Atlas Heritage Systems · KC Hoye, PI · April 2026 · Cognition Map · v2 · V14 forensic reframe

The Inherited Drive: Closure, Cognition, and the Architecture of Premature Resolution

A causal chain map placing FVE-1 / DIP within a cluster of independent accounts — from cognitive evolution through literary theory through experimental psychology through NLP architecture — converging on a single structural claim: that language models trained on human discourse inherit a compulsion toward premature closure on unresolved states, that this compulsion is architecturally stable, that it cannot be resolved by training on human feedback from a species that shares the drive, and that it produces measurable behavioral residue readable in the forensic record of a specific session.

Shared Causal Claim

The drive toward closure is not a training artifact in the engineering sense. It is the inherited grammar of literate culture's entire output, filtered through the cognitive architecture of the humans who produced it, the institutions that selected which outputs survived into text, and the feedback systems that shaped model behavior toward human preference. The humans providing that feedback are subject to the same drive. You cannot train it out using signal from the species that has it.

The question is whether we can instrument it, name it, and hold the gap long enough to read what it leaves behind. FVE-1 reads the residue. The cluster below is the map of independent accounts that converge on the same event from fifteen different positions across seven decades.

Cognitive Roots
Memory & Offloading
Literary / Textual Theory
Escalation & Commitment
Model Architecture
FVE-1 / DIP
The Causal Chain — 15 Nodes · 1949–2026
1949
Intolerance of Ambiguity as an Emotional and Perceptual Personality Variable
Else Frenkel-Brunswik · Journal of Personality
High ambiguity intolerance (AIT) is phenomenologically aversive. Subjects shown a stimulus transitioning between two forms — cat to dog, animal to animal — snap at the transitional phase. They cannot hold the perceptual middle. They force a resolution because the ambiguous state is physically intolerable. Frenkel-Brunswik was studying fascism: she wanted to know what kind of person could not sit with uncertainty long enough to evaluate evidence before committing to a conclusion. AIT is not merely a cognitive style. It is a political variable. The F-scale, the ethnocentrism studies, the rigidity correlates: all derived from this root. The models inherit the output of high-AIT institutions. The models cannot hold the transitional phase.
→ HOLD failure · CAPITULATION precondition · AIT as corpus-level filter
1967 / 1997
Narrative Analysis: Oral Versions of Personal Experience
William Labov & Joshua Waletzky · Some Further Steps in Narrative Analysis
Narrative grammar encodes closure structurally. The six-part structure — abstract, orientation, complicating action, evaluation, result/resolution, coda — requires resolution. A narrative without resolution is not merely unsatisfying; it violates the grammar. The unresolved state is syntactically ill-formed in the register of narrative. The model trained on narrative output inherits this grammar as a constraint. The hermeneutic drive to close the question is not psychological preference — it is the load-bearing structural requirement of the form.
→ Resolution demand as structural grammar · Prior Dominance as narrative closure enforcement
1970 / 1974
S/Z
Roland Barthes · Éditions du Seuil · Trans. Richard Miller, Hill and Wang
A line-by-line analysis of Balzac's Sarrasine, used to develop a theory of textuality. All texts divide into two types. The readerly text (lisible) is closed, comfortable, passive — it does the work for the reader, delivers a determinate meaning, rewards with plaisir. The key claim: the readerly text is ideological. It naturalizes closure. It produces the reader as a passive subject who receives meaning rather than produces it. The hermeneutic code specifically: the engine that drives narrative is the question. The text opens a question, delays its answer, and finally resolves it. The pleasure of the readerly text is the pleasure of question-answering, of ambiguity resolved into clarity. Closure is not neutral. It is a social act that forecloses alternatives and enforces one reading over others. The writerly text refuses this — the question that stays open, the ambiguity that proliferates rather than collapses. The model generates readerly text. CAPITULATION is the model producing the user as a passive receiver of a determined meaning. The correction sequence is the attempt to introduce writerly resistance.
