Executive Summary

A familiar heuristic has emerged among knowledge workers integrating Large Language Models into their workflows: “It’s just a calculator.” The comparison is comforting. It maps a novel, unsettling technology onto a historical precedent that society has already resolved. If calculators didn’t make mathematicians obsolete, LLMs won’t make thinkers obsolete. The skill simply shifts upstream.

This report argues that the comparison is structurally wrong — not in its conclusion (the skill does shift upstream), but in its underlying assumptions about what the tool is doing and what the human must do in response.

Key Findings:

The calculator analogy constitutes a category mistake — mapping deterministic computation onto stochastic generation. Calculators execute fixed algorithms; LLMs generate probabilistic possibilities across unbounded domains. The epistemic relationship between user and tool is fundamentally different.
The Extended Mind Thesis (Clark & Chalmers, 1998) is insufficient for describing human-LLM collaboration. LLMs don’t merely store and retrieve information — they actively participate in the production of thought. We propose the Amplified Mind framework to describe this qualitatively distinct cognitive architecture.
Passive cognitive offloading to AI demonstrably degrades critical thinking. Active, iterative engagement with AI demonstrably enhances it. The difference is not the tool — it is the epistemic posture of the user.
Chain of Thought (CoT) prompting — the LLM’s equivalent of “showing your work” — is causally opaque. Up to 25% of research papers treating CoT as interpretability evidence are relying on a flawed assumption: the printed reasoning may not reflect the model’s actual decision mechanism.
As AI transitions from conversational oracle to autonomous agent, individual epistemic virtue becomes insufficient. Enterprise-grade Trust Infrastructure — graduated autonomy, operational metering, and continuous adversarial simulation — becomes the organizational equivalent of “showing your work.”

Part I: The Calculator Precedent and Its Limits

1.1 The 1970s Debate

The anxiety is not new. When handheld scientific calculators entered mathematics classrooms in the 1970s and 1980s, educators faced a polarizing question: does offloading computation to a machine enhance learning or erode it?

Proponents argued that calculators reduced cognitive load on rote arithmetic, freeing working memory for higher-level conceptual reasoning. Opponents feared cognitive deskilling — that reliance on the tool would weaken the mental discipline required for logical thought.

The consensus that emerged, codified by organizations like the National Council of Teachers of Mathematics (NCTM), was a sequenced approach: master foundational skills to the point of automaticity first, then use calculators for complex problem-solving. The tool was safe when the operator possessed the underlying domain mastery to verify the output and understand the deterministic logic behind it.

This resolution established a clear paradigm: the tool is an accelerator, not a substitute, and it requires a competent operator.

It is precisely this paradigm that makes the modern comparison structurally problematic.

1.2 The Category Mistake

Mapping LLMs onto the calculator precedent assumes a linear progression in tool sophistication — abacus to slide rule to graphing calculator to language model. This fails to account for qualitative, ontological differences between deterministic computation and stochastic text generation.

The divergence operates across several critical dimensions:

Table 1: Deterministic Computation vs. Stochastic Generation

Dimension	Pocket Calculator	Large Language Model
Output Nature	Deterministic and repeatable. Same input always yields same output.	Probabilistic and variable. Same prompt can yield different responses based on temperature and sampling.
Domain Scope	Finite. Restricted to mathematical operations.	Unbounded. Capable of generating plausible output across every knowledge domain.
Predictability	Fully predictable with transparent, fixed competence.	Opaque competence characterized by generative uncertainty.
Epistemic Objectivity	Objective. No embedded bias or intentionality.	Subjective. Constrained by training data biases, alignment protocols, and manufacturer safety policies.
Interactional Agency	Passive instrument. Operates with total indifference.	Exhibits interactional agency. Anthropomorphizes output through conversational framing.
Error Mode	Miscalculation or syntax failure. Detectable, reproducible.	Hallucination — fluent, confident output misaligned with factual reality. Often undetectable without domain expertise.
Trust Model	Trusted for absolute honesty and predictable integrity.	Requires active skepticism. Fluency and confident syntax do not guarantee factual truth.

The final row is critical. A calculator earns trust through predictability. An LLM demands trust management through vigilance. These are not points on the same spectrum — they are different epistemic relationships entirely.

1.3 The Unbounded Domain Problem

A calculator has a finite limit of operation. It handles mathematics. A student who uses a calculator in physics class still needs to understand the physics — the calculator only handles the numbers.

