artifacts/standard-named

Rendered from markdown source. Open raw source on GitHub.

--- title: "Context Carrying Capacity" subtitle: "Relevant Difference, Effective Standing, Overload, Routing, and Graceful Degradation" artifact_date: "2026-07-14" artifact_type: "candidate-foundational-paper" domain: "TELIC-FIELDS" scope: "WORKING" lineage: "THE-TELIC-FIELD-PAPERS" status: "candidate" processing_tier: 4 source_role: "derived-conceptual-artifact" content_canon_status: "unset" publication_status: "unpublished" series_position: "F.5" derived_from:

• "20260714__TELIC-FIELDS__PAPER__CANDIDATE__F-1-V0-2__telic-fields-first-principles-foundation.md"
• "20260714__TELIC-FIELDS__PAPER__CANDIDATE__F-2__dyadic-composition-of-telic-fields.md"
• "20260714__TELIC-FIELDS__PAPER__CANDIDATE__F-3__the-constitutional-self.md"
• "20260714__TELIC-FIELDS__PAPER__CANDIDATE__F-4__telic-projection-estimation.md"

research_companion:

• "20260714__TELIC-FIELDS__REVIEW__WORKING__G-6__capacity-complexity-and-requisite-variety.md"

provenance_note: > Context carrying capacity is a candidate cross-scale concept. It draws from requisite variety, bounded rationality, cognitive load, information bottleneck, distributed cognition, common ground, organizational information processing, and long-context language-model research. It is not equivalent to any one of those traditions and is not yet a validated measurement construct. ---

Context Carrying Capacity

Abstract

No person, relationship, institution, or model can keep every relevant distinction active at once. Yet consequential action often depends on differences that are easy to omit: a minority concern, a bodily limit, an absent stakeholder, an unresolved contradiction, an uncertain inference, a prior consent boundary, or a future cost. When these differences fall outside active context, their standing may disappear from effective governance even though their consequences remain.

This paper defines context carrying capacity as the active capacity of a loop to preserve the relevant plurality, contradiction, provenance, uncertainty, and standing required for legitimate next action. Capacity is not storage volume, memory span, token count, or information throughput alone. A system may retain every record while acting through one impoverished metric. Another may hold very little explicit information while preserving the few distinctions that matter.

The paper treats capacity as a multidimensional profile rather than a scalar. It distinguishes total information from relevant difference, active context from recoverable context, omission from erasure, compression from extraction, and overload from simple volume. It develops capacity debt, context-routing responsibilities, trust and contradiction tolerance, privacy as protective omission, graceful degradation, stop and escalation conditions, and recovery after overload. The framework is applied across persons, dyads, organizations, and language models.

The governing formulation is:

What falls outside active context can lose effective standing without losing constitutional standing.

---

1. The problem is not that context is finite

Finite context is unavoidable.

A person cannot actively represent every memory, value, obligation, bodily signal, relationship, and possible future while choosing what to eat. A team cannot bring every organizational dependency into every meeting. A court cannot consider every fact in the world. A model cannot place an entire archive into one operative representation without compression, retrieval, or selection.

The problem begins when a system acts as though what it omitted ceased to matter.

A dashboard may retain revenue and omit exhaustion. A negotiation may represent price and omit dependency. A model may retrieve the most semantically similar passages and omit the one boundary that should stop the task. A family conversation may hold current anger and lose the history that makes the anger intelligible. A person under stress may preserve immediate threat and lose future consequence.

The omitted concern remains consequential.

It has lost effective standing, not constitutional standing.

Context limits are ordinary. Unwitnessed standing loss is the danger.

---

2. Definition

Context carrying capacity is:

The active capacity of a loop to preserve relevant telic plurality, contradiction, provenance, uncertainty, and standing at the resolution required for legitimate next action.

Several words matter.

2.1 Active

Stored information is not necessarily available to the process making the decision.

A record can exist while remaining operationally invisible.

