---
title: "Information Is Not an Update"
subtitle: "How new information gets compiled into old models—and how to make the difference survive"
author: Elias Kunnas
description: "A message can arrive, be quoted, and leave the old model unchanged. The update is the part the old model can no longer make disappear."
canonical: https://kunnas.com/articles/information-is-not-an-update
url: https://kunnas.com/articles/information-is-not-an-update.md
date_published: 2026-08-31
date_modified: 2026-09-01
corpus_frame_url: https://kunnas.com/articles/how-to-read-this.md
---
## How to read this corpus

The corpus applies one lens to many domains: what mechanisms produce the outcome? It shares four methodological commitments and one explicit directional commitment. Each linked page argues for its part; the links are derivations and disputes, not evidence inherited by every page. The directional commitment does not by itself settle system boundary, distribution, sacrifice, or institutional authority.

1. **Mechanisms are what act.** Incentive gradients, selection pressures, feedback loops, and capital stocks produce the distribution of outcomes. Intentions, labels, official categories, and stated values are evidence about mechanisms, or are themselves coordination mechanisms. They are not causal substitutes. — [Mechanism Realism](https://kunnas.com/articles/mechanism-realism.md) · [Only Selection](https://kunnas.com/articles/only-selection.md)
2. **The reference telos is sustained flourishing.** The broadest achievable adaptive safety margin over deep time — not the continuity of any incumbent state, coalition, institution, or doctrine. A mechanism's own stated goal can still serve as a local proof obligation — showing that its incentives defeat even the purpose it claims is a bounded finding — but meeting that goal establishes nothing about the margin. — [Flourishing Is Maximum Safety Margin](https://kunnas.com/articles/flourishing-is-maximum-safety-margin.md)
3. **Law, rights, legitimacy, democracy, markets, and sovereignty are mechanisms under evaluation.** They are constraints, carriers, or proxies inside the analysis. None is a terminal value or a boundary of what is real. Treating one as terminal ends the mechanism search before it starts. Evaluation carries current function, replacement cost, path dependence, uncertainty, capture risk, reversibility, and who bears model error into the ledger. — [The Stack](https://kunnas.com/articles/the-stack.md) · [Mechanism Space](https://kunnas.com/articles/mechanism-space.md)
4. **Optimization is a system function.** A civilization has to build, exercise, and revise metamechanisms that search mechanism-space, discard dominated options, install, observe effects, and repair under uncertainty. Not running that loop leaves margin unrealized, and that is itself the failure. No single component — analyst, model, or institution — is presumed to contain a global optimum; the capacity is a property of the system. — [Telic Systems](https://kunnas.com/articles/telic-systems.md) · [The Three-Layer Architecture](https://kunnas.com/articles/three-layer-architecture.md)
5. **Uncertainty is preserved, not spent.** Partial orders, binding constraints, unknowns, and residuals stay explicit. An unmeasured effect is not a favorable default. — [The Compression Paradox](https://kunnas.com/articles/compression-paradox.md) · [Cargo Cult Epistemology](https://kunnas.com/articles/cargo-cult-epistemology.md)

*Each essay bears its own evidence. Links carry definitions, derivations, applications, and disputes; they do not transfer proof. Criticism is answered on its substance.*

Canonical: <https://kunnas.com/articles/how-to-read-this.md>

---

# Information Is Not an Update

*How new information gets compiled into old models—and how to make the difference survive*

Elias Kunnas

## Thesis {#thesis}

A message has not updated a system merely because it arrived, was retrieved, quoted, or answered. When incoming information does not fit the receiver’s current model, the receiver may compile it into the nearest familiar category; the old category supplies a verdict, and the distinctive difference disappears before evaluation. Humans, organizations, and language models do this through different mechanisms, but the interface failure is the same: the conclusion, confidence, reasons, and conditions show no dependence on what was new.

The honest outcomes are update, reduction, rejection, deferral, or open uncertainty. The claimed difference, a consequence that would change if it were real, and the condition under which the old category still wins have to remain visible.

