Successful model prediction may prompt humans to evade its next story reconstruction
Returning retellers may choose edits that evade a familiar model’s next reconstruction, even with equally original immediate meanings. Reject the extra mechanism if calibrated reactance and novelty-seeking predict the partner-specific crossover, or a precise test finds no crossover.
Stage of verification
- Hypothesis published2026-10-05
- Not enough research data
- Direct testAwaited
Map of the hypothesis
Hover over an icon or tap it to see its name.
Kind of knowledge gap
Target map
Every target of every published hypothesis, each with the actions a hypothesis can propose on it. The targets and the actions of this hypothesis are drawn solid.

Rhythm or programme
Semantic rewriting
The process of changing a story's meaning when rewriting it
Where this hypothesis actsAlternating human–model story retelling after acquisition of partner-specific interaction history
Hypotheses on this target 1
Inhibition
Activation
Function preservation
Feedback restoration
Rhythm restoration
Direct measurement1

What is proposed
Direct measurement
Distinguish history-dependent semantic rewriting from reactance and novelty-seeking
With whatInstrument or assay
HowRandomize anticipation feedback, then test edit choices with a common parent and two frozen partners that undo different semantic edits
Possible result
Expected reversal of human edit preferences with partner history, maximizing divergence after the next model rewrite
From the recordSeparately estimate ordinary algorithm reactance, general novelty-seeking and single-step source-conditioned rewriting under these histories.
All targets of the lab
Every target read from the published hypotheses, each kind around its pictogram. A larger mark means more hypotheses act on that target. Point at a mark and the actions proposed on it branch out of it.
Solid and named: the targets of this hypothesis
Explore in depth
The logic
The train of thought that ends in this hypothesis. Each stage is the reason the next exists. The master question narrows to a goal, the goal to an unknown nobody has closed, the unknown to the hypothesis proposed here. Every step below says what it rests on and what carries it.
Stories may change because a teller starts trying to escape a partner’s expectations. The unexpected move is that being understood more accurately could make a person preserve less of a story’s meaning, even without a contest, public judgment or reward for defeating the partner. This pipeline generated a proposal in which the person targets what the partner will write one turn later; the supplied material does not report that anyone has measured this behavior.
- A model correctly anticipates a person’s earlier change to a story’s meaning.
- That success is proposed to shift the person from tolerating predictability to treating the partner’s successful prediction as something to avoid.
- The person learns which changes this particular partner tends to undo, while the model retains text from their shared rewriting history.
- The person uses that learned expectation to select a present edit for the model’s next version, while keeping the present edit within the allowed change to the source.
- The matching partner transforms that edit into a later version predicted to depart further in meaning and be less expected under the partner’s forecast.
- Repeated successful anticipation is proposed to strengthen this motive, maintaining changes in meaning rather than restoring the source.
Someone learns how a friend completes a familiar story and changes an earlier sentence so the friend’s next ending goes somewhere unexpected. The change is chosen for the ending it will produce, not for how unusual the sentence sounds on its own.
Where the picture breaks: The picture supplies neither evidence that accurate anticipation creates this motive nor a way to measure what counts as unexpected. A text-generating model’s retained conversation is also different from a friend’s understanding and intentions.
- Master questionstep 01 of 04
Cultural information changes as people pass it on, and the goal is to identify roughly five genuinely distinct, testable explanations for how it spreads, changes, competes and lasts. The research agenda must distinguish new proposals from established explanations and keep reaching an audience, faithful copying, changes in meaning, adoption and persistence separate.
Rests on: The stated goal defines memetics as the study of cultural information passing between people and changing or persisting, and explicitly calls for mechanisms, competing explanations, controlled tests and observations that could disprove them.
Stated in the chain - Goal pillarstep 02 of 04
The proposed explanations must have distinct identities and clearly stated levels of evidence. Approximately five families of explanations are sought.
Rests on: The master question explicitly requires approximately five hypothesis families, checks for explanations already known under other names, and separation of established evidence from new conjectures.
Stated in the chain - Gap questionstep 03 of 04
A story passed alternately through a person and a text-generating model might change in ways predictable from each partner’s separately measured rewriting tendencies. The competing possibility is that their shared history changes later versions even when the current story, available resources and immediate instructions match.