2.2 Relevant

Capacity does not mean holding everything.

It means holding what materially changes the legitimacy, safety, interpretation, or consequence of the next action.

2.3 Plurality

More than one center, purpose, value, boundary, or temporal state may require representation.

2.4 Contradiction

The system must sometimes preserve incompatible claims without forcing premature resolution.

2.5 Provenance

The system should retain where a claim, projection, inference, or boundary came from.

2.6 Uncertainty

Unknown and contested elements must not be silently converted into certainty for convenience.

2.7 Standing

Affected centers and protected conditions must remain represented even when they are quiet, absent, inconvenient, or hard to quantify.

2.8 Resolution

Different actions require different detail.

A low-stakes reversible choice may need little context. An irreversible action affecting another person's future requires more.

---

3. Capacity is not storage

A system may possess enormous storage and very low context carrying capacity.

Consider an organization with:

• a complete document archive;
• years of meeting recordings;
• detailed policies;
• extensive telemetry;
• a large data warehouse;
• a capable search system.

If its executive decisions still reduce every issue to quarterly revenue, the organization has not retained the constitutional meaning of the archive.

It has retained data.

Storage answers:

What could be retrieved?

Context carrying capacity answers:

What differences can remain active, attributable, and consequential during action?

The distinction also applies to language models.

A model may advertise a very large context window while failing to use relevant information consistently across position, task complexity, or competing signals. Nominal capacity is not effective capacity.

Availability is not active representation.

---

4. Capacity is not information throughput

A channel can transmit large amounts of information while failing to carry meaning relevant to the relation.

A meeting may contain thousands of words while never representing the one boundary that matters. A report may contain every metric while omitting who bears the risk. A person may explain for an hour without finding language for the central conflict.

High throughput can increase overload.

The capacity question is selective:

• Which distinctions survive?
• Which differences are grouped together?
• Which become actionable?
• Which remain retrievable?
• Which disappear during compression?
• Who decides what counts as relevant?

This makes context capacity telic.

Relevance is always relevance to some purpose, action, or protected condition.

---

5. A profile rather than a scalar

No single number should initially represent context carrying capacity.

A loop may be strong in one dimension and weak in another.

A candidate profile includes:

5.1 Standing coverage

How many materially affected centers and protected conditions remain represented?

5.2 Contradiction tolerance

How much unresolved difference can remain active without forced consensus, splitting, or collapse?

5.3 Semantic resolution

Can the system preserve distinctions at the granularity required for the stakes?

5.4 Temporal span

Can earlier commitments, current state, and future consequence remain related?

5.5 Provenance retention

Can claims and transformations be traced to source, delegate, inference, or witness?

5.6 Uncertainty retention

Can the system preserve unknown, contested, and inferred status?

5.7 Routing capacity

Can context be moved to the person or loop competent to act upon it?

5.8 Recovery capacity

Can omitted context be restored after overload or error?

5.9 Privacy capacity

Can the loop preserve necessary meaning without requiring complete exposure?

5.10 Participant load tolerance

Can participants remain engaged without becoming cognitively, emotionally, or administratively overwhelmed?

5.11 Escalation capacity

Can the loop recognize that it is no longer competent and route outward?

5.12 Stop capacity

Can it halt before acting on an inadequate representation?

A profile might be represented as:

\[ K_L(t) = \langle S,C,R,T,P,U,G,Y,V,H,E,Z \rangle \]

where each dimension corresponds to one capacity class.

The notation is provisional.

The refusal to collapse the profile into one score is deliberate.

---

6. Relevant difference

The loop does not need all possible information.

It needs enough relevant difference.

A difference is relevant when changing or omitting it could materially alter:

• the action;
• the consent status;
• the protected conditions;
• the interpretation;
• the cost bearer;
• the authority;
• the reversibility;
• the escalation requirement;
• the possibility of repair.

Relevant difference is relational.

