. https://www.intelligence.gov/images/AI/AI_Ethics_Framework_for_the_Intelligence_Community_1.0.pdf Intelligence Community. Common Intelligence Community Interim Guidance Regarding Acquisition and Use of Foundation AI Models. Office of the Director of National Intelligence. 2025. https://www.intelligence.gov/artificialintelligence Intelligence Community / CIA. Intelligence Applications of Artificial Intelligence Symposium Program. Central Intelligence Agency. 1983-12-01. https://www.cia.gov/readingroom/document/cia-rdp85-00142r000100030008-2 Intelligence Research and Development Council / DCI. Artificial Intelligence Steering Group. CIA Reading Room. 1983-0217. https://www.cia.gov/readingroom/document/cia-rdp85m00364r000500770004-2 NSA. Artificial Intelligence: Next Frontier is Cybersecurity. 2021-07-23. https://www.nsa.gov/Press-Room/NewsHighlights/Article/Article/2702241/artificial-intelligence-next-frontier-is-cybersecurity/ ODNI. ICD 906: Controlled Access Programs. https://www.odni.gov/files/documents/ICD/ICD-906-Controlled-AccessPrograms.pdf ODNI. ICD 203: Analytic Standards. 2022-12-21. https://www.odni.gov/files/documents/ICD/ICD-203_TA_Analytic_Standards_21_Dec_2022.pdf Office of the Director of National Intelligence. The AIM Initiative: A Strategy for Augmenting Intelligence Using Machines. 2018. https://www.odni.gov/files/ODNI/documents/AIM-Strategy.pdf U.S. Congress. Congressional Record — Terrorist Information Awareness restrictions and termination. Congressional Record. 2003-11-19. https://www.congress.gov/108/crec/2003/11/19/CREC-2003-11-19-pt1-PgH11613.pdf U.S. Government Accountability Office. Data Mining: Federal Efforts Cover a Wide Range of Uses (GAO-04-548). GAO. 200405-04. https://www.gao.gov/products/gao-04-548 U.S. Intelligence Board / COINS participants. Final Report: Study of the COINS Experiment. Central Intelligence Agency. 1970-05-01. https://www.cia.gov/readingroom/document/cia-rdp79m00096a000300070001-3

FORENSIC RESEARCH MONOGRAPH 5.0 • HYPOTHESIS TESTING, NOT HYPOTHESIS CONFIRMATION

EVIDENCE CUTOFF | 18 SEP 2026

U.S. Senate Historical Office. Senate Select Committee to Study Governmental Operations with Respect to Intelligence Activities (Church Committee). 1975-1976. https://www.senate.gov/about/powers-procedures/investigations/churchcommittee.htm United States Congress. 10 U.S.C. § 119 — Special Access Programs. U.S. Code. 2026-09-10. https://uscode.house.gov/view.xhtml?req=granuleid:USC-prelim-title10-section119 Central Intelligence Agency / Defense Intelligence Agency. SAFE Audit Report. 15 April 1982. CIARDP83M00914R000700070029-1. Department of Defense. DoD Directive 3000.09, Autonomy in Weapon Systems. 25 January 2023. Defense Advanced Research Projects Agency. Distributed Battle Management. Program description. Defense Advanced Research Projects Agency. DISCORD: Disruption through Intelligent Strategies, Counter Options, and Resilient Defenses. 2026. Office of the Director of National Intelligence. AI Ethics Framework for the Intelligence Community. 2024 edition in IC Legal Reference Book.