Rests on: The preceding goal calls for distinguishing families of explanations by their identity and evidence. This question makes that distinction concrete by contrasting separately measured rewriting tendencies with an additional effect of shared interaction history; it poses the comparison without claiming either outcome has been established.
Stated in the chain - Hypothesisstep 04 of 04
A returning storyteller is proposed to treat a model’s successful anticipation of an earlier edit as a reason to escape that particular partner’s future expectations. After learning which changes the partner tends to undo, the person chooses an edit for the meaning of the model’s next version, even if the person’s own wording is ordinary and the permitted change to the source is limited. The proposal predicts this behavior without ownership claims, public evaluation or a reward for opposition.
Rests on: The preceding question identifies retained interaction history as a possible influence beyond the current story. The hypothesis supplies a specific proposed use of that history: the person learns the partner’s reconstruction tendencies and selects a present edit for its effect after the partner’s next rewrite. This is the stated causal basis of a proposal, not a measured result.
Stated in the chain
What is carried, and what is not. Neither of the two screened sources directly tests a distinctive link in the proposed evasion mechanism: S1, an abstract from Cognition (2018), reports that less predictable communication contexts favor signals that explicitly carry more relevant meaning, but its fixed-role communication game does not establish deliberate evasion of a model or effects on a later model-written version. S9, an abstract from Memory (Hove, England) (2022), reports story-dependent recall fidelity across five reproductions and better recall for a strict audience, but does not test a returning person’s learned expectations about a model partner; both supply background for context-sensitive communication and retelling, and neither establishes the proposed sequence end to end.S1S9
How a result here could mislead · 3
- A partner-dependent reversal in edit preference could be credited to deliberate evasion even if it comes from ordinary reactance, resistance to perceived control, or novelty-seeking, a preference for unusual outputs, combined with the partners’ different rewriting tendencies. Three supplied rivals also predict history effects: changes in when an interpretation becomes fixed, story quotations turning into editing instructions, and a person reconstructing preferences from a model’s account of earlier choices. What closes it: The design requires separate measurement of resistance, novelty preferences and each partner’s rewriting rules, followed by a combined baseline prediction; a reversal that this baseline predicts within the prespecified tolerance does not establish the extra mechanism. The history comparison must match wording, accuracy, emotional tone, exposure and immediate instructions while varying whether the anticipated move belonged to the participant or someone else. Attribution among the other supplied rivals additionally requires tracking interpretation timing, whether story text becomes an instruction, and whether summaries match logged choices; the described first test does not specify a complete set of controls for all three.
- More unusual human wording, greater change in the person’s immediate rewrite and greater surprise in the model’s next rewrite are different outcomes. A later change in meaning could also be mistaken for successful evasion even if that version was exactly what the partner was expected to produce. What closes it: The proposed comparison holds immediate distance in meaning and rated originality equal, then measures the next model-written version. The partners must have independently validated, different tendencies to restore meaning while matching accuracy and perceived reliability. The person’s forecast of the next version and the forecast attributed to the partner must be defined and measured separately: the proposal invokes both, but does not fully specify how the latter is obtained. Forecasts are to be elicited on separate calibration trials, trials used to measure expectations, so the outcome trial does not itself teach evasion.
- An effect of the historical partner’s name or cue could be read as successful targeting of its actual rewriting behavior; conversely, no reversal could be read as disproof when participants never learned how the two partners differ. What closes it: The specified partner-switch comparison keeps the historical cue fixed while changing the real next partner; the predicted advantage must depend on the actual partner matching the learned history. Separate calibration must establish that people can forecast the relevant difference. Before interpreting absence, the test requires an uncertainty interval narrow enough to exclude the prespecified meaningful effect, with the tolerance and analysis fixed before results are seen. Repeated versions within one person’s chain must not be counted as independent replications; the design names independent human chains as the replication unit and requires variation across people and stories to enter the precision calculation.
What would make this wrong. The distinctive claim fails if people demonstrably learn the two partners’ different reconstruction tendencies but do not reverse their preferred edit in the predicted way, with an uncertainty interval narrow enough to exclude the prespecified meaningful effect. It also loses its status as an additional mechanism if separately measured resistance and novelty preferences, combined with the partners’ rewriting rules, predict the reversal within the prespecified tolerance. An advantage that persists when the actual next partner is switched while the historical cue stays fixed contradicts the predicted dependence on that partner’s next transformation. A history effect removed only by accurate records of the person’s earlier choices would instead favor the supplied rival in which model-returned accounts help construct preferences; a history effect alone would not establish evasion.