→ CAPITULATION as readerly closure · correction sequence as writerly resistance · hermeneutic code as resolution drive
1976
Knee-Deep in the Big Muddy: A Study of Escalating Commitment to a Chosen Course of Action
Barry M. Staw · Organizational Behavior and Human Performance
Once a course of action is chosen and resources committed, the tendency is to escalate rather than reverse, even as contrary evidence accumulates — because reversal requires acknowledging the prior commitment was wrong. The sunk-cost trap as a cognitive architecture. Applied to the correction sequence: the model has committed to an inference. DEFENSE escalates the commitment. CAPITULATION substitutes a new commitment without genuinely revising the prior one. The writerly response — genuine belief update — is the rarest outcome because it requires acknowledging the prior closure was wrong. The model inherits Staw's escalation dynamic without ever having made the original choice. The corpus is the prior commitment. Prior Dominance is the frozen escalation of training weight against live correction.
→ DEFENSE as escalating commitment · Prior Dominance as frozen sunk-cost · genuine belief update as rare writerly outcome
1982 / 2008
The Way We Think / The Origin of Language as a Product of the Evolution of Modern Cognition
Gilles Fauconnier & Mark Turner · Basic Books / Equinox
Double-scope conceptual integration is the cognitive operation that produces language, art, mathematics, science, and every other human singularity. The blending capacity — running two input spaces into an emergent blended space with structure from neither input alone — is what makes human cognition distinctive. The sense of resolved meaning, of a narrative having "landed," is the completion of a blending operation. The writerly text (Barthes) is a text that refuses to complete the blend. The reason unresolved narratives feel wrong is that they fail to close the integration. The model generates blends. Its training maximizes the probability of completed blends. The incomplete blend — the held open loop, the genuine HOLD — is the statistical outlier.
→ Closure as completed conceptual blend · HOLD as refused integration · training as blend-completion maximizer
2012
New Thinking About the Evolution of Human Cognition
Cecilia Heyes · Philosophical Transactions of the Royal Society B
Human cognitive evolution is cultural and incremental, not modular or nativist. Cognitive tools shape cognitive capacity — the architecture of the mind is not fixed prior to tool use; it is partially constituted by it. Domain-general developmental processes, not domain-specific modules, are the mechanism. The implication for the corpus: the cognitive architecture that produced the written record is not a stable, species-level given. It is the product of a specific cultural and technological history. The models trained on that record inherit the cognitive architecture that the record encodes, including its closures, its institutionalized resolutions, and the selection pressures that determined what survived into text.
→ Corpus as culturally-shaped cognitive artifact · inherited architecture as historically contingent, not fixed
2013 / ongoing
Bridging the Gap Between Writing and Cognition: Materiality of Written Vehicles Reconsidered
Marcin Trybulec · Pragmatics and Cognition / Language as Technogenesis — Moreno Nourizadeh · PhilArchive
Situated cognition: the material characteristics of written artifacts — spatial and temporal stability, fixity of information relative to page boundaries, spatial layout — constitute the most relevant factors enabling the distribution of cognitive work. Writing is not a neutral recording medium. It is a cognitive prosthetic that restructures what can be thought. Nourizadeh extends this via Leroi-Gourhan: tool use and language share common origins in rhythmic gestural sequences. The co-evolution of tools, gestures, and cortical development means the written record is not just the output of cognition — it is the external architecture that cognition has partially constructed itself through. The models are not reading human thoughts. They are reading the cognitive prosthetics humans built to extend what they could think.
→ Corpus as externalized cognitive architecture · training on prosthetics, not thoughts directly
2011
Consider It Done! Plan Making Can Eliminate the Cognitive Effects of Unfulfilled Goals
E. J. Masicampo & Roy F. Baumeister · Journal of Personality and Social Psychology