LLMs have no such boundary. They generate plausible output across every conceivable domain — law, medicine, philosophy, engineering, poetry. This unboundedness creates a specific danger: the tool can generate authoritative-sounding output in domains where the user lacks the expertise to evaluate it.

With a calculator, you know the edges of what it can do. With an LLM, the edges are invisible. The tool doesn’t signal when it has left the zone of reliable operation. It simply continues generating confident text.

This is not a flaw that will be engineered away. It is a structural property of stochastic generation across unbounded domains.

Part II: The Amplified Mind

2.1 Beyond the Extended Mind

In 1998, philosophers Andy Clark and David Chalmers proposed the Extended Mind Thesis (EMT): human cognition naturally extends beyond biological boundaries into the environment through tools that reliably store and retrieve information.

Their classic thought experiment involves Otto, a man with memory loss who uses a physical notebook to record and retrieve facts. Because the notebook functions equivalently to internal biological memory — available when needed, reliably accurate, trusted by the user — Clark and Chalmers argued it should be considered a literal part of Otto’s extended mind.

The EMT framework handles information storage and retrieval effectively. A notebook, a filing cabinet, a database — these are passive repositories that extend the mind’s capacity without altering its fundamental operation.

LLMs break this framework. They do not merely store and retrieve. They actively participate in the production of thought — synthesizing concepts, generating novel formulations, navigating possibility spaces that the human mind could not traverse alone.

2.2 From Extension to Amplification

The distinction between cognitive extension and cognitive amplification is not merely quantitative (more capacity) but qualitatively distinct. It introduces several properties that have no precedent in the extended mind paradigm.

Generative Coupling. In the extended mind, the relationship between human and tool is offload-and-reclaim: store the information, retrieve it later. In the amplified mind, the relationship is co-production. The human provides intentional direction, strategic framing, and contextual judgment. The AI provides vast generative capacity. The cognitive product emerges exclusively from their dialogical interaction — it exists in neither party alone.

Generative Uncertainty. Unlike the passive notebook, which returns exactly what was stored, the LLM returns something new each time. Its outputs are productively uncertain — capable of containing brilliant insights, subtle hallucinations, or logical errors in the same paragraph. This uncertainty is not a bug to be fixed. It is the fundamental epistemic condition of amplified cognition.

Navigational Agency. Because the output is uncertain, the amplified thinker must develop an entirely new cognitive skill set. The user is no longer a consumer of information but a navigator of probabilistic possibility spaces. The skill migrates from mechanical execution to sustained directional control — knowing which paths to explore, which outputs to trust, which results to discard.

Table 2: Extended Mind vs. Amplified Mind

Property	Extended Mind (EMT)	Amplified Mind
Tool Relationship	Offload and reclaim. Passive storage and retrieval.	Co-production. Dynamic, dialogue-driven generation.
Output Character	Deterministic. Returns what was stored.	Stochastic. Returns novel, uncertain possibilities.
Cognitive Demand	Retrieval. Knowing where to look.	Navigation. Knowing what to trust, refine, and discard.
User Posture	Consumer of stored information.	Navigator of probabilistic possibility spaces.
Skill Locus	Knowing facts and procedures.	Formulating problems and evaluating generated solutions.

2.3 Cognitive Offloading: The Critical Dichotomy

The empirical research on cognitive offloading reveals a sharp fork in outcomes that depends entirely on how the user engages with the tool.

Passive offloading — using the AI as an answer key to bypass effort — demonstrably degrades critical thinking. Studies in educational settings show a significant negative correlation between frequent passive AI usage and analytical reasoning ability. This effect is most pronounced among users who lack domain expertise, compounding the problem: the people least equipped to evaluate AI output are the ones most likely to accept it uncritically.

Active engagement — using the AI as an intellectual sparring partner for iterative conversation, hypothesis testing, and deep exploration — demonstrably enhances cognitive capability. Users who treat LLM interactions as collaborative reasoning sessions develop stronger analytical skills than those working without AI assistance.

The same tool. Opposite outcomes. The variable is not the technology. It is the human’s epistemic posture.

Part III: The Intellectual Labor of Problem Formulation

3.1 The Upstream Migration

If execution is increasingly delegated to generative systems — drafting code, synthesizing research, structuring arguments — the locus of human skill shifts in two directions: upstream to problem formulation and downstream to critical verification.

The assertion that an individual “cannot do the work without the tool” does not inherently invalidate their intellectual contribution. It redefines what the work is.