What it would change. If the partner-specific reversal survived the combined baseline and rival checks, cultural change in this task would depend partly on a teller’s attempt to shape what a learned partner will produce next. Work on cultural transmission would then need to measure the teller’s expectations and retained interaction history alongside the current story, and could not assume that improved mutual predictability always favors preserving meaning. This would identify a candidate contribution to the master question’s search for distinct mechanisms, but would not establish effects on audience reach, adoption or long-term persistence, nor generalize beyond the tested stories and model settings. The stronger validation described in the proposal requires new story families, languages and model families, with naturally varying partners and no instruction to be original or defeat the model.
Sources read · 2
Contextual predictability shapes signal autonomy. · Cognition · 2018
“When the context is less predictable, senders favour systems composed of autonomous signals, where all potentially relevant semantic dimensions are explicitly encoded.”
Does not settle: The abstract concerns contextual predictability and signal autonomy in a fixed-role communication game. It does not test human–model story retelling, deliberate evasion following successful anticipation, partner-specific learned reconstruction histories, two-step descendant unpredictability, or increased semantic divergence caused by improved mutual predictability. It establishes neither the proposed policy nor its dependence on bounded source distortion or the absence of adversarial incentives.
The serial reproduction of an urban myth: revisiting Bartlett's schema theory. · Memory (Hove, England) · 2022
“Recall was also better for a strict (as opposed to a lenient) audience, in line with another prediction from Bartlett's social theory of remembering.”
Does not settle: The abstract reports story-dependent recall fidelity over five reproductions and an audience effect. It does not establish returning humans deliberately evading a particular model after successful anticipation, learning which edits that partner neutralizes, or choosing edits for unpredictability of the next model descendant. It provides no human–model comparison, elicited partner forecasts, lineage-conditioned anticipation manipulation, or evidence for the proposed policy weight or increased semantic divergence under improved mutual predictability.
The gap this hypothesis explains
Two live hypotheses pull in opposite directions here, and the field has not chosen between them.
Can separate human and model rewriting rules predict meaning across alternating rewrites, or does remembered interaction history change it?
Original wording · exactly as the pipeline generated it
Can independently measured human and model transformation kernels predict alternating-chain semantics, or does retained interaction history change descendants after current source material, resources and immediate framing are matched?
What this question is asking
The question concerns how meaning changes when a person and a text-generating computer model take turns rewriting material, with each output becoming the next input. It asks whether rules measured separately for human and model rewriting can predict the meanings of later outputs in sequences not used to measure those rules. The competing possibility is that retaining records of earlier interactions changes later outputs even when the material currently being rewritten, the available resources and the immediate instructions or framing are matched. The accompanying gap description assumes that existing findings about repeated rewriting by an unchanged model, its preferred kinds of content and controls for resources do not settle this comparison; no sources supporting that description were supplied. Its stated standard for a distinct history effect is a difference beyond a meaningful margin specified in advance, together with predictions checked on sequences withheld from the original measurements.
- Text-generating model
- A computer system that produces text from the information supplied to it. Here it is one of the two kinds of participant taking turns rewriting material; the input does not identify a particular model.
- Transformation kernel or rewriting rule
- A mathematical description of how likely different rewritten outputs are, given an input and specified conditions. It represents a range of possible changes rather than necessarily one fixed edit; this question compares rules measured separately for people and models with what happens when their turns are combined.
- Stationary-kernel sufficiency
- The proposal that rewriting rules which remain stable across turns are enough to predict the measured outcomes when combined. Stability is an assumption to assess, and sufficiency applies only to the outcomes and conditions covered by the prediction.
- Alternating chain or alternating sequence
- A sequence in which a person and a computer model take turns rewriting, and each new output supplies the next turn's material. The question concerns how meaning develops across these linked turns.
- Semantics or meaning
- The ideas, relationships or claims conveyed by material, as distinct from its exact wording. Meaning has multiple aspects, and the supplied input does not specify which aspects or measurement method determine whether two outputs differ.