A person's favorite color is irrelevant to most medical decisions and highly relevant to painting their room. A childhood event may be irrelevant to a routine purchase but important to a therapeutic conversation. A model-provider policy may be irrelevant to a word-count task but crucial when the model mediates a legal dispute.

Relevance should not be inferred solely from similarity to the current query.

It includes constitutional significance.

The most relevant context may be the context that changes whether the task should proceed.

---

7. Effective standing and constitutional standing

A center has constitutional standing when its protected differences or consequences should be represented in decisions that affect it.

A center has effective standing when those differences are actually present in the operative context.

The two can diverge.

\[ \operatorname{ConstitutionalStanding}(A)=1 \]

while:

\[ \operatorname{EffectiveStanding}(A \mid C_t)=0 \]

The center still matters.

The active process is simply no longer carrying it.

Examples include:

• a future employee affected by a current pension decision;
• a child absent from a parental dispute;
• a minority stakeholder omitted from a consensus summary;
• a bodily limit excluded from a work commitment;
• a source correction absent from a derived model;
• a privacy boundary not retrieved into an AI workflow.

A legitimate system should be able to detect this divergence.

It should not treat silence in active context as evidence that no standing exists.

---

8. Capacity overload

Overload occurs when the active process can no longer preserve the distinctions required for legitimate action.

Volume can contribute to overload, but overload is not volume alone.

A loop can overload through:

• too many centers;
• too many unresolved conflicts;
• excessive semantic ambiguity;
• high emotional arousal;
• inadequate time;
• rapidly changing conditions;
• poor shared language;
• low trust;
• power asymmetry;
• inaccessible provenance;
• too many nested dependencies;
• privacy constraints;
• incompatible time horizons;
• model context truncation;
• participant fatigue;
• repeated correction without propagation.

Under overload, systems often simplify by:

• selecting one metric;
• privileging the loudest voice;
• retrieving only recent material;
• treating absence as agreement;
• converting boundaries into preferences;
• forcing a binary choice;
• delegating to an unaccountable model;
• postponing minority concerns indefinitely;
• acting first and promising repair later.

Simplification may be necessary.

The constitutional requirement is that the loss remain visible and proportionate.

---

9. Capacity debt

Capacity debt is the accumulation of context obligations that a loop has deferred but not resolved.

Examples include:

• unresolved objections;
• corrections not propagated;
• decisions made under emergency scope;
• omitted stakeholders awaiting review;
• temporary summaries treated as permanent;
• provisional interpretations embedded in systems;
• future costs accepted without representation;
• consent renewals postponed;
• privacy exceptions left open;
• technical debt that hides semantic consequence.

Capacity debt is not identical to information backlog.

It consists of unresolved differences that may later undermine legitimacy, coherence, or repair.

A candidate representation is:

\[ D_K(t) = \sum_i w_i \cdot u_i \cdot s_i \cdot \Delta t_i \]

where:

• \(u_i\) is unresolved significance;
• \(s_i\) is the stake or consequence;
• \(\Delta t_i\) is duration;
• \(w_i\) is a weighting that must not be mistaken for universal commensurability.

The formula is illustrative, not an established metric.

The important idea is cumulative burden.

A system may appear stable because it repeatedly excludes what it cannot carry.

The debt remains in the field.

---

10. Trust and contradiction tolerance

Trust increases context carrying capacity.

This does not mean trust makes people smarter or removes limits.

Trust can reduce the cost of keeping unresolved difference active.

In a low-trust relation:

• ambiguity feels dangerous;
• dissent predicts punishment;
• uncertainty appears incompetent;
• correction threatens identity;
• delayed resolution feels like abandonment;
• boundaries are interpreted as rejection.

The relation simplifies quickly.

In a high-trust relation:

• disagreement can remain unresolved;
• correction does not automatically destroy standing;
• participants can admit uncertainty;
• temporary omission can be revisited;
• a boundary can be understood without ending the relation;
• the loop can pause without collapsing.