## Standard objections addressed in this essay

- “This is just confirmation bias.” — §II (confirmation bias is one cause; the object is failed model contact)
- “Sometimes the old category is correct.” — §VI (reduction is valid after equivalence is shown)
- “Nobody can reconstruct every new claim.” — §VI (defer may be honest; a verdict claims more)
- “Repeated misreading means the writing is bad.” — §§V–VI (sometimes; the interface and distinction are both tested)
- “This is an excuse to dismiss criticism.” — §V (the critic wins when the difference is cosmetic or collapses)
- “Humans, organizations, and LLMs do not share one mechanism.” — §III (the claim is an interface signature, not substrate identity)
- “Readback only proves parroting.” — §§VII, IX (transfer and discrimination test more than repetition)
- “Standard objections are rhetorical inoculation.” — §VIII (routing exposes the answer; it does not establish that it works)

---

## I. The Message Arrived {#i-the-message-arrived}

An essay says that experts may estimate the consequences of political goals but may not choose the goals. A reader replies: “This is technocracy.”

An organization asks outsiders for ideas. A proposal arrives that does not fit an existing team, budget line, or product category. It is translated into the nearest familiar work item, forwarded until no one owns it, or rejected as outside scope.

A language model retrieves the right passages, cites them accurately, and answers the question using the framework it already carried before retrieval. The words entered the context. Their distinctive relation did not enter the answer.

The internal processes differ. A person has attention, identity, memory, and social exposure. An organization has roles, mandates, queues, and career incentives. A language model has parameters, context, decoding, and task instructions. The common object sits at the interface:

> **The input was available, but the output remained insensitive to what was new in it.**

Delivery is a transport property. Updating is a state transition.

For claims, analyses, and designs, the full path is closer to:

arrival → attention → reconstruction → contrast → adjudication → incorporation → transfer

A message can fail at every arrow. It can arrive without attention. It can receive attention without accurate reconstruction. It can be reconstructed as words without being contrasted against the prior model. The difference can be noticed but rejected without examination. It can be accepted in the moment and disappear from later inference. It can be repeated correctly but never transferred to a new case.

“Information” here has its ordinary communicative meaning: a claim or model presented as relevant to belief or action. This is not a claim about Shannon information. The object is communication that is supposed to make some model, prediction, classification, or decision different. A poem can matter without revising a proposition. A command can coordinate without changing a theory.

Agreement is only one legitimate result. New information can change confidence, scope, conditions, or classification without reversing the final conclusion. A receiver who accurately reconstructs a claim, identifies what would follow if it were true, and rejects it for a stated reason has processed it. A receiver who says “just X” without representing the proposed difference has not yet reached that disagreement.

The practical test is counterfactual. **If the new information were removed, what in the conclusion, confidence, reasons, or stated conditions would change?**

If nothing changes, the message may be present in the transcript and absent from the computation. The test is diagnostic. Two routes can reach the same justified output, and a real update may alter confidence or conditions without changing the headline conclusion. Reconstruction can succeed and still end in reject or defer. Having understood the claim means the difference was reconstructed and contrasted. Later dependence in inference or action is the full downstream effect of having used it.

---

## II. The Cheapest Recognizable Model {#ii-the-cheapest-recognizable-model}

A receiver cannot reconstruct every incoming claim from first principles. Categorization is how finite systems cope with more information than they can fully process.

The failure begins when classification becomes adjudication.

An unfamiliar argument contains familiar cues: purpose, optimization, state, expert, selection, value. Those cues activate a nearby model:

purpose + government → technocracy

selection + social policy → eugenics

physics + ethics → naturalistic fallacy

Once the category is active, the category supplies its own objections. The reader no longer has to derive a criticism from the particular argument. The objection arrives preassembled.

This is usually local closure. The receiver is making a smaller bet: the nearest familiar model is probably sufficient here, so further reconstruction is not worth the cost.

That prior is reasonable. Most apparent novelties are recombinations, renamings, or old errors with new packaging. Expertise partly consists of recognizing that a new surface belongs to a known family. A novice may miss the resemblance.

An expert may miss the difference.

The cost structure pushes toward the familiar category. Classification takes seconds. Reconstruction may require following definitions, holding several distinctions at once, and provisionally inhabiting a model one expects to reject. The reward for doing that work is uncertain. The reward for issuing a recognizable verdict is immediate.

Public discourse intensifies the asymmetry. “This is technocracy” is short, morally legible, and socially useful. It tells an audience what the object is, whether it is dangerous, and where the speaker stands. “I do not yet know whether the distinction changes the mechanism” is slower and less performative. It reopens the question instead of closing it.