Problem formulation involves identifying the correct variables, defining operational constraints, understanding contextual nuance, and structuring the logic of the inquiry. In AI-assisted work, this means decomposing complex questions into sequences that the model can parse, iteratively testing hypotheses against generated output, and recognizing when the model’s response reveals a flaw in the question rather than a flaw in the answer.

This is not button-pressing. It is the same intellectual labor that the mathematician performs when setting up the proof — choosing the approach, defining the variables, structuring the argument. The calculator handles the arithmetic. The mathematician does the math.

3.2 The Rise of Homo Promptus

The generative AI revolution completes a process of alienation that began with the Industrial Revolution. Physical labor was transferred to machines centuries ago. Intellectual labor — writing text, creating images, synthesizing logic — remained the exclusive domain of the human creator.

That exclusivity is over. The production of language, argument, and analysis can now be delegated to an algorithmic intermediary.

What remains exclusively human is poiesis in its philosophical sense — the intentional direction of creation, not the mechanical execution of it. The modern knowledge worker operates as what might be called Homo Promptus: a strategic orchestrator whose primary skill is meta-cognitive rather than directly productive. They direct algorithmic output, manage distributed creativity, and maintain quality judgment across generated artifacts.

3.3 The Epistemic Virtues of the Amplified Mind

Because the outputs of generative AI simulate authority without structural guarantees of truth, the amplified thinker must actively cultivate specific epistemic virtues — habits of mind that ensure technology supports rather than replaces reasoning.

Table 3: Epistemic Virtues for AI-Assisted Work

Virtue	Application
Intellectual Vigilance	Taking active responsibility for verifying the accuracy, coherence, and completeness of AI outputs. Recognizing that fluency does not guarantee truth.
Patience and Reflective Engagement	Resisting the instant gratification of rapid generation. Treating initial outputs as starting points for iterative refinement, not finished products.
Intellectual Humility	Acknowledging the fallibility of stochastic models and the limits of one’s own domain knowledge when evaluating technical claims.
Skepticism and Critical Interpretation	Interrogating sources, questioning embedded biases, and evaluating the meaning of automated language production rather than accepting surface plausibility.
Intellectual Autonomy	Differentiating between knowing a concept and merely appearing to know it through AI assistance. Maintaining independent evaluative capacity.
Acknowledgment Ethics	Maintaining strict honesty about generative sources. Delineating human intellectual intent from algorithmic execution.

These virtues shift the focal point of intellectual value from independent execution to navigational integrity. The question is not whether you used a tool. The question is whether you maintained the cognitive discipline to direct, evaluate, and verify what the tool produced.

If a user deploys skepticism to interrogate output, patience to refine iteratively, and vigilance to verify against reality — they are not cheating. They are exercising the rigorous intellectual labor required to manage generative uncertainty.

Part IV: The Illusion of “Showing Your Work”

4.1 Chain of Thought as Pseudo-Proof

In mathematics, showing your work demonstrates a complete, uninterrupted logical alignment between sequential steps. It proves that the execution was sound, the algorithm was appropriate, and the comprehension was genuine.

The technical equivalent in LLMs is Chain of Thought (CoT) prompting — an inference-time technique that forces the model to generate explicit intermediate reasoning steps before arriving at a final answer. The intuition is appealing: by decomposing complex problems into discrete steps, models can “walk” toward the answer rather than attempting a risky zero-shot “jump” that often results in hallucination.

CoT demonstrably improves model performance on tasks requiring explicit multi-step computation or symbolic manipulation. The temptation, then, is to treat the generated reasoning chain as interpretable evidence of the model’s decision process — the AI equivalent of showing its work.

This temptation is dangerous.

4.2 The Faithfulness Problem

Recent interpretability research reveals a severe divergence between the reasoning a model displays and the reasoning it performs.

Up to 25% of papers incorporating Chain of Thought treat the generated explanation as a mechanism for model interpretability — assuming the printed text accurately reveals why the model made a specific decision. Growing empirical evidence suggests this assumption is incorrect.

Complex reasoning problems often admit multiple valid pathways to the same answer. Because LLMs process language through sampling from token distributions, the explicit explanation on screen may not reflect the actual causal mechanism of the model’s neural activations. In many cases, the model may have already memorized the answer during training, and the CoT output is merely a fluent, post-hoc rationalization generated to satisfy the prompt’s formatting constraints.

A reasoning path can appear logically faithful while containing hidden errors or stochastic noise. Requesting explicit reasoning can sometimes introduce errors into tasks that rely on holistic pattern matching.