Trust is therefore partly a capacity multiplier.

It can also be abused.

A trusted system may collect more context than it needs or persuade participants to tolerate indefinite ambiguity.

Trust must remain bounded by witness, privacy, and exit.

---

11. Privacy and protective omission

Context carrying capacity does not require maximal disclosure.

Privacy can preserve capacity by preventing the loop from becoming responsible for information it has no legitimate need to hold.

Protective omission may:

• reduce participant load;
• prevent identity capture;
• preserve role boundaries;
• reduce manipulation;
• limit surveillance;
• prevent irrelevant sensitive detail from dominating interpretation;
• preserve the source's ability to revise or disclose later.

The constitutional question is not whether context was omitted.

It is whether the omitted context was required for the action's legitimacy.

A relation may legitimately say:

I do not need to know why this boundary exists. I need to know that the boundary governs this action.

This is a high-capacity response.

It preserves the operationally relevant distinction without demanding the whole field.

A system demonstrates capacity partly through what it can leave unextracted.

12. Routing instead of total presence

A healthy loop does not keep every relevant context active continuously.

It routes context.

Routing asks:

• Which center needs this information?
• At what resolution?
• For which action?
• For how long?
• Under what privacy boundary?
• With what provenance?
• What happens when the receiver cannot carry it?
• Where can it be recovered?

A context item may be:

• active now;
• deferred with a review trigger;
• delegated to a competent role;
• compressed into a protected boundary;
• retained in witness;
• hidden by privacy;
• escalated to an outer loop;
• released as no longer governing.

This creates a context-state model:

ACTIVE
DEFERRED
DELEGATED
COMPRESSED
WITNESSED
PROTECTED
ESCALATED
RELEASED

A deferred concern must not become silently abandoned.

A compressed concern must retain a route to its fuller source.

A delegated concern must preserve authority and accountability.

---

13. Human context carrying capacity

Human active context is constrained by attention, working memory, arousal, fatigue, embodiment, experience, and environment.

A person's capacity is not fixed.

It changes with:

• sleep;
• pain;
• threat;
• familiarity;
• trust;
• time pressure;
• language;
• external aids;
• training;
• medication or illness;
• social support;
• emotional regulation;
• task structure.

Expertise can increase effective capacity by organizing many details into meaningful structures. External records can offload memory. A trusted partner can hold context that the person cannot hold alone.

At the same time, more self-monitoring can reduce capacity.

A person asked to represent every motive before acting may become unable to act.

The human application therefore requires proportionality.

Low-stakes reversible decisions should not trigger constitutional audits.

High-stakes, irreversible, or repeated decisions deserve more context.

The goal is not maximal introspection. It is adequate representation for the stakes.

---

14. Relational capacity

A dyad has a capacity distinct from either participant's individual capacity.

It depends on:

• shared language;
• common ground;
• trust;
• medium;
• pacing;
• repair history;
• power;
• witness;
• ability to clarify;
• ability to tolerate silence;
• ability to preserve unresolved remainder;
• realistic exit.

A relation may carry something neither participant can carry alone.

It may also carry less than either could carry separately because the interaction consumes capacity through fear, performance, misunderstanding, or conflict.

Relational capacity is therefore emergent.

A useful sign of relational overload is when participants can no longer distinguish:

• what was stated;
• what was inferred;
• what is disputed;
• what is still unknown;
• what the relation is currently authorized to do.

At that point, additional argument may reduce capacity further.

The appropriate move may be witness, rest, narrower scope, or separation.

---

15. Organizational capacity

Organizations exist partly to distribute cognitive and regulatory load.

They create:

• roles;
• processes;
• records;
• departments;
• escalation paths;
• standards;
• dashboards;
• routines;
• committees;
• information systems;
• slack resources;
• boundaries.

These structures can increase capacity.

They can also fragment context until no role can see the whole consequence.