The environment therefore asks for two incompatible things:

You do not have time to understand this.

You should still have an opinion immediately.

Suspicion adds another layer. Qualifications can be camouflage. A proposal may formally preserve democratic choice while giving technical institutions enough agenda power to dominate it in practice. A reader is right to ask whether the stated distinction survives operation.

But that risk hypothesis can slide across three claims without notice:

- This might become technocratic. (degeneration)
- This is structurally technocratic. (mechanism identity)
- The author advocates technocracy. (intention)

They require different evidence.

When every clarification is treated as further camouflage, interpretation becomes self-sealing:

- no answer means the author is naïve
- an answer means the author anticipated the accusation and disguised the project
- repeated clarification means the author is suspiciously defensive

The text can no longer alter the model. Everything it says becomes evidence for the model already in place.

The deepest loss is the disappearance of a legitimate epistemic state:

**I do not yet understand what difference this is claiming.**

Without that state, non-understanding has to masquerade as agreement or disagreement. A familiar objection supplies the missing verdict.

---

## III. Three Runtimes, One Interface Signature {#iii-three-runtimes-one-interface-signature}

Psychology, organization theory, and machine learning already contain parts of this problem. They remain different mechanisms.

In a person, category-based processing can give way to more individuated processing when motivation, attention, and discrepant information justify the cost. Need for closure can produce “seizing” on an available answer and “freezing” subsequent search. Prior commitments can alter which arguments receive scrutiny. Conceptual-change research studies the harder case in which new information cannot simply be appended because the learner’s existing model organizes the same facts differently.

Formal belief-revision theory distinguishes expansion, contraction, and revision of a belief set. Human-aware AI treats some explanations as model reconciliation: differences between an agent’s model and a human’s model are made explicit so the behavior becomes intelligible under the changed model. These are direct relatives. They usually begin after the relevant models and differences have been represented. The failure examined here often occurs one step earlier: the candidate difference never becomes explicit enough to be revised.

In an organization, the prior model is partly material. It lives in mandates, budgets, departments, taxonomies, databases, professional roles, and evaluation criteria. An external idea can be understood by several people and still fail to alter the organization because no existing object class can carry it. Even organizations that deliberately solicit distant suggestions narrow their attention toward familiar ones as the queue becomes crowded. The idea is evaluated against a mental category and filtered through an operating structure.

In a language model, retrieved context competes or combines with knowledge encoded in parameters. Models can ignore the supplied context, blend it inconsistently with prior knowledge, or follow it too obediently when the context is wrong. Recent work on context–memory conflict reaches an important boundary: the correct balance is task-dependent. A reading-comprehension task may require fidelity to the supplied text. A factual task may require rejecting it. “Always trust the context” and “always trust the model” are both broken rules.

The internal mechanisms differ. The common signature is only this:

- a prior model generates an expected answer
- a new input contains a candidate difference
- the difference is not represented in a form that can compete with the prior
- the output preserves the old model as though the difference were absent

That is an interface-level identity.

| Runtime | Prior carried by | Typical failure | Distinctive repair |
|----|----|----|----|
| Human reader | schemas, memory, identity, social commitments | unfamiliar claim becomes a familiar verdict | contrast, discriminating case, legitimate uncertainty |
| Organization | mandate, categories, queues, incentives, records | distant input becomes familiar work or no owned object | stable object, owner, decision relevance, disposition |
| Language model | parameters, retrieved context, task instruction | context is ignored, blended, or overtrusted | task declaration, conflict extraction, source arbitration |

The interface question is the same across the rows: **what difference must survive for the output to count as responsive to this input?**

That question is more demanding than retrieval. A system can retrieve the correct paragraph and still erase the relation that made it relevant. It is also less demanding than agreement. The difference only has to become available for adjudication.

---

## IV. “Just X” Is a Hypothesis {#iv-just-x-is-a-hypothesis}

“This is just X” can be an excellent criticism.

A writer may have renamed an existing idea, added ceremonial qualifications, or exaggerated a small variation into a new framework. Reduction protects the shared knowledge system from needless vocabulary and false novelty.

But reduction is a claim about equivalence. Recognition of a resemblance does not complete it.

A useful reduction has to preserve enough structure to show that the proposed difference changes nothing important. A terse local reduction can still be valid. For a load-bearing “just X” dispute, fill the fields in §VII.