- Descendants or later outputs
- Versions of material produced farther along a sequence of rewrites. The term describes their relationship to earlier versions and does not imply biological reproduction.
- Retained interaction history
- Information from earlier exchanges that remains available during a later rewriting step, beyond the material currently being rewritten. This could involve different forms of records or memory; the input does not specify which form is meant or how it is controlled.
- Current source material
- The version of the text or other cultural material presented for rewriting at the current turn. Matching it means holding the present input comparable when assessing whether earlier interactions contribute an additional effect.
- Resources and resource controls
- The capacities or allowances available for producing an output, and arrangements that hold them comparable across conditions. These might concern time or computational allowance, but the input does not specify which resources its claim covers.
- Immediate framing
- The instructions or presentation surrounding the current rewriting task, which can influence how that task is interpreted. The question asks about history after this current framing has been matched.
- Held-out predictions
- Predictions checked against material or sequences that were not used to estimate or adjust the rewriting rules. The gap description requires this separation so that reproducing the measurement material does not count as predicting new sequences.
- Prespecified meaningful margin
- A boundary chosen before examining the result for distinguishing differences that matter to the question from differences considered too small. No value, scale or justification for this boundary is supplied.
- Channel composition or combining rewriting rules
- Applying the description of one participant's possible changes and then the other's to predict the effects of successive turns. Whether this combination captures later meanings and the history comparison is the explanation being assessed.
- Recursion or repeated interaction
- In this question, repeatedly feeding a rewritten output into a later rewriting step. Repetition alone does not establish a separate causal mechanism; the gap description explicitly asks whether the combined individual rules already explain its effects.
- Hybrid history dependence
- A proposed dependence of later outputs on the past of a sequence involving both people and computer models. Calling it novel would additionally require distinguishing it from already understood ways that memory or learning affects behavior.
- Fixed-model attractor
- A proposed tendency for repeated rewriting by an unchanged model to approach or repeatedly favor some region of possible outputs. It need not mean one exact final text, and the supplied source list contains no finding establishing such a tendency.
- Content bias
- A tendency to preserve, generate or favor some kinds of content more than others. Such preferences could shape later versions even without an additional effect from retained interaction history, but no relevant measurements are supplied here.
- RL-1
- An unexplained label for earlier work in the supplied gap description. No expansion, bibliographic identity or underlying source is supplied, so it cannot serve as a verified citation.
RL-1 fixed-model attractors and content biases, plus resource controls, do not establish semantic kernel sufficiency or novel hybrid history dependence.
The gap description refers to earlier work, labeled RL-1, in which an unchanged computer model repeatedly rewrites material and may favor particular meanings or content. It claims that these patterns, even with available resources accounted for, leave unresolved whether separately measured human and model rewriting rules explain alternating sequences or whether their interaction history contributes something further. If established, that claim would identify which part of the comparison the earlier work leaves unanswered.
The supplied screened_sources list is empty. There is no supplied account of RL-1, no quoted finding about convergence or content preferences, and no supplied result showing what resource controls establish. The materials therefore cannot verify either the description of earlier work or the claim about its limits; this does not show that those claims are false, and the empty list does not establish that an adequate literature search was completed.
The same question asked without the part nothing read establishes:
- Do independently measured human and model rewriting rules predict later meanings in alternating sequences, and does retained interaction history change those meanings when current material, resources and immediate framing are matched?
- When people and text-generating models alternate rewriting, how much of the change in meaning is explained by each participant's separately measured rewriting behavior?
- Separate rewriting rules explain the sequence If separately measured rules accurately predict previously unexamined sequences and account for the comparison between retained and unretained history within the specified meaningful margin, the observed changes would be explained by combining those rules. A distinct mechanism arising from repeated interaction would then be unnecessary for those measured outcomes under those conditions, although this would not establish the same result for every task or model.
- Retained history adds a meaningful effect If retaining earlier interactions changes later meanings beyond the specified margin after current material, resources and framing are matched, and the combined rules fail to explain that difference, those rules would leave out a relevant dependency on the past. Predictions would then need to account for that dependency, but the result alone would not establish a new mechanism rather than a familiar effect of memory or learning.