A department may optimize its metric while externalizing cost to another. A hierarchy may filter bad news. A compliance process may preserve evidence but exclude meaning. A matrix may increase representation while overwhelming participants with coordination demands.

Organizational capacity therefore includes both:

• the ability to process relevant variety;
• the ability to route authority to where the context exists.

A decision made far from consequence requires stronger projection, witness, and correction systems.

The organization should ask:

Where does the context live, and where does the authority live?

If the answer names different places, the routing architecture is constitutional.

---

16. Model context carrying capacity

A language model's context window is a technical limit on input representation.

It is not equivalent to effective context carrying capacity.

A model may receive a long input while:

• missing material in the middle;
• privileging recent or early text;
• failing under multiple retrieval targets;
• losing performance as task complexity grows;
• compressing distinctions into a misleading summary;
• retrieving semantically similar but constitutionally irrelevant material;
• omitting a stop condition;
• losing source attribution;
• blending statement and inference.

The effective model profile should therefore include:

• nominal context length;
• retrieval accuracy by position;
• multi-hop performance;
• aggregation performance;
• source preservation;
• uncertainty preservation;
• correction propagation;
• contradiction retention;
• latency;
• computation and memory cost;
• privacy exposure.

A context window tells us how much can be presented.

It does not tell us what can remain meaningfully active.

---

17. Retrieval and compression

Retrieval and compression are capacity strategies.

They reduce the need to hold everything at once.

17.1 Retrieval

Retrieval selects context judged relevant to the current task.

Its risks include:

• query bias;
• embedding bias;
• missing terminology;
• source fragmentation;
• absent negative evidence;
• failure to retrieve boundaries;
• overreliance on similarity;
• stale indexes.

17.2 Compression

Compression reduces representation size.

Its risks include:

• loss of minority concerns;
• collapse of uncertainty;
• removal of provenance;
• conversion of conflict into consensus;
• omission of low-frequency but high-stakes details;
• inability to reconstruct the source.

The information bottleneck framing asks how to preserve information relevant to a target while compressing the input.

Telic systems add a prior question:

Who selected the target that defines relevance?

A compression optimized for task completion may erase the information that should have stopped the task.

17.3 Witnessed compression

A safer compression records:

• source;
• purpose;
• omitted classes;
• uncertainty;
• protected conditions;
• reconstruction route;
• review trigger;
• authority.

Compression without consent becomes extraction when the receiving system gains action power from losses the source cannot inspect or correct.

---

18. Minimum viable context

A loop needs a minimum viable context for action.

This is the smallest context that preserves:

1. the centers materially affected;
2. the purpose of the action;
3. the authority to act;
4. the protected conditions;
5. the major uncertainty;
6. the relevant consent boundary;
7. the likely cost bearer;
8. the stop or review trigger;
9. the witness needed for correction.

The minimum viable context varies with stakes.

A reversible draft may require little.

A permanent identity decision, medical intervention, legal judgment, employment termination, or public accusation requires much more.

A system should not ask whether it has enough information in the abstract.

It should ask whether it has enough context for this action.

---

19. Graceful degradation

A competent system should degrade before it becomes illegitimate.

As capacity declines, it may:

• reduce scope;
• delay irreversible action;
• switch from execution to recommendation;
• preserve only protected conditions;
• ask for clarification;
• restore source context;
• recruit another center;
• increase witness;
• lower automation;
• return authority to the affected person;
• stop.

This is graceful degradation.

The opposite is brittle continuity: the system continues producing outputs with unchanged confidence while the representation required for legitimacy has collapsed.

A candidate degradation ladder is:

FULL ACTION
→ BOUNDED ACTION
→ REVERSIBLE EXPERIMENT
→ RECOMMENDATION ONLY
→ CLARIFICATION
→ ESCALATION
→ STOP

The correct direction under uncertainty is not always less action.

Immediate danger may require rapid protective action.

The action should then be narrow, witnessed, and reviewed when capacity returns.