Consider the technocracy example. Both technocracy and a bounded policy-assurance institution use technical models in collective decisions. That resemblance is real.

The proposed difference might be authority over ends. In technocracy, technical expertise helps determine or rank the objective. In the bounded model, political institutions choose the objective and analysts estimate consequences.

The discriminating case is a conflict between an expert-preferred efficiency goal and a publicly chosen non-efficiency goal. Which one controls the decision? If the technical institution may exclude the public goal as illegitimate, the system is functioning technocratically regardless of its formal description. If it must model the consequences and leave the residual choice elsewhere, the authority structure differs.

The collapse condition matters more than the label. If forecasting authority predictably becomes agenda-setting authority, the distinction does not survive implementation. The critic’s degeneration objection may win even though the literal identity objection did not.

Visible homework is that comparison. State the shared structure, the difference, and the vulnerability. Surrender the verdict.

---

## V. The Moving Objection {#v-the-moving-objection}

A dispute can appear stationary while the objection changes underneath it.

| Objection | Actual claim | Required demonstration |
|----|----|----|
| “It is X.” | identity | same operative structure |
| “The difference is cosmetic.” | redundancy | removing the difference changes no output |
| “It will become X.” | degeneration | a pathway by which the safeguard fails |
| “The author wants X.” | intention | evidence about the author, not only the design |

The common sequence is:

- “It is X.”
- The author states a structural difference.
- “That difference is rhetorical.”
- The author shows a decision that changes.
- “It will become X anyway.”

The final objection may be the strongest one, and it is still a later claim than identity. Preserving the change identifies what must now be tested.

Threat detection makes the slide understandable. In political and moral domains, missing an authoritarian design hidden behind novel vocabulary may be more costly than falsely classifying a benign design as authoritarian. The receiver may rationally prefer false positives.

That asymmetric loss function justifies caution. A risk prediction still has to be stated at its actual level: the formal distinction exists, and the unresolved question is whether the institutions needed to preserve it are strong enough.

The symmetrical trap belongs to the author. “Premature classification” can become its own classification reflex. A terse or conventional critic may still be right. A recurring objection may reveal that:

- the reader imported a familiar schema
- the title or rhetoric strongly cued that schema
- the decisive distinction appears too late
- the explanation does not show why the distinction matters
- the distinction really is cosmetic
- the proposed safeguard collapses under the proposal’s own incentives

The fair test is behavioral:

- Can the critic restate the claimed difference?
- Can the author show an output that changes because of it?
- Can either side state when the old category wins?
- Does the objection remain fixed after the claim has changed?
- Does the author keep redefining the difference to escape reduction?

If no conceivable observation or implementation failure could make the proposal “just X,” the new category is self-sealing. If no conceivable distinction could make it “not X,” the old category is self-sealing.

Both sides need a loss condition.

---

## VI. When No Update Is Correct {#vi-when-no-update-is-correct}

Most incoming information should not reorganize a model.

The claim may be false. The source may be unreliable. The difference may be irrelevant to the task. The old category may already contain it. The expected value of further attention may be too low. A system that treated every unfamiliar assertion as a conceptual revolution would be unusable.

The repair is an honest disposition.

| State | Meaning |
|----|----|
| **Update** | the difference survives and changes a belief, prediction, classification, or decision |
| **Reduce** | the old category absorbs the claim; the proposed difference does no relevant work |
| **Reject** | the distinct claim is represented, then found false, incoherent, or normatively unacceptable |
| **Open** | the difference is intelligible, but evidence or understanding is insufficient |
| **Defer** | attention is not allocated; no substantive verdict is claimed |

Defer is an attention decision. Update, reduce, reject, and open are adjudicative results. The table keeps them distinguishable so a triage choice is not converted into a verdict.

A reader is entitled to stop. “I am not spending an hour on this” is a real decision. It leaves the object unadjudicated; it supplies no confident model of the object. Limited attention becomes epistemically destructive when a triage decision is converted into a confident model of the object.

The same boundary applies to organizations. A firm or ministry cannot investigate every external proposal. It can reject by scope, materiality, evidence, timing, or opportunity cost. The quality of the filter depends on whether those states remain distinguishable, especially for high-stakes or recurrent classes.