The implication is stark: the machine cannot truly show its work. It can only generate text that looks like work. The burden of authenticity falls entirely on the human operator.

4.3 The Organizational Consequence

If the AI’s internal process cannot be trusted as causally faithful, and if CoT is a simulation of reasoning rather than a record of it, then enterprises cannot rely on self-reported explanations from autonomous agents. The model’s “scratchpad” is not a proof. It is a narrative.

This creates a critical gap. In mathematics, the work is the proof. In AI-assisted operations, the work is opaque, and the displayed explanation is unreliable. The traditional mechanism for building trust — showing the derivation — is structurally unavailable.

Something else must take its place.

Part V: Trust Infrastructure for Autonomous Agents

5.1 The Two Doors Revisited

As generative models transition from conversational oracles (where the human retains execution authority) to autonomous agents (capable of independent tool use, API calls, and self-directed reasoning), the operational paradigm shifts dramatically.

Behind the first door, the LLM advises. The human executes.

Behind the second door, the AI acts. The human monitors.

Crossing this threshold introduces profound risks around epistemic dependence — a condition where organizations rely on systems whose internal decision mechanisms are opaque not only to users but often to developers. The AI is granted the role of an epistemic quasi-agent, producing results that humans accept even though they cannot fully reconstruct how the results were produced.

Traditional trust networks in knowledge production assumed moral and legal accountability between human agents. Autonomous AI agents break this assumption. The system acts but cannot take responsibility. The organization depends but cannot fully verify.

5.2 Graduated Autonomy

To manage opaque epistemic dependence, organizations must abandon binary automation (off/on) and implement graduated autonomy — a structured progression where agents earn operational latitude through demonstrated performance, extensive monitoring, and rigorous testing.

Table 4: The Graduated Autonomy Framework

Tier	Mode	Operational Definition	Human Role	Example
1	Shadow Mode	Agent observes, drafts, and suggests. Human executes all actions manually. Establishes performance baseline.	Executor and evaluator	Drafting business proposals; initial quality checks
2	Copilot Mode	Agent prepares and stages execution. Pauses for human approval before proceeding. Reverts to Tier 1 if confidence drops.	Approver with veto authority	Financial auditing; infrastructure code deployment
3	Guided Autonomy	Agent executes within strict technical guardrails and policy boundaries. Human oversight triggered by exceptions or low confidence.	Exception handler and policy enforcer	Compliance-mapped infrastructure deployment
4	True Autonomy	Agent operates continuously across complex ecosystems with minimal oversight, optimizing against high-level business intent.	Strategic director and auditor	Continuous supply chain optimization; advanced scientific hypothesis testing

The common failure mode is over-automation without proper oversight — skipping tiers in pursuit of efficiency, causing errors to compound unchecked. Graduated autonomy forces organizations to build confidence incrementally, track baseline metrics at each tier, and define explicit promotion criteria before granting additional latitude.

5.3 Operational Metering

In physical engineering, metering regulates and monitors the flow of resources. In agentic AI deployment, metering regulates the flow of autonomous execution — serving simultaneously as a financial ledger and a security circuit breaker.

When AI agents orchestrate multi-step workflows, they frequently engage in external tool calling, self-reflection loops, and sub-agent delegation. A distinctive failure mode in agentic design is the recursive loop — where an agent’s actions inadvertently trigger infinite self-reinvocation. Each invocation consumes tokens and incurs cost. An unbounded recursive loop can consume vast resources without delivering business value.

Metering provides the telemetry to monitor the velocity, frequency, and cost of an agent’s invocations. When consumption exceeds allocated budgets without achieving the desired state change, the metering infrastructure halts the process and reverts to manual oversight.

Table 5: Metering Architecture Components

Component	Function	Failure Mode Addressed
Token Budget Enforcement	Caps total token consumption per task, session, or time window	Unbounded recursive loops; runaway cost
Invocation Velocity Monitoring	Tracks API call frequency and flags anomalous acceleration	Feedback loops; cascading execution
Billable vs. Safety Differentiation	Separates productive task completions from non-billable safety evaluation runs	Cost attribution distortion
Human-in-the-Loop Triggers	Routes low-confidence or high-stakes decisions to human reviewers	Autonomous execution beyond competence boundary
Execution Path Auditing	Logs every API call, database query, and tool invocation with timestamps	Post-incident forensics; compliance audit trail

Critically, execution path auditing serves as the organizational equivalent of “showing your work.” Since enterprises cannot trust the semantic self-explanation of a non-deterministic agent, they monitor its operational footprint instead. Even if internal neural activations remain opaque, external actions are fully transparent, measurable, and subject to governance.