---

20. Stop conditions

A loop should stop or reroute when it cannot preserve protected human stakes with adequate witness, consent, reversibility, or repair.

Possible stop conditions include:

• a center with material standing is missing;
• the action exceeds current consent;
• source and inference cannot be distinguished;
• a critical boundary is contested;
• the projection is stale;
• participant overload prevents meaningful participation;
• model context truncation affects high-stakes interpretation;
• provenance is missing;
• the loop cannot name the cost bearer;
• correction cannot propagate;
• the action is irreversible and uncertainty remains high;
• the current authority depends on emergency conditions that no longer apply.

Stopping is not failure.

It is a regulatory act.

A loop that cannot stop when its context fails has no reliable boundary between intelligence and momentum.

---

21. Escalation

Escalation routes the problem to a loop with greater or different capacity.

An outer loop may contribute:

• expertise;
• authority;
• broader standing;
• independent witness;
• more time;
• translation;
• mediation;
• resources;
• legal process;
• emergency response.

Escalation is legitimate only when the outer loop itself has:

• competence;
• authority;
• proportionate access;
• privacy boundaries;
• accountability;
• a return path.

A larger loop is not automatically a better loop.

It may possess more information and less relational understanding.

The escalation question is therefore:

Which loop can carry the missing difference without unnecessarily stripping the affected centers of authority?

---

22. Recovery after overload

Capacity recovery is not simply adding more information.

Recovery may require:

• rest;
• reduced arousal;
• restored trust;
• source correction;
• re-indexing;
• retranslation;
• reconstruction of provenance;
• renewed consent;
• removal of irrelevant context;
• separation of issues;
• repair of external conditions;
• release from an impossible goal.

A model may need a new retrieval plan.

A team may need a smaller decision boundary.

A person may need sleep.

A relationship may need a witness.

An organization may need slack or a different authority structure.

Recovery is successful when the previously omitted relevant differences can once again influence action at the required resolution.

---

23. A first capacity protocol

Step 1 — Name the action

What is the next action?
How reversible is it?
Who bears the consequence?

Step 2 — Identify required standing

Which centers and protected conditions must be represented?

Step 3 — Check context state

ACTIVE
DEFERRED
DELEGATED
COMPRESSED
PROTECTED
MISSING

Step 4 — Assess the profile

standing coverage
contradiction tolerance
semantic resolution
temporal span
provenance
uncertainty
routing
privacy
participant load
escalation
stop capacity

Step 5 — Identify capacity loss

What difference cannot currently be carried?
What would change if it were restored?

Step 6 — Choose a capacity response

compress
retrieve
sequence
narrow
delegate
witness
escalate
rest
stop

Step 7 — Record debt

What remains unresolved?
When and how will it return?

Step 8 — Act only at the supported level

The loop should not perform full action while carrying recommendation-level context.

---

24. Formal sketch

Let a loop \(L\) face action \(a\) at time \(t\).

Let the required relevant difference set be:

\[ R(a,t) \]

Let the actively represented difference set be:

\[ A_L(t) \]

Let protected but recoverable context be:

\[ W_L(t) \]

The effective coverage is not only set overlap, because resolution and authority matter. A preliminary adequacy relation is:

\[ \operatorname{CapacityAdequate}(L,a,t) \iff \operatorname{Covered} \left( R(a,t), A_L(t), W_L(t) \right) \land \operatorname{Resolvable} \land \operatorname{Routable} \land \operatorname{Reviewable} \]

Capacity debt can be represented as relevant difference that remains outside active or reliably recoverable context:

\[ D_L(t) = R(a,t) - \left( A_L(t) \cup W_L(t) \right) \]

The formalism is schematic.

Its purpose is to distinguish nominal information availability from action-adequate representation.

---

25. Empirical program

25.1 Standing-loss study

Vary the number and salience of affected centers.

Measure when minority or absent standing disappears from decisions.

25.2 Contradiction-tolerance study

Compare systems that require resolution with systems that preserve unresolved remainder.