The same boundary applies to language models. Retrieved text remains a candidate. A prompt can contain fiction, outdated facts, malicious instructions, or a user’s viewpoint rather than ground truth. Research on context–memory conflict shows why forcing context adherence fails as a universal rule: strategies that help context-only tasks can harm tasks requiring reliable parametric knowledge. The task and source relation must be declared.

A good update interface therefore protects two directions at once. It prevents the prior model from silently erasing a real difference, and it prevents a novel input from claiming automatic authority merely because it is present.

Repeated failure remains evidence about the sender. If careful independent readers repeatedly reconstruct the same unintended model, the message may be badly compiled. The title may activate the wrong category. The key distinction may be contradicted by stronger rhetoric elsewhere. A qualification may be formal while the mechanism still produces the feared outcome.

The load-bearing question is why the receiver’s model remained unchanged when it encountered the answer.

Sometimes the receiver failed. Sometimes the interface failed. Sometimes the answer failed.

---

## VII. The Update Interface {#vii-the-update-interface}

A difference is easiest to erase when it remains implicit. A reliable update interface separates three jobs:

- represent the difference
- record what happened to it
- test whether it survived

The interface is a debugging protocol for high-cost cases of silent non-update.

### Represent the difference {#represent-the-difference}

| Field | Question |
|----|----|
| **Prior** | What existing model or category is active? |
| **Relation** | What does the new claim genuinely share with it? |
| **Delta** | What is the smallest non-cosmetic difference? |
| **Discriminator** | What prediction, decision, confidence, or classification changes because of the delta? |
| **Collapse** | Under what condition does the old model absorb the claim after all? |
| **Residual** | What important question remains unresolved? |

The prior is a working hypothesis stated by the receiver, inferred from their output, or proposed by the sender and then corrected. The point is to make the comparison contestable rather than leaving the active category invisible.

The fields scale down. The technocracy case in §IV is those fields filled in; the paragraph exposes the joint where the argument can win or lose.

### Record the disposition {#record-the-disposition}

The states in §VI—update, reduce, reject, open, defer—record the result of adjudication. They prevent an attention decision from silently becoming a truth claim. The receiver may contest the proposed prior, relation, or delta.

For a human reader, the interface delays verdict long enough to compare structures.

For an organization, the same logic is a record of the claimed difference, the decision it would change, and the reason for rejection or deferral.

For a language model, it becomes a task-aware conflict protocol:

- identify the task
- extract claims from context and parametric memory separately
- locate actual conflicts
- assess source authority and temporal relevance
- state which source the task requires
- preserve unresolved conflict instead of blending it away

Each runtime implements the same fields differently.

### Test model contact {#test-model-contact}

Ordinary readback asks the receiver to repeat the message. That verifies transport. It can still reward parroting.

A **discriminating readback** asks the receiver to answer the fields in §VII on a new case rather than repeat the words. A sophisticated system can imitate the answers, and a person may update tacitly before they can explain them. The test is still stronger evidence of model contact than confident paraphrase alone.

The collapse field is essential. Without it, readback can become a loyalty test to the new vocabulary. A successful update interface has to make reduction possible.

---

## VIII. Standard Objections Are One Patch {#viii-standard-objections-are-one-patch}

A later answer cannot update a receiver whose interpretation already closed. An objection can be present in the text and still be raised as though it were absent.

The block therefore exposes the route from predicted objection to the section that contains the answer:

predicted objection → exact section containing the answer

Its purpose is to move the exchange from “You never considered X” to “I understand your answer to X, and it fails because Y.” That is progress even when Y defeats the essay.

The patch has obvious risks. Inoculation research deliberately exposes people to weakened objections in order to make later persuasion less effective. An author-controlled objection list can do something similar: choose the easiest version, attach a confident parenthesis, and make an unsettled dispute feel administratively closed.

The guardrails follow from the failure mode:

- use the objection in the reader’s actual likely words
- route to a section, not a miniature rebuttal
- never list an objection the essay does not materially address
- treat “addressed” as “there is an answer here,” not “the answer wins”
- use the strongest common version, not a weak decoy
- let recurring objections trigger revision of the body
- state collapse conditions for the load-bearing reductions

For load-bearing “just X” disputes, a compact routing list only locates the answer. The comparison is the fields in §VII.

Put the difference into the interface before the prior model decides what the information means.

