Do human–machine retellings need a new explanation for meaning changes?

HERETICAL CANDIDATE: lineage-specific resistance to accurate repair. In a recurring human–AI collaboration, a human's earlier successful correction establishes an implicit right to determine that lineage's meaning. When a later model correction overrides that right, the human deliberately produces a competing interpretation, even when they can privately identify the source-correct interpretation. A history of being overruled, rather than the current story or an error-prone memory, changes the production objective. The substrate is an agent–lineage commitment, operationalized through consequential acceptance/rejection and subsequent edits, not inferred from introspective explanations alone. The extra dependency is R(person, lineage, history of exercised revision rights): P(next proposition | current artifact, full supplied context, R) differs when R is experimentally changed while artifact, exposure, model accuracy, motor work and available source evidence are matched. It predicts a paradoxical negative causal value of additional accurate corrections in the threatened-rights condition. This is a conjecture about deliberate semantic counter-production, not evidence that human–AI chains generally need a new state.

In development sessions, establish either participant final approval or neutral editorial approval using matched stories and identical accepted text. Subsequently provide identical verified source-correct model repairs while experimentally retaining or overriding the previously exercised approval right. Include a yoked observer with the same texts, actions, timing and accuracy evidence but no ownership of that lineage; model-versus-human source labels are independently counterbalanced. At an identical current-artifact checkpoint, the rights hypothesis predicts more intentional correct-to-incorrect or correct-to-incompatible transitions after accurate override than after accurate authorized repair, despite equivalent private source-question accuracy. This negative correction-dose slope should transfer with assignment of the lineage's revision right and disappear when the right is prospectively relinquished; an unrelated right on another lineage should not suffice. Binding fatigue predicts dependence on conflicting revision cycles, not legitimate versus illegitimate authority; embodied reinstatement predicts action matching; contract ambiguity predicts audit payoffs. Compare with independently calibrated reactance, algorithm-aversion, endowment, ordinary learning and belief-conditioned H kernels, not only an unconditioned H. If those established components predict the authority-by-history contrast within the meaningful margin on held-out lineages, retire the proposed distinct family. No residual interaction alone identifies a new norm mechanism.

Randomize whether an equal number of incompatible intermediate corrections repeatedly touches one connected neighborhood or dispersed unrelated links; restore the identical correct full text at a checkpoint and equate final exposure, total conflicting propositions, task time and output tokens. With the same human returning, the localized condition should show accelerating, spatially adjacent causal-role failures predicted by independently estimated a_t and revision-load amplitude, despite matched checkpoint text. Fresh humans should reset that excess; reinstating a gesture without repairing the affected bindings should not. Estimate load and binding accessibility in separate calibration participants to avoid the diagnostic test becoming retrieval practice. Compare the law against arbitrary flexible item-level learning/interference models, graph-conditioned one-step H and A, repeated-individual reconstruction, and exposure-position controls. A stable power-law relation fitted on one load/graph range must predict another without refitting its exponent. If endpoint text, ordinary interference and causal connectivity explain the trajectories; if growth is unrelated to connected damage; or if a_t merely redescribes the same scored errors, reject fatigue as a distinct mechanism. A good curve fit alone is insufficient.

Ordinary transformations may explain retelling chains without an extra recursive state predicts instead: Freeze nulls and their uncertainty before seeing evaluation chains. In human-only, model-only, HA and AH chains, held-out proposition transitions, causal-role reversals, exception loss and private-source retention fall inside the propagated predictive envelope, and any candidate extension improves proper predictive scores or prespecified discrepancies by less than the smallest meaningful margin. Order effects are allowed. At an identical-current-artifact checkpoint, matched context, resources and independently measured ordinary learning explain later differences; no additional lineage-right, local-fatigue, motor-history or joint-contract state earns predictive value. Intervene on ancestry access and source regeneration with explicit resource matching: context-conditioned kernels predict the changes without chain-specific refitting. Under predictive equivalence on independent seeds and genuinely held-out contexts, remove the distinct recursive family's novelty and priority, while retaining the observed attractor phenomenon. This IH loses if a reproducible intervention-specific discrepancy exceeds uncertainty and the meaningful margin after competent state enrichment and absolute model checks; that loss does not automatically identify which extension is right.