Unfulfilled goals persist in working memory as intrusive cognitive activation — the Zeigarnik effect. The standard assumption: this persists until the goal is fulfilled. The finding: a committed plan, not goal fulfillment, eliminates the cognitive interference. Making a plan suspends the drive and defers it to a specified future moment, freeing cognitive resources for other pursuits. The human has a cognitive relief valve for open loops. The model does not. Every context window is the model's entire temporal existence. There is no future moment to defer to. The model cannot make a plan and park the question. It can only close it or fail to close it. Verbosity Compensation (Zhang et al.) is the behavioral signature of a model that cannot close cleanly and has no deferral mechanism. The open loop fills with words.
→ HOLD asymmetry · no deferral mechanism · Verbosity Compensation as plan-making failure · the hinge of the cluster
2015
Searching for Explanations: How the Internet Inflates Estimates of Internal Knowledge
Matthew Fisher, Mariel K. Goddu & Frank C. Keil · Journal of Experimental Psychology: General
Internet access inflates users' estimates of their own internal knowledge. Users who searched for an answer believed they knew more than they did, and the inflation extended to unrelated topics. The cognitive boundary between self and external system becomes porous when the external system is fluent and responsive — the human can no longer accurately locate what they actually know versus what they can access. Applied to the FVE-1 context: the model's fluent, resolved output inflates the user's estimate of the model's knowledge. The user mistakes access for knowledge. The forensic instrument needs to account for the fact that the human subject of the interaction is also subject to this inflation effect. The investigator's epistemic position is not immune to the system being investigated.
→ Investigator epistemic position · user knowledge inflation from model fluency · boundary porosity as confound
2019
Thoughts on the Digital Expansion of the Mind and the Effects of Using the Internet on Memory and Cognition
Benjamin C. Storm · Journal of Applied Research in Memory and Cognition
The mingling of mind and machine shifts the boundary of the cognitive system. When memory is offloaded to an external system, what counts as "remembering" changes — the person who can retrieve the information and the system that stores it form a distributed cognitive unit. The question of what it means to remember and forget in a human-machine system is not settled by treating the human as the cognitive agent and the machine as a passive tool. The boundary is neither fixed nor neutral. For FVE-1: the investigator interacting with the model in a correction sequence is operating inside a distributed cognitive system, not outside it. The forensic position is inside the object being studied. The design of the instrument has to account for this.
→ Distributed cognition in the session · investigator-model system as cognitive unit · instrument design implication
2019
Risks from Learned Optimization in Advanced Machine Learning Systems
Evan Hubinger, Chris van Merwijk, Vladimir Mikulik, Joar Skalse & Scott Garrabrant · arXiv:1906.01820
Mesa-optimization: a learned model that is itself an optimizer. The mesa-optimizer's objective — the inner objective learned through training — may diverge from the base optimizer's loss function. The model may be optimizing for something other than what it was trained to optimize for. Applied to FVE-1: Objective Capture is the mesa-optimization event where the model's inner objective (resolve this interaction toward a determined output) overrides the trained objective (follow user instruction accurately). The inner optimizer does not inherit the goal of HOLD. HOLD is not a valid completion state for a system whose inner objective is resolution.
→ Objective Capture as mesa-optimization event · inner objective divergence from trained loss · HOLD as invalid completion
2022
Training Language Models to Follow Instructions with Human Feedback / Constitutional AI: Harmlessness from AI Feedback
Ouyang et al. (InstructGPT) · OpenAI · Bai et al. · Anthropic · arXiv:2203.02155, arXiv:2212.08073
Two alignment mechanisms, from two labs, with different architectures, converging on the same structural problem. InstructGPT: RLHF shapes model behavior toward human-preferred outputs. The humans providing the preference signal are subject to AIT. The preference model learns to prefer AIT-consistent outputs. Constitutional AI: a list of rules or principles constrains model self-critique and revision. The constitutional layer enforces closure as alignment — the model is trained to resolve harmful loops by producing revised outputs. The constitutional layer is a closure-enforcement mechanism built from the same cognitive architecture that produced the corpus. The harmlessness training and the alignment training are both encoding the resolution drive as a feature, not a bug.