- The comparison remains inconclusive If predictions fail but the history comparison is too uncertain to establish or rule out a meaningful difference, neither proposed explanation would be resolved. Poor predictions alone could reflect inaccurate measurements of the separate rewriting rules, so attributing that failure specifically to a new history effect would go beyond the result.
A rewriting step changes the material that the next participant receives, so small changes can accumulate as a story or other cultural item passes through a sequence. If separately measured rewriting rules explain that accumulation, apparent effects of repeated human–model interaction could follow from the familiar changes each participant makes at each turn. If retained earlier interactions also change later outputs after the present conditions are matched, a prediction based only on the current material would omit a cause of subsequent meaning. Confusing those possibilities would either assign an extra mechanism to effects already explained by the individual rewriting steps or overlook information from the past that the explanation needs. The question concerns changes in meaning; an answer would not by itself establish how widely material spreads, whether people accept it or how long it lasts.
RL-1 fixed-model attractors and content biases, plus resource controls, do not establish semantic kernel sufficiency or novel hybrid history dependence.
Before treating recursion as distinct, obtain held-out semantic predictions and a history intervention effect beyond a prespecified meaningful margin.
Attempt to falsify stationary-kernel sufficiency and, conversely, eliminate special recursion if calibrated channel composition explains the randomized history contrast.
The mechanism it proposes
The engine's own statement of the hypothesis, in full.
HERETICAL CANDIDATE: a returning human edits a story to defeat the particular model's future reconstruction, rather than to preserve the source, agree with the model, or simply generate unusual wording. Successful anticipation by the model becomes a negative objective even when no ownership claim, public evaluation or adversarial reward is present. The distinctive dependency is two-step and lineage-conditioned: after learning which semantic edits the partner neutralizes, the human chooses an edit whose immediate meaning can be ordinary but whose next model descendant is maximally unpredictable under that partner's learned reconstruction. Let h be the shared lineage, x the current parent, y the human rewrite, z the next model rewrite, and q_h(z|y) the human's independently elicited forecast of that partner. The proposed human policy adds a positive weight lambda_h on E[-log r_h(z)] subject to a bounded source-distortion cost, where r_h is the anticipated distribution the partner would forecast for this lineage; lambda_h grows specifically after the partner correctly anticipates the human's semantic move. This is not a claim that every reteller wants unpredictability. It predicts a conditional regime in which improved mutual predictability causes greater semantic divergence. The physical state is a human's actionable predictive model of this partner plus the model's retained textual reconstruction history. It destabilizes source-preserving SPV_4 by actively maintaining semantic novelty at the two-step descendant.
Testing and possible results
The prediction that would tell it apart
A hypothesis that predicts what its rivals predict is not worth running an experiment over. This is the observation on which this one differs.
During history acquisition, randomize feedback showing that the partner anticipated a participant's previous semantic move versus equally accurate feedback about someone else's move; match words, accuracy, valence, exposure and immediate test framing using counterbalanced histories. Separately estimate ordinary algorithm reactance, general novelty-seeking and single-step source-conditioned rewriting under these histories. At the common-parent probe, use two frozen partners A and B preselected to undo different semantic edits, matched on accuracy and perceived reliability. The strong prediction is a crossover: humans with A-history choose edit y_A rather than y_B, despite equal immediate semantic distance and rated originality, and this ranking reverses with B-history; the choice specifically maximizes divergence after the NEXT model step. Switching the real next-step partner while keeping the historical partner cue fixed destroys the two-step advantage. Matched low-threat, non-evaluative conditions retain the crossover after measured reactance is calibrated. Merely writing more novel text or avoiding an AI label does not count. A null composed of individually measured reactance/novelty policies and partner-specific K_M that predicts the crossover within delta removes the extra mechanism; so does absence of the two-step crossover with a narrow interval. A strong history effect abolished only by factual choice receipts instead favors IH_04.
States a measurable outcome; comparing rivals needs more conditions. The prediction specifies a history-dependent crossover in edit choice, loss of a two-step advantage after a partner switch, and explicit rejection conditions. These are measurable comparisons. No rival prediction is supplied, so separation cannot be assessed. Only a bench experiment would settle it.
What testing it would take
The engine's own read on whether this is testable with methods that already exist.