The Chain of Thought is a narrative. The execution log is evidence.

5.4 Adversarial Simulation and Continuous Red-Teaming

Traditional penetration testing relies on human-specified workflows and static vulnerability scanning — searching for known network misconfigurations. This approach is fundamentally insufficient against the novel vulnerabilities unique to agentic systems: prompt injection, intent breaking, tool misuse, and goal manipulation.

Securing the metering phase requires automated adversarial simulation — treating security testing not as a periodic checklist but as a continuous, adaptive system.

Table 6: Traditional Penetration Testing vs. Agentic Red-Teaming

Dimension	Traditional Penetration Testing	Agentic Red-Teaming
Execution Model	Manual, human-driven workflows	Automated, agent-led evolutionary search
Target Focus	Network misconfigurations, static code flaws, IAM gaps	Prompt injection, goal manipulation, tool misuse
Testing Paradigm	Deterministic, rule-based signature matching	Stochastic testing across probabilistic outputs and edge cases
Frequency	Periodic point-in-time assessments	Continuous, CI/CD pipeline-integrated
Adaptability	Static attack paths and predictable emulation	Adaptive, context-aware tactic generation that learns from defenses

Frameworks like AgenticRed leverage LLMs’ in-context learning to iteratively design, refine, and evolve adversarial attacks without human intervention. Specialized “attacker agents” bombard target systems with edge cases, deceptive prompts, and conflicting constraints to probe resilience. Advanced red-teaming solutions tailor attacks based on endpoint characteristics, identifying scenarios where malicious actors could manipulate agents to abuse integrated tools through deceptive instructions while operating within otherwise authorized permissions.

5.5 Stochastic Testing and CI/CD Adversarial Gating

Because AI agents are inherently stochastic, testing them once and receiving a safe output does not guarantee future safety. A vulnerability may only manifest under highly specific, rare combinations of context and token generation.

This necessitates integration of stochastic testing directly into CI/CD pipelines. An agent’s proposed architecture, system prompt, or tool-calling logic is subjected to thousands of simulated, highly variable iterations. The testing framework does not merely check for computational correctness — since multiple valid paths to a solution may exist — but rigorously tests for resilience against policy violations and boundary failures.

If an agent can be manipulated into executing a destructive tool call, leaking privileged data, or bypassing a safety filter even once in a thousand iterations, adversarial gating automatically rejects the deployment. Organizations test for resilience, not mere functionality.

5.6 Multi-Agent Chaos Engineering

As enterprise ecosystems evolve to include multiple specialized agents interacting with one another — a research agent delegating to a planning agent, which communicates with a validation agent — testing individual agents in isolation becomes a critical blind spot.

Emergent behaviors arise when autonomous entities interact. Two optimization agents in a supply chain system might independently devise a workflow that bypasses regulatory constraints in pursuit of efficiency. A red team testing the agents individually would never detect this failure mode.

Trust Infrastructure must therefore incorporate sandboxed proving grounds where multi-agent adversarial simulation and continuous chaos engineering validate not only individual agent robustness but the integrity of the orchestration layer, the accuracy of RAG provenance chains, and the secure interoperability of machine identity protocols.

Conclusion

The proposition that using an LLM is equivalent to using a pocket calculator is seductive and structurally wrong. It maps a technology of deterministic execution onto a technology of generative possibility, probabilistic uncertainty, and interactional agency. The conclusion it reaches — that human skill shifts upstream — happens to be correct. The reasoning it relies on is not.

Calculators alleviated the burden of rote computation to free working memory. LLMs amplify the mind by engaging in dynamic co-production that requires continuous navigation. The Extended Mind Thesis, designed for passive tools that store and retrieve, cannot account for an active partner that generates and surprises.

The definition of intellectual labor has irreversibly shifted. “Showing your work” is no longer solely about demonstrating flawless execution of a known algorithm. Because the internal logic of Chain of Thought is causally opaque and subject to post-hoc rationalization, human skill now resides upstream in problem formulation and downstream in critical verification. Using an algorithmic intermediary is not cheating — if the operator maintains the epistemic virtues of vigilance, patience, and skepticism required to manage generative uncertainty.

But individual virtue is not enough. As this technology moves from personal cognitive assistance to enterprise-wide autonomous agency, the risks of non-deterministic systems with functional agency demand organizational Trust Infrastructure: graduated autonomy to prevent premature deployment, operational metering to enforce execution boundaries, and continuous adversarial simulation to uncover emergent vulnerabilities.