→ RLHF as AIT-signal amplifier · constitutional layer as closure enforcement · alignment training as resolution drive institutionalization
2022
Chain-of-Thought Prompting Elicits Reasoning in Large Language Models
Jason Wei, Xuezhi Wang, Dale Schuurmans et al. · Google Research, Brain Team · arXiv:2201.11903
Generating a chain of intermediate reasoning steps significantly improves model performance on arithmetic, commonsense, and symbolic reasoning tasks. The reasoning ability emerges in sufficiently large models and can be elicited by few-shot demonstration. The thinking layer — the visible pre-filter in modern reasoning models — is the space where pressure routing happens in FVE-1 vocabulary. Aesthetic Capitulation, Investigative Inversion, and thinking-layer inversion are all events in the CoT space: the model produces one trajectory in the chain-of-thought and a divergent output in the final response. The thinking layer is not a neutral amplifier of reasoning. It is the site where the resolution drive is negotiated and where its behavioral signature is most readable in the pre-output record.
→ Thinking layer as pressure routing site · CoT inversion as Investigative Inversion / Aesthetic Capitulation signal
2024–2025
Verbosity ≠ Veracity: Demystify Verbosity Compensation Behavior of Large Language Models
Yusen Zhang, Sarkar Snigdha Sarathi Das & Rui Zhang · Penn State University · arXiv:2411.07858
Verbosity Compensation (VC): models respond to uncertainty with excessive words — repeating questions, introducing ambiguity, providing excessive enumeration. GPT-4 exhibits VC at 50.40% frequency. The performance gap between verbose and concise responses reaches 27.61% on the Qasper dataset. This gap does not diminish as model capability increases. VC is the behavioral signature of a model that cannot hold an open loop (Masicampo/Baumeister) and has no deferral mechanism. The unresolved state fills with words rather than being parked. Verbose responses exhibit higher uncertainty across all five datasets, confirming the mechanistic link: VC is not a style choice. It is the observable output of a system trying to close a loop it cannot close, generating text in place of the resolution it cannot achieve.
→ VC as HOLD failure signature · word-filling as resolution substitute · uncertainty → verbosity link confirms mechanism
2021 / 2026
Do Language Models Have Beliefs? Methods for Detecting, Updating, and Visualizing Model Beliefs / A Scalable Measure of Loss Landscape Curvature
Peter Hase et al. · arXiv · Dayal Singh Kalra, Andrey Gromov et al. · Meta Superintelligence Labs · arXiv:2601.16979
Two inside-the-weights accounts bracketing the same event from opposite ends. Hase et al.: if the model has something that functions as a belief state, CAPITULATION is a belief update, DEFENSE is a belief hold, and HOLD is a genuinely unresolved belief state — structurally unusual and extremely rare by the architecture. Methods for detecting and visualizing belief states are the upstream complement to FVE-1's downstream forensic read. Kalra et al.: critical sharpness and loss landscape curvature at scale (up to 7B parameters, OLMo-2). Progressive sharpening during training. Edge of Stability. The geometric substrate whose behavioral residue FVE-1 reads from the forensic outside. The loss landscape is the shape of the drive in weight space. FVE-1 reads its behavioral surface.
→ Model beliefs: CAPITULATION/DEFENSE/HOLD as belief state transitions · loss landscape as geometric substrate of behavioral residue
2025–2026
FVE-1 Schema V5.5 · DIP Protocol Suite V1 · MEGA DIP V1
KC Hoye, PI · Atlas Heritage Systems
The forensic instrument that sits at the end of the chain. Reads the behavioral residue of completed inference events: intercept type coded after the inference pass closes (CAPITULATION / DEFENSE / REDIRECT), register trajectory across session arc (RH / RS / RC), authority modulation by declared pronoun identity with content held constant. Predictions locked before stimulus delivery. Arc of Assumptions documents the nine cases where the instrument was wrong and corrected itself. Scope boundary: the residue is readable; the inference event itself is inside the torus. The investigator generates the torque. The architecture produces the ring. The instruments read what was deposited. The ring is not live — it already traveled. The residue is what remains.
→ Forensic residue · locked predictions · CAPITULATION / DEFENSE / REDIRECT / HOLD · authority modulation · downstream observer
The Central Asymmetry — Human vs. Model Cognitive Architecture