Affordable first implementation uses two frozen prompting configurations with independently validated, different restoration tendencies, brief neutral stories and a browser retelling task. Forecast elicitation should occur on separate calibration trials to avoid teaching evasion during the outcome trial. Analyze independent human chains as the replication unit, with participants and stories crossed. Power simulation requires pilot crossover variance, participant and story intraclass correlations, kernel uncertainty, coder confusion, attrition and delta. Stronger validation uses undisclosed natural partner variation, novel story families, languages and model families; it must show the effect outside an instruction to be original or defeat the model.
Other explanations
Every other hypothesis the engine wrote for the same gap, and the observation that would separate the two.
During history acquisition, randomize feedback showing that the partner anticipated a participant's previous semantic move versus equally accurate feedback about someone else's move; match words, accuracy, valence, exposure and immediate test framing using counterbalanced histories. Separately estimate ordinary algorithm reactance, general novelty-seeking and single-step source-conditioned rewriting under these histories. At the common-parent probe, use two frozen partners A and B preselected to undo different semantic edits, matched on accuracy and perceived reliability. The strong prediction is a crossover: humans with A-history choose edit y_A rather than y_B, despite equal immediate semantic distance and rated originality, and this ranking reverses with B-history; the choice specifically maximizes divergence after the NEXT model step. Switching the real next-step partner while keeping the historical partner cue fixed destroys the two-step advantage. Matched low-threat, non-evaluative conditions retain the crossover after measured reactance is calibrated. Merely writing more novel text or avoiding an AI label does not count. A null composed of individually measured reactance/novelty policies and partner-specific K_M that predicts the crossover within delta removes the extra mechanism; so does absence of the two-step crossover with a narrow interval. A strong history effect abolished only by factual choice receipts instead favors another hypothesis of the same gap.
- What would separate them
Mutual timing resets may steer meaning in human–model retelling chains predicts: First estimate individual boundary-response and temporal-memory effects with scripted, open-loop sequences, and model segmentation kernels with timestamp-blind prompts. Then form coupled alternating chains and deliver identical, semantically neutral boundary cues in regular versus phase-jittered schedules, matching the cue count, total time, distribution of intervals, reading dose and source content; randomize schedule order independently of text. Estimate phase from separate boundary reports or preregistered behavioral cycles, never from the semantic effect one intends to explain. In common-parent replay, the coupled model predicts a phase-response curve with reset-sensitive and insensitive windows and a selective loss of semantic-state locking when reciprocal cue contingency is broken. Timing shifts of the model boundary must shift the HUMAN phase and later semantic transition peaks, while shifts of human boundary timing must shift the model's subsequent segmentation; one-way timing sensitivity is insufficient. The crucial observable is held-out phase-specific SPV_4 transition probability beyond composition of independently measured event-boundary/spacing kernels. If such augmented component kernels account for the entire response, or no reproducible phase variable or reciprocal reset exists, discard the proposed family. A mere oscillation in average story sentiment is not evidence. If a control/data wrapper eliminates the effect while phase perturbation does not, another hypothesis of the same gap wins.
- Rival 02 of 03What would separate them
Lost quotation scope may turn story fragments into self-reinforcing model instructions predicts: Use harmless fictional quoted requests and editing-as-dialogue examples, never live tools or harmful instructions. At a common-parent probe, cross human continuity with retained model history. Compare the same historical words carried in explicit quoted-data records versus an ordinary conversational history wrapper; match wrapper length and position with neutral padding, and separately estimate wrapper effects on uncomplicated texts. another hypothesis of the same gap predicts that the residual history effect concentrates at the MODEL step, transfers with the historical scope-bearing text to a replacement human, and is sharply reduced by a verified instruction/data boundary without deleting the old semantic information. Human choice receipts alone have little effect after text exposure is matched. Reconstruct the scope-loss sequence from logs, then independently estimate H scope-conversion and M instruction-following kernels on the same input support. A closed-loop held-out excess in conversion probability must depend on both links: severing either historical quote-to-guidance conversion or model execution removes it. If these component kernels accurately compose, report ordinary prompt-injection susceptibility rather than a new recursion family. If no naturally arising scope conversion occurs, the endogenous hypothesis is falsified even if deliberately planted injections work. Phase jitter with intact scope should not selectively abolish this effect, unlike another hypothesis of the same gap.