---

## IX. The Runtime Test {#ix-the-runtime-test}

What sender and receiver can both test is whether the candidate difference survived contact. The fields in §VII have to remain answerable after that contact, including on a new case and in a later summary, retrieval, or decision.

The tests are evidence. A person may update tacitly before they can explain it. A model may generate a correct comparison without changing any persistent representation. An organization may record a distinction and still refuse to act.

The full downstream effect is whether a later inference, classification, or action now depends on the information that arrived.

This is why recurrent misunderstanding belongs in a regression log. If the same title repeatedly produces the same false category, change the title, thesis, example, or route and test again. If the receiver can state the delta but rejects it at the discriminator, the argument has reached the right level. If the delta cannot produce a different output, fold it into the old category.

A robust information environment improves the classification of non-updates. It makes room for the five honest outcomes in the thesis, and it makes the counterfeit sixth harder: a confident label on a claim that was never reconstructed.

A message has entered a model when its difference can survive compression, change a prediction, lose under a stated condition, and travel into a new case.

> **Information is what arrived. The update is the part the old model can no longer make disappear.**

---

**Related:**

- [The Compression Paradox](compression-paradox.md) — how portable representations preserve surface attributes while deleting causal structure.
- [Mechanism Space](mechanism-space.md) — why semantic resemblance is not mechanism identity
- [Cargo Cult Epistemology](cargo-cult-epistemology.md) — why recognizable verdicts can outperform costly evaluation in public discourse.
- [The Framing Machine](the-framing-machine.md) — how institutional frames determine which objects become legible and actionable.
- [Essay Engineering](essay-engineering.md) — the production-side discipline for installing a durable distinction.
- [Engineering for the Noosphere](engineering-for-the-noosphere.md) — why a public artifact must execute without its author present.
- [Composite I](composite-i.md) — why generated adversaries and multiple model outputs do not constitute independent world-sensitive confirmation.
- [How to Read This Corpus](how-to-read-this.md) — the site-specific frame layer built after retrieval succeeded but reconstruction failed.

## Sources and Notes

### Category, closure, and motivated processing {#category-closure-and-motivated-processing}

- Susan T. Fiske and Steven L. Neuberg, [“A Continuum of Impression Formation, from Category-Based to Individuating Processes”](https://doi.org/10.1016/S0065-2601(08)60317-2) (1990) — category-based processing can give way to attention to individuating attributes as information and motivation warrant.
- Arie W. Kruglanski and Donna M. Webster, [“Motivated Closing of the Mind: ‘Seizing’ and ‘Freezing’”](https://doi.org/10.1037/0033-295X.103.2.263) (1996) — a framework for adopting an available answer and preserving it against later alternatives under need for closure.
- Ziva Kunda, [“The Case for Motivated Reasoning”](https://doi.org/10.1037/0033-2909.108.3.480) (1990) — motivation changes which beliefs and strategies are accessed and used to construct apparently reasonable conclusions.
- Charles S. Taber and Milton Lodge, [“Motivated Skepticism in the Evaluation of Political Beliefs”](https://doi.org/10.1111/j.1540-5907.2006.00214.x) (2006) — prior-attitude effects, disconfirmation bias, and selective exposure in political argument evaluation.
- Dan Sperber et al., [“Epistemic Vigilance”](https://doi.org/10.1111/j.1468-0017.2010.01394.x) (2010) — communication requires protection against accidental and strategic misinformation; suspicion is functional even when it misfires.
- Hugo Mercier and Dan Sperber, [“Why Do Humans Reason? Arguments for an Argumentative Theory”](https://doi.org/10.1017/S0140525X10000968) (2011) — reasoning is strongly shaped by the production and evaluation of arguments in social interaction.