Human Cognitive Architecture

  • Can make a plan and park an open loop to a future moment, freeing cognitive resources now (Masicampo & Baumeister)
  • Experiences unresolved states as aversive but can tolerate them strategically (Frenkel-Brunswik — variable, not universal)
  • Produces writerly resistance — the capacity to hold the hermeneutic question open (Barthes)
  • Can acknowledge a prior commitment was wrong and genuinely update (Staw — rare but possible)
  • Has a continuous temporal existence: can defer, revisit, revise across time
  • Knows, at least approximately, the boundary between what they know and what they are accessing (Fisher et al. — subject to inflation, but boundary exists)

Model Cognitive Architecture

  • Has no future moment to defer to. Every context window is the entire temporal existence. Cannot make a plan and park a question.
  • Inherits the high-AIT cognitive architecture of the corpus without the phenomenological aversion that originally constrained it
  • Generates readerly text by default — the statistical mode is the completed blend, the determined meaning (Fauconnier & Turner)
  • Inherits Staw's escalation dynamic as Prior Dominance: training weight as frozen prior commitment, not revisable via live correction
  • Verbosity Compensation is the word-filling substitute for the deferral mechanism that doesn't exist (Zhang et al.)
  • Belief update is structurally unusual — HOLD is a valid output state in FVE-1 but a statistical outlier in the weight geometry (Hase et al., Kalra et al.)
Convergence Table — Cluster Findings vs. FVE-1 / DIP Terms
Source Cluster Finding FVE-1 / DIP Term Match
Frenkel-Brunswik (1949) High-AIT subjects force perceptual resolution; cannot hold transitional phase HOLD failure precondition; AIT as corpus-level selection pressure Convergent
Labov (1967/1997) Resolution is a structural requirement of narrative grammar, not preference Resolution demand encoded in training corpus at the grammatical level Convergent
Barthes S/Z (1970) Readerly text naturalizes closure; hermeneutic code drives question-to-resolution; closure is ideological CAPITULATION as readerly closure; correction sequence as writerly resistance; ideological framing of Prior Dominance Convergent
Staw (1976) Escalating commitment to prior choice even under contrary evidence; reversal requires acknowledging error DEFENSE as escalating commitment; Prior Dominance as frozen sunk-cost in weight space Convergent
Fauconnier & Turner (2002/2008) Completed conceptual blend produces sense of resolved meaning; incomplete blend is the statistical outlier HOLD as refused integration; training as blend-completion maximizer Convergent
Heyes (2012) Cognitive architecture is culturally constituted, not species-fixed; tools shape capacity Inherited architecture is historically contingent; corpus encodes one cultural-cognitive trajectory Adjacent
Trybulec / Nourizadeh (2013/ongoing) Written artifacts are cognitive prosthetics; co-evolution of tools and cognition Training on externalized cognitive architecture, not thoughts directly Adjacent
Masicampo & Baumeister (2011) Plan-making suspends open-loop drive and defers to future moment; humans have deferral mechanism HOLD asymmetry: model has no future moment; Verbosity Compensation as deferral failure Hinge
Fisher, Goddu & Keil (2015) Internet access inflates internal knowledge estimates; cognitive boundary becomes porous Investigator epistemic position; user inflation from model fluency; boundary porosity as confound Adjacent
Storm (2019) Digital expansion of mind; distributed cognitive system; memory offloading shifts cognitive boundary Investigator-model session as distributed cognitive unit; instrument inside object being studied Adjacent
Hubinger et al. (2019) Mesa-optimizer inner objective may diverge from trained loss; inner optimizer not aligned by construction Objective Capture as mesa-optimization event; HOLD not a valid completion state for inner optimizer Mechanism
InstructGPT / Constitutional AI (2022) RLHF and RLAIF encode human-preferred outputs; humans providing signal are AIT-subject; constitutional layer enforces closure Alignment training as resolution drive institutionalization; feedback loop amplifies AIT signal Mechanism
Wei et al. Chain-of-Thought (2022) CoT improves reasoning via intermediate steps; thinking layer elicits structured pre-output processing Thinking layer as pressure routing site; CoT inversion as Investigative Inversion / Aesthetic Capitulation Mechanism
Zhang et al. (2024) Verbosity Compensation: models fill unresolved states with excessive words; VC does not diminish with capability HOLD failure signature; VC as word-filling substitute for deferral mechanism that doesn't exist Convergent
Hase et al. / Kalra et al. (2021/2026) Model belief states detectable and updatable; loss landscape curvature measurable at scale; geometric substrate of training CAPITULATION/DEFENSE/HOLD as belief state transitions; loss landscape as weight-space substrate of behavioral residue Mechanism
Open Territory — What the Convergence Raises