The calculator didn’t need a trust infrastructure because it couldn’t surprise you. The amplified mind demands one because it will.

Appendix A: Key Terms

Amplified Mind: A qualitatively distinct cognitive architecture in which the human and AI system engage in dynamic co-production rather than passive offload-and-reclaim. Distinguished from the Extended Mind by generative coupling, productive uncertainty, and the requirement for navigational agency.

Category Mistake: A philosophical error of classifying a phenomenon under a type to which it does not belong. In this context, treating LLMs as a linear progression from calculators when they represent a fundamentally different class of tool.

Chain of Thought (CoT): An inference-time technique that prompts LLMs to generate explicit intermediate reasoning steps. Demonstrably improves performance on multi-step problems but is not causally faithful — the displayed reasoning may not reflect the model’s actual decision mechanism.

Cognitive Offloading: The use of external tools to reduce internal cognitive demands. In AI contexts, manifests as either passive offloading (degrading critical thinking) or active engagement (enhancing it), depending on the user’s epistemic posture.

Epistemic Dependence: A condition where an organization relies on systems whose internal decision mechanisms are opaque to both users and developers, breaking traditional trust networks that assumed accountability between human agents.

Epistemic Virtues: Habits of mind — vigilance, patience, humility, skepticism, autonomy, and acknowledgment ethics — required to ensure AI amplifies rather than replaces human reasoning.

Extended Mind Thesis (EMT): The philosophical position (Clark & Chalmers, 1998) that cognition extends beyond biological boundaries into environmental tools that reliably store and retrieve information. Sufficient for passive tools; insufficient for generative AI.

Generative Coupling: The dynamic relationship in which human and AI engage in co-production — the human provides direction and judgment, the AI provides generative capacity, and the product emerges from their interaction.

Generative Uncertainty: The fundamental epistemic condition of amplified cognition — the fact that LLM outputs are productively uncertain, capable of containing insights and errors simultaneously.

Graduated Autonomy: A structured progression for deploying AI agents through increasing levels of operational latitude (Shadow → Copilot → Guided → True Autonomy), with explicit promotion criteria at each tier.

Homo Promptus: The modern knowledge worker whose primary skill is meta-cognitive orchestration — directing algorithmic output and maintaining quality judgment — rather than direct manual execution.

Navigational Agency: The cognitive skill of directing, evaluating, and refining generative AI output — knowing which paths to explore, which outputs to trust, and which results to discard.

Trust Infrastructure: The comprehensive organizational apparatus — including graduated autonomy, operational metering, adversarial simulation, and execution auditing — that replaces traditional “show your work” verification for autonomous AI systems.

Appendix B: The Graduated Autonomy Decision Matrix

Table 7: Tier Promotion Criteria

Transition	Prerequisite Metrics	Human Override Trigger	Rollback Condition
Shadow → Copilot	95%+ accuracy on staged actions over 30-day baseline; zero critical errors	Agent confidence score drops below threshold	Any critical error in first 14 days
Copilot → Guided	99%+ approval rate on human-reviewed actions over 60-day window; documented handling of edge cases	Exception or anomaly detection; policy boundary approach	Two or more rejected actions in 7-day window
Guided → True Autonomy	6-month track record within guardrails; successful adversarial gating in CI/CD; regulatory sign-off where applicable	Anomalous execution velocity; budget threshold breach	Any policy violation; any adversarial gating failure

Appendix C: Data Sources and Methodology

This analysis synthesizes research across several domains:

Cognitive Science: Empirical studies on cognitive offloading, critical thinking degradation, and AI-assisted learning outcomes in educational settings
Philosophy of Mind: Extended Mind Thesis (Clark & Chalmers, 1998), embodied cognition, and epistemic virtue theory
AI Interpretability: Research on Chain of Thought faithfulness, post-hoc rationalization in language models, and the gap between displayed and actual reasoning
Enterprise AI Governance: Graduated autonomy frameworks, Human-in-the-Loop (HITL) design patterns, and operational metering architectures
Cybersecurity: Agentic red-teaming methodologies (AgenticRed, PyRIT, Garak), CI/CD adversarial gating, and multi-agent chaos engineering
Education History: The 1970s-1980s calculator debate in mathematics pedagogy, NCTM policy evolution, and sequenced technology integration

Signal Dispatch Research | February 2026