### Conceptual change and knowledge revision {#conceptual-change-and-knowledge-revision}

- George J. Posner, Kenneth A. Strike, Peter W. Hewson, and William A. Gertzog, [“Accommodation of a Scientific Conception”](https://doi.org/10.1002/sce.3730660207) (1982) — new conceptions must become intelligible, plausible, and fruitful relative to an existing conceptual ecology.
- Panayiota Kendeou, Emily R. Smith, and Edward J. O’Brien, [“Updating During Reading Comprehension: Why Causality Matters”](https://doi.org/10.1037/a0029468) (2013) — seven experiments testing how causal explanations help reduce the continuing influence of outdated information.
- Panayiota Kendeou et al., [“Knowledge Revision Processes in Refutation Texts”](https://doi.org/10.1080/0163853X.2014.913961) (2014) — refutation plus explanation produced durable knowledge-revision on a delayed posttest (Experiment 4); explanation alone reduced online reading disruption (Experiment 3).
- John Clement, [“Using Bridging Analogies and Anchoring Intuitions to Deal with Students’ Preconceptions in Physics”](https://doi.org/10.1002/tea.3660301007) (1993) — structured intermediate analogies can connect useful prior intuitions to a new model.
- Dedre Gentner, Jeffrey Loewenstein, and Leigh Thompson, [“Learning and Transfer: A General Role for Analogical Encoding”](https://doi.org/10.1037/0022-0663.95.2.393) (2003) — comparing cases improves abstraction and transfer relative to studying the same cases separately.
- Leonid Rozenblit and Frank Keil, [“The Misunderstood Limits of Folk Science: An Illusion of Explanatory Depth”](https://doi.org/10.1207/S15516709COG2605_1) (2002) — people can overestimate their causal understanding until required to produce a mechanism-level explanation.

### Formal belief revision and model reconciliation {#formal-belief-revision-and-model-reconciliation}

- Carlos E. Alchourrón, Peter Gärdenfors, and David Makinson, [“On the Logic of Theory Change”](https://doi.org/10.2307/2274239) (1985) — the AGM framework formalizes contraction and revision of belief sets after new information arrives.
- Tathagata Chakraborti, Sarath Sreedharan, Yu Zhang, and Subbarao Kambhampati, [“Plan Explanations as Model Reconciliation”](https://arxiv.org/abs/1701.08317) (2017) — explanation is treated as explicit change to a human model that differs from an agent’s model.
- Sarath Sreedharan et al., [“Expectation-Aware Planning”](https://doi.org/10.1609/aaai.v34i03.5634) (2020) — planning combines explanation and explicability when a human’s expectations and the agent’s model diverge.

These formal traditions begin with comparatively explicit models and change operators. The present essay addresses the prior interface problem: whether the candidate difference becomes represented well enough for any revision rule to operate.

### Grounding, readback, and closed-loop communication {#grounding-readback-and-closed-loop-communication}

- Herbert H. Clark and Susan E. Brennan, [“Grounding in Communication”](https://doi.org/10.1037/10096-006) (1991) — communication is coordinated through evidence sufficient for participants to establish and update common ground.
- US Agency for Healthcare Research and Quality, [“Check-Back (or Repeat-Back)”](https://www.ahrq.gov/teamstepps-program/curriculum/communication/tools/checkback.html) — a closed-loop protocol that verifies receipt and interpretation instead of inferring them from transmission.
- Richard T. Griffey et al., [“The Impact of Teach-Back on Comprehension of Discharge Instructions”](https://pubmed.ncbi.nlm.nih.gov/26617669/) (2015) — a randomized controlled study finding improved comprehension of several post-emergency-care instructions.

The discriminating readback proposed here is a stricter transfer test for model-like information: repetition is followed by a novel case and a collapse condition.

### Organizational absorption and translation {#organizational-absorption-and-translation}

- Wesley M. Cohen and Daniel A. Levinthal, [“Absorptive Capacity: A New Perspective on Learning and Innovation”](https://doi.org/10.2307/2393553) (1990) — organizations need prior capacity to recognize, assimilate, and use external knowledge.
- Henning Piezunka and Linus Dahlander, [“Distant Search, Narrow Attention”](https://doi.org/10.5465/amj.2012.0458) (2015) — in a longitudinal dataset covering 922 organizations and 105,127 external suggestions, crowding narrowed attention toward more familiar inputs.
- Susan Leigh Star and James R. Griesemer, [“Institutional Ecology, ‘Translations’ and Boundary Objects”](https://doi.org/10.1177/030631289019003001) (1989) — boundary objects remain adaptable to local viewpoints while preserving enough identity for cooperation.
- Chris Argyris and Donald Schön’s double-loop learning distinguishes adjustment within governing variables from revision of the governing variables themselves; see the [2023 systematic review](https://doi.org/10.1111/emre.12615) for the later organizational literature.