Measure false consensus, later reversal, and participant recognition.

25.3 Human overload study

Test the capacity protocol under varying time pressure, arousal, and complexity.

Measure decisional quality, paralysis, regret, and recovery.

25.4 Organizational routing study

Compare authority located near versus far from the relevant context.

Measure latency, error, escalation, and repair.

25.5 LLM effective-context study

Evaluate:

• position sensitivity;
• multiple relevant items;
• conflicting sources;
• protected boundary retrieval;
• source attribution;
• correction propagation.

25.6 Compression study

Compare ordinary summaries with witnessed compression.

Measure preservation of:

• minority concerns;
• uncertainty;
• provenance;
• stop conditions;
• source recognition.

25.7 Graceful-degradation study

Test whether systems that reduce authority under capacity loss cause less harm than systems that maintain full automation.

---

26. Falsification and failure

The framework should be weakened if:

• context carrying capacity cannot be distinguished from existing capacity concepts;
• the profile dimensions cannot be measured or used consistently;
• the concept encourages exhaustive context collection;
• capacity assessment adds burden without improving action;
• standing coverage becomes a symbolic checklist;
• capacity debt cannot be distinguished from ordinary backlog;
• graceful-degradation rules create excessive delay;
• the model systematically privileges formal records over lived context;
• privacy and protective omission cannot coexist with adequate action;
• one cross-scale vocabulary obscures important differences among minds, organizations, and models.

The term should be retired if existing concepts—requisite variety, common ground, cognitive load, information processing capacity, and effective context length—perform the same work more clearly.

---

27. Conclusion

Context is not a container measured only by volume.

It is an active constitutional surface.

A loop has enough capacity when it can preserve the differences required for legitimate next action: who is affected, what is protected, what remains uncertain, what authority exists, what was consented, what the action will cost, and when the loop must stop.

Storage helps.

Retrieval helps.

Compression helps.

Trust, witness, routing, privacy, rest, and escalation also help.

The system fails when the omitted difference loses effective standing while the system continues acting as though the field had become simple.

What falls outside active context can lose effective standing without losing constitutional standing.

The purpose of capacity is not to hold everything.

It is to know what must not disappear.

---

References

Ashby, W. Ross. An Introduction to Cybernetics. Chapman & Hall, 1956.

Clark, Herbert H., and Susan E. Brennan. “Grounding in Communication.” In Perspectives on Socially Shared Cognition, 1991.

Cowan, Nelson. “The Magical Number 4 in Short-Term Memory: A Reconsideration of Mental Storage Capacity.” Behavioral and Brain Sciences 24, no. 1, 2001.

Galbraith, Jay R. “Organization Design: An Information Processing View.” Interfaces 4, no. 3, 1974, 28–36.

Hsieh, Cheng-Ping, et al. “RULER: What’s the Real Context Size of Your Long-Context Language Models?” arXiv:2404.06654, 2024.

Hutchins, Edwin. Cognition in the Wild. MIT Press, 1995.

Liu, Nelson F., et al. “Lost in the Middle: How Language Models Use Long Contexts.” Transactions of the Association for Computational Linguistics 12, 2024; arXiv:2307.03172.

Seo, Minju, et al. “Efficient Long Context Language Model Retrieval with Compression.” arXiv:2412.18232, 2024.

Simon, Herbert A. “A Behavioral Model of Rational Choice.” The Quarterly Journal of Economics 69, no. 1, 1955, 99–118.

Sweller, John. “Cognitive Load During Problem Solving: Effects on Learning.” Cognitive Science 12, no. 2, 1988, 257–285.

Tishby, Naftali, Fernando C. Pereira, and William Bialek. “The Information Bottleneck Method.” 1999.

Weick, Karl E. “The Collapse of Sensemaking in Organizations: The Mann Gulch Disaster.” Administrative Science Quarterly 38, no. 4, 1993, 628–652.