Reinstating learned gestures may preserve causal roles during human–model retelling predicts instead: At an identical-artifact checkpoint, cross semantically congruent role enactment at encoding with matched or swapped spatial enactment at later human production. Include no-enactment and equal-amplitude meaningless-movement controls, the same verbal generation and source-question practice, matched delays and workload, and fresh-human handoffs. Gesture instructions must not reveal any missing proposition; assign counterbalanced arbitrary locations to already supplied characters. This IH predicts an encoding-by-reinstatement interaction: congruent motor reinstatement selectively preserves the earlier causal roles, whereas swapping the learned locations increases role reversals even with the same current text. The interaction should persist after balancing ordinary verbal generation/retrieval practice and be absent for an unlearned movement mapping. Fatigue predicts localized cycle-dose deficits and fresh-person reset, but not this sign-changing mapping interaction; the rights and contract accounts predict their social manipulations instead. Calibrate an ordinary multimodal encoding-specificity model on nonrecursive tasks and replay controls. If it predicts the entire chain interaction, the motor explanation may be useful but the proposed new recursive family is eliminated. If matched motor perturbations have no meaningful role-specific effect despite a successful action-memory manipulation, reject this scout in favor of other models.

Continuing evaluation with coarse checks may reward strategic omissions in retelling predicts instead: Cross continuation of the same evaluation relationship versus a one-shot handoff with coarse whole-story evaluation versus prespecified proposition-specific audit. Hold expected reward, current source, evidence access, output length/time, task wording and candidate policy fixed as far as feasible; report residual incentive differences. Explicitly distinguish descriptive accuracy from public approval. The contract account predicts more omission of auditable exceptions/attributions under continuing relationships with coarse verification, and a selective reversal under item-level audit. The expected gain from possible later reinterpretation should predict WHICH details disappear, even when those details are causally central, easy to recall and accurately answered in private. Remove future evaluation or assign liability to an independent editor while preserving authorship rights: strategic omissions should shrink; merely transferring revision ownership should not suffice. Compare with independently calibrated one-shot incentive, audience-design, risk-aversion, self-presentation and accountability kernels composed across rounds. Only a held-out continuation-by-verifiability effect beyond those components supports the proposed extra state. If the effects are fully predicted by ordinary task-conditioned editing, remove the recursive/new-family claim. No contract-specific selectivity, despite verified incentive comprehension, falsifies this mechanism in the task.

CROSS-DOMAIN TRANSFER: localized fatigue of causal bindings during repeated correction. Repeatedly reconciling incompatible versions in the same human can damage a connected cluster of role/exception bindings even when the current written text has been restored exactly. Surface repair conceals a growing internal gap in the causal event representation; subsequent reconstruction loses adjacent relations disproportionately. The hypothesized extra state is a_t, the size of a connected low-accessibility region in a separately probed causal-binding graph, together with the history of revision load. This differs from an exhausted general resource: unrelated relations and overall response speed need not deteriorate. It also differs from simply dropping an exception in the artifact. A fresh reader of the identical current artifact should not inherit a hidden binding deficit unless some changed artifact or accessible history transmits it. A Paris-type growth law specifies the risky quantitative extension. No physical cracking of brain tissue is proposed.

Randomize whether an equal number of incompatible intermediate corrections repeatedly touches one connected neighborhood or dispersed unrelated links; restore the identical correct full text at a checkpoint and equate final exposure, total conflicting propositions, task time and output tokens. With the same human returning, the localized condition should show accelerating, spatially adjacent causal-role failures predicted by independently estimated a_t and revision-load amplitude, despite matched checkpoint text. Fresh humans should reset that excess; reinstating a gesture without repairing the affected bindings should not. Estimate load and binding accessibility in separate calibration participants to avoid the diagnostic test becoming retrieval practice. Compare the law against arbitrary flexible item-level learning/interference models, graph-conditioned one-step H and A, repeated-individual reconstruction, and exposure-position controls. A stable power-law relation fitted on one load/graph range must predict another without refitting its exponent. If endpoint text, ordinary interference and causal connectivity explain the trajectories; if growth is unrelated to connected damage; or if a_t merely redescribes the same scored errors, reject fatigue as a distinct mechanism. A good curve fit alone is insufficient.