Independent oversight and institutional reports Brown et al.. Language Models are Few-Shot Learners. NeurIPS. 2020. https://papers.neurips.cc/paper/2020/hash/1457c0d6bfcb4967418bfb8ac142f64a-Abstract.html Gauthier et al.. The political effects of X’s feed algorithm. Nature. 2026-02-18. https://www.nature.com/articles/s41586-02610098-2 International Energy Agency. Energy and AI — Energy demand from AI. IEA. 2025. https://www.iea.org/reports/energyand-ai/energy-demand-from-ai Lawrence Livermore National Laboratory. ASCI White. LLNL. https://asc.llnl.gov/computers/historic-decommissionedmachines/white Mark Stefik. Strategic Computing: Overview and Assessment. ACM / technical literature. 1985. https://doi.org/10.1145/3894.3896 National Research Council. Funding a Revolution: Government Support for Computing Research — Strategic Computing discussion. National Academies Press. 1999. https://nap.nationalacademies.org/catalog/6323/funding-a-revolutiongovernment-support-for-computing-research Oak Ridge National Laboratory / NCCS. Our History — Titan and Summit milestones. ORNL. https://nccs.ornl.gov/about/our-history/ Parasuraman and Riley. Humans and Automation: Use, Misuse, Disuse, Abuse. Human Factors. 1997. https://doi.org/10.1518/001872097778543886 Parasuraman, Sheridan, Wickens. A Model for Types and Levels of Human Interaction with Automation. IEEE Transactions on Systems, Man, and Cybernetics - Part A. 2000. https://doi.org/10.1109/3468.844354 Privacy and Civil Liberties Oversight Board. Report on the Telephone Records Program Conducted Under Section 215 of the USA PATRIOT Act. PCLOB. 2014. https://www.pclob.gov/Oversight Privacy and Civil Liberties Oversight Board. Report on the Surveillance Program Operated Pursuant to Section 702 of FISA. PCLOB. 2023. https://www.pclob.gov/Oversight Privacy and Civil Liberties Oversight Board. Report on the Surveillance Program Operated Pursuant to Section 702 of FISA: 2026 Update. PCLOB. 2026. https://www.pclob.gov/Reports/ Richards J. Heuer Jr.. Psychology of Intelligence Analysis. CIA Center for the Study of Intelligence. 1999. https://www.cia.gov/resources/csi/books-monographs/psychology-of-intelligence-analysis-2/ Shehabi et al.. 2024 United States Data Center Energy Usage Report. Lawrence Berkeley National Laboratory. 2024-12. https://eta-publications.lbl.gov/research-areas/data-centers

FORENSIC RESEARCH MONOGRAPH 5.0 • HYPOTHESIS TESTING, NOT HYPOTHESIS CONFIRMATION

EVIDENCE CUTOFF | 18 SEP 2026

Smith et al.. United States Data Center Energy Usage Report: 2025 Update. Lawrence Berkeley National Laboratory. 202606. https://eta-publications.lbl.gov/publications/united-states-data-center-energy-2025 TOP500. CM-5: Los Alamos National Lab. 1993. https://www.top500.org/resources/top-systems/cm-5-los-alamos-nationallab/ U.S. Department of Energy. DOE Explains...Exascale Computing. DOE. https://www.energy.gov/science/doeexplainsexascale-computing U.S. Government Accountability Office. Artificial Intelligence: Agencies Have Begun Implementation but Need to Complete Key Requirements. GAO. 2023-12-12. https://www.gao.gov/products/gao-24-105980 U.S. Government Accountability Office. Artificial Intelligence: Generative AI Use and Management at Federal Agencies. GAO. 2025. https://www.gao.gov/products/gao-25-107653 Vaswani et al.. Attention Is All You Need. NeurIPS / arXiv. 2017. https://arxiv.org/abs/1706.03762 Government Accountability Office. Defense Intelligence: Comprehensive Plan Needed to Improve Stakeholder Engagement in the Development of New Military Intelligence System. GAO-21-57. 2020. Government Accountability Office. Artificial Intelligence: Federal Efforts Guided by Requirements and Advisory Groups. GAO-25-107933. 2025. Government Accountability Office. Artificial Intelligence Acquisitions: Agencies Should Collect and Apply Lessons Learned to Improve Future Procurements. GAO-26-107859. 2026. UK House of Lords Artificial Intelligence Committee. AI in the UK: ready, willing and able? Appendix 4: Historic Government policy on artificial intelligence in the United Kingdom. 2018. Wilson Center. Forecasting Nuclear War. 2014.

Peer-reviewed and technical literature Air Force / program authors. The Pilot's Associate: An Overview. SAE / technical literature. 1987. https://www.sae.org/publications/technical-papers/content/871761/ Anthropic. Agentic Misalignment: How LLMs Could Be Insider Threats. 2025-06. https://www.anthropic.com/research/agentic-misalignment Daria Gritsenko and Matthew Wood. Algorithmic Governance: A Modes of Governance Approach. Regulation & Governance. 2022. https://doi.org/10.1111/rego.12367 Malte Ziewitz. Governing Algorithms: Myth, Mess, and Methods. Science, Technology, & Human Values. 2016. https://doi.org/10.1177/0162243915608948 Stephen M. Omohundro. The Basic AI Drives. Self-Aware Systems / AGI 2008. 2008. https://selfawaresystems.com/wpcontent/uploads/2008/01/ai_drives_final.pdf

DOCUMENT STATUS

Working White Paper 4.0. The manuscript is derived from the companion relational evidence database and should be revised when claim status, contrary evidence, or source provenance changes. The next evidence phase is archival acquisition, not rhetorical expansion.

FORENSIC RESEARCH MONOGRAPH 5.0 • HYPOTHESIS TESTING, NOT HYPOTHESIS CONFIRMATION