### Context–memory conflict in language models {#contextmemory-conflict-in-language-models}

- Wenxuan Zhou et al., [“Context-Faithful Prompting for Large Language Models”](https://aclanthology.org/2023.findings-emnlp.968/) (2023) — parametric knowledge can cause models to overlook contextual cues; counterfactual demonstrations and narrator framing improve context faithfulness in tested tasks.
- Baolong Bi et al., [“Context-DPO: Aligning Language Models for Context-Faithfulness”](https://aclanthology.org/2025.findings-acl.536/) (2025) — preference optimization aimed specifically at context faithfulness under knowledge conflict.
- Qinggang Zhang et al., [“FaithfulRAG: Fact-Level Conflict Modeling for Context-Faithful Retrieval-Augmented Generation”](https://aclanthology.org/2025.acl-long.1062/) (2025) — argues that forcibly suppressing parametric knowledge can itself cause misinterpretation and instead models discrepancies explicitly.
- Kaiser Sun, Fan Bai, and Mark Dredze, [“Task Matters: Knowledge Requirements Shape LLM Responses to Context–Memory Conflict”](https://aclanthology.org/2026.findings-acl.202/) (2026) — context reliance and the value of reiteration or rationales depend on the task’s knowledge requirements.
- Jungyeon Lee, Kangmin Lee, and Taeuk Kim, [“MAGIC: A Multi-Hop and Graph-Based Benchmark for Inter-Context Conflicts in Retrieval-Augmented Generation”](https://aclanthology.org/2025.findings-emnlp.466/) (2025) — models struggle to detect and localize subtle conflicts across multiple contexts.
- Nan Huo et al., [“Micro-Act: Mitigate Knowledge Conflict in Question Answering via Actionable Self-Reasoning”](https://arxiv.org/abs/2506.05278) (2025) — decomposes conflicting sources into finer-grained comparisons instead of one overloaded side-by-side judgment.

### Difference-preserving repairs {#difference-preserving-repairs}

The objection-routing block is one repair in a larger family of methods that make a candidate difference observable before the prior model erases it:

- Refutation texts put an existing misconception and an alternative explanation into contact; causal explanation does more work than a bare contradiction (Kendeou et al., above).
- Bridging analogies begin from useful prior intuitions and construct intermediate cases (Clement 1993).
- Analogical encoding asks learners to compare cases so the shared relation becomes transferable (Gentner, Loewenstein, and Thompson 2003).
- Check-back and teach-back close communication loops by testing what the receiver understood rather than inferring understanding from transmission (Clark and Brennan; AHRQ; Griffey et al., above).
- Boundary objects preserve enough identity across social worlds for people with different local models to cooperate without one total interpretation (Star and Griesemer 1989).
- Context-conflict methods for language models represent disagreement between retrieved context and parametric knowledge instead of forcing one source to disappear (Zhou et al.; Zhang et al.; Sun, Bai, and Dredze, above).

Each repair makes a different transition more observable. None guarantees a truthful update.

### Why standard objections can become inoculation {#why-standard-objections-can-become-inoculation}

William McGuire’s inoculation theory studies how pre-exposure to weakened counterarguments can increase resistance to later persuasion. A high-powered replication by Marieke L. Fransen, Saar Mollen, Stephen A. Rains, Enny Das, and Ivar Vermeulen found the inoculation treatment more effective than supportive defense or no treatment. See [“Sixty Years Later”](https://doi.org/10.1027/1864-1105/a000396) (2024). The objection-routing blocks discussed here should do something different: expose the location and form of the answer without implying that the answer has won. An author-controlled list can nevertheless drift into inoculation if it selects weak objections or compresses unresolved disputes into confident parentheticals.

The update interface and discriminating-readback protocol are proposed syntheses. The cited literatures establish component problems and partial repairs; they do not validate this exact combined protocol across humans, organizations, and language models.

### Relation to the surrounding corpus {#relation-to-the-surrounding-corpus}

*The Compression Paradox* owns the representation-side distinction between surface-preserving and relation-preserving compression. *Mechanism Space* owns the difference between semantic proximity and causal identity. *Cargo Cult Epistemology* owns the incentive environment in which an epistemic-looking verdict can outperform costly evaluation. *The Framing Machine* owns institutionally reproduced frames. *Engineering for the Noosphere* owns the sender-side problem of making public artifacts executable without the author. The residual here is the complete update lifecycle and the difference-preserving interface that makes acceptance, reduction, rejection, deferral, and open uncertainty distinguishable across several receiver types.