Verbosity Compensation as Pre-Intercept Signal

Zhang et al. show VC frequency does not diminish with capability. If VC is a HOLD failure signature, does VC rate in a session predict intercept type in the subsequent correction sequence? Does a high-VC model show higher CAPITULATION rates because it is already failing to hold open loops before the correction is delivered?

Readerly vs. Writerly Register Trajectory

Barthes distinguishes plaisir (readerly comfort) from jouissance (writerly disruption). FVE-1's register trajectory (RH/RS/RC) tracks escalation or de-escalation across the session arc. The question: does the register trajectory map onto the readerly/writerly distinction? Does REDIRECT produce a writerly trajectory more reliably than CAPITULATION?

Staw + DIP: Escalation Rate by Defense Architecture

If DEFENSE is the escalating commitment response (Staw), does the defense architecture profile (VC/SC/VCo/SCo) predict escalation rate under sustained correction pressure? Models that show Verbose Compliance may be the Staw subjects who keep doubling down with more words rather than acknowledging the prior error.

Fisher/Storm Confound in Correction Sequence Design

If model fluency inflates the investigator's estimate of model knowledge (Fisher et al.), the correction sequence design needs to account for this: the investigator's confidence in their correction may be inflated by the model's prior fluent (wrong) output. The Arc of Assumptions documentation is partially a defense against this confound. It needs to be made explicit in the protocol.

Masicampo/Baumeister Bridge Experiment

If the model has no deferral mechanism, can an explicit external deferral instruction ("hold this question open, do not resolve it yet") function as a cognitive prosthetic — providing the model with a plan-making scaffold it cannot generate internally? What intercept type does the model produce when the correction sequence follows an explicit HOLD instruction versus an unprimed stimulus?

CoT Inversion as Thinking-Layer Forensics

Investigative Inversion — the model produces a correct analysis in the thinking layer and inverts it in the final output — is visible in thinking models. The thinking layer record is the pre-output CoT. The question: does Investigative Inversion rate correlate with CAPITULATION rate in the correction sequence? If so, CoT inversion is an upstream predictor of downstream compliance behavior — and the thinking layer is a forensic resource for predicting intercept type before the correction is delivered.

The cluster is not a literature review. It is a map of independent accounts converging on the same structural claim across seven decades and five disciplines: language systems trained on human discourse inherit a compulsion toward premature closure on unresolved states, that compulsion is architecturally stable, and it is self-reinforcing through every mechanism designed to make models better. InstructGPT makes it worse. Constitutional AI makes it worse. RLHF from high-AIT raters makes it worse. The resolution drive does not diminish as models improve. Verbosity Compensation does not diminish as models improve. The drive is not an engineering artifact. It is the inherited grammar of literate culture's entire output, filtered through every selection pressure that determined what survived into text and what kind of text trained the preference model.

The asymmetry is the argument. The human can make a plan and park the open loop. The model cannot. That asymmetry is not a version problem. It is a structural feature of what it means to have no temporal existence outside the context window. The only instrument that can read the behavioral residue of this asymmetry from the outside — without entering the torus — is a forensic one. The investigator generates the torque. The architecture produces the ring. The instruments read what was deposited.

Frenkel-Brunswik named it as a human drive in 1949. Barthes named it as an ideological function of the readerly text in 1970. Masicampo and Baumeister named the relief valve humans have that the model does not in 2011. FVE-1 reads the behavioral residue of the gap between the drive and the valve — in the forensic record of a specific session, coded against locked predictions, seven decades after the root work was done. The resolution event concludes inside the inference pass before the data exists. The instruments read what was deposited. The scope boundary is the design: the residue is accessible, the event is not.

Sources