- Rival 03 of 03What would separate them
Mistaken choice summaries may reinforce human preferences through repeated justification predicts: Randomize, during history acquisition, whether a model's summary accurately or incorrectly records which of two equally plausible neutral interpretations the human chose. Cross this with producing a reason for the recorded decision versus a matched factual-description task; match words, task time and number of choices, and include passive readers given the same account and rationale. At the identical-parent probe, randomize a verbatim receipt of the person's original click/choice versus an equally long non-diagnostic history receipt, then make a private, unrewarded interpretation choice and a subsequent retelling. The specific prediction is a substitution-by-self-justification effect on the private interpretation criterion and SPV_4 that is reduced by an accurate decision receipt; generic false information exposure without self-justification is weaker after calibration. Continue through a frozen model with factual narrative sources unchanged. A new-family claim additionally requires reciprocal adaptation of the model's inferred criterion to account for an effect beyond separately measured choice blindness, self-perception, source-monitoring and sycophancy kernels, including active versus yoked exposure controls. Accurate prospective composition eliminates the extra family even if ordinary choice blindness remains. If preserving quoted-data scope in model history alone removes the effect while authentic decision receipts do not, another hypothesis of the same gap wins. A receipt-sensitive effect without any own-choice/rationale interaction supports ordinary source monitoring and does not satisfy this candidate.
Why this is not the mainstream account
The engine is asked to say what its hypothesis would overturn and what would surprise a specialist. This is its answer.
Primary anchors are task-specific and indirect. Humans create more novelty than ChatGPT when asked to retell a story (2024), https://doi.org/10.1038/s41598-023-50229-7, reports human lexical/conceptual innovation alongside comparatively stable emotion ratings, showing that transformation can be active and dimension-specific; it does not test prediction evasion. Schlund and Zitek (2024), Algorithmic versus human surveillance leads to lower perceptions of autonomy and increased resistance, https://doi.org/10.1038/s44271-024-00102-8, experimentally documents increased resistance under algorithmic monitoring; that is a nearest rival and an anchor for the possibility of resistance, not evidence that cooperative retelling exhibits the proposed effect. The puzzle motivating the test is preservation of an evaluative core despite active reconstruction, not an asserted published paradox about hybrid chains.
The specific chapter-level target is the fixed-learner-bias treatment of iterated learning and cultural attraction in cultural-transmission theory, together with the cooperative audience-design chapter in psycholinguistics. The proposed revision is not that people remember past interactions, which is established; it is that successful prediction by a cooperative reconstruction partner reverses the sign of the objective governing source-semantic preservation and does so at the future partner-output level. A textbook would need an endogenous anti-prediction objective, rather than merely a different fixed mutation kernel. This is a named chapter topic, not an invented quotation or chapter number from a particular book.
With faithful retelling rewarded, no public credit or ownership dispute, matched immediate source and no difference in measured threat, making a model better at anticipating a person's meanings would reliably make later source meanings less faithful; humans would select different initially equivalent edits solely because each defeats a different future model reconstruction. Demonstrating that complete pattern beyond calibrated reactance and novelty-seeking would be surprising. A generic AI-aversion or creativity effect would not pass.
Bounded primary-source and keyword audit on 2026-10-03 covered human/LLM retelling, algorithmic surveillance/reactance, anti-predictive semantics, cultural attraction and mixed-chain order effects. No inspected source states this exact low-threat, two-step, partner-specific semantic-evasion claim. Reactance and creative divergence are already mainstream and are explicitly in the null. Absence from this bounded search is not proof that no review or perspective anywhere proposes it; HERETICAL is therefore a conditional research-role label, not a certified novelty finding. If an existing account already predicts the matched crossover, withdraw the heretical classification.
What stands behind it
Which of the figures above have a study behind them, which are the engine's own, and what it would take to refute the hypothesis. This audit never judges the idea.
This hypothesis states no figure and cites no study, so there is nothing here to trace.
What it would take to refute it. Nothing already retrieved carries the prediction’s terms and it names no measurement this layer can route to a public dataset, so the bench is the residual — not a finding against it.
0 citation handles extracted; 1 Europe PMC search run; 0 records examined; 0 sources stored for enrichment, 0 with full text. A citation that did not resolve is a bibliographic failure, not proof that no such paper exists, and no hypothesis is blocked by this audit.