In development sessions, establish either participant final approval or neutral editorial approval using matched stories and identical accepted text. Subsequently provide identical verified source-correct model repairs while experimentally retaining or overriding the previously exercised approval right. Include a yoked observer with the same texts, actions, timing and accuracy evidence but no ownership of that lineage; model-versus-human source labels are independently counterbalanced. At an identical current-artifact checkpoint, the rights hypothesis predicts more intentional correct-to-incorrect or correct-to-incompatible transitions after accurate override than after accurate authorized repair, despite equivalent private source-question accuracy. This negative correction-dose slope should transfer with assignment of the lineage's revision right and disappear when the right is prospectively relinquished; an unrelated right on another lineage should not suffice. Binding fatigue predicts dependence on conflicting revision cycles, not legitimate versus illegitimate authority; embodied reinstatement predicts action matching; contract ambiguity predicts audit payoffs. Compare with independently calibrated reactance, algorithm-aversion, endowment, ordinary learning and belief-conditioned H kernels, not only an unconditioned H. If those established components predict the authority-by-history contrast within the meaningful margin on held-out lineages, retire the proposed distinct family. No residual interaction alone identifies a new norm mechanism.

Ordinary transformations may explain retelling chains without an extra recursive state predicts instead: Freeze nulls and their uncertainty before seeing evaluation chains. In human-only, model-only, HA and AH chains, held-out proposition transitions, causal-role reversals, exception loss and private-source retention fall inside the propagated predictive envelope, and any candidate extension improves proper predictive scores or prespecified discrepancies by less than the smallest meaningful margin. Order effects are allowed. At an identical-current-artifact checkpoint, matched context, resources and independently measured ordinary learning explain later differences; no additional lineage-right, local-fatigue, motor-history or joint-contract state earns predictive value. Intervene on ancestry access and source regeneration with explicit resource matching: context-conditioned kernels predict the changes without chain-specific refitting. Under predictive equivalence on independent seeds and genuinely held-out contexts, remove the distinct recursive family's novelty and priority, while retaining the observed attractor phenomenon. This IH loses if a reproducible intervention-specific discrepancy exceeds uncertainty and the meaningful margin after competent state enrichment and absolute model checks; that loss does not automatically identify which extension is right.

Reinstating learned gestures may preserve causal roles during human–model retelling predicts instead: At an identical-artifact checkpoint, cross semantically congruent role enactment at encoding with matched or swapped spatial enactment at later human production. Include no-enactment and equal-amplitude meaningless-movement controls, the same verbal generation and source-question practice, matched delays and workload, and fresh-human handoffs. Gesture instructions must not reveal any missing proposition; assign counterbalanced arbitrary locations to already supplied characters. This IH predicts an encoding-by-reinstatement interaction: congruent motor reinstatement selectively preserves the earlier causal roles, whereas swapping the learned locations increases role reversals even with the same current text. The interaction should persist after balancing ordinary verbal generation/retrieval practice and be absent for an unlearned movement mapping. Fatigue predicts localized cycle-dose deficits and fresh-person reset, but not this sign-changing mapping interaction; the rights and contract accounts predict their social manipulations instead. Calibrate an ordinary multimodal encoding-specificity model on nonrecursive tasks and replay controls. If it predicts the entire chain interaction, the motor explanation may be useful but the proposed new recursive family is eliminated. If matched motor perturbations have no meaningful role-specific effect despite a successful action-memory manipulation, reject this scout in favor of other models.

Continuing evaluation with coarse checks may reward strategic omissions in retelling predicts instead: Cross continuation of the same evaluation relationship versus a one-shot handoff with coarse whole-story evaluation versus prespecified proposition-specific audit. Hold expected reward, current source, evidence access, output length/time, task wording and candidate policy fixed as far as feasible; report residual incentive differences. Explicitly distinguish descriptive accuracy from public approval. The contract account predicts more omission of auditable exceptions/attributions under continuing relationships with coarse verification, and a selective reversal under item-level audit. The expected gain from possible later reinterpretation should predict WHICH details disappear, even when those details are causally central, easy to recall and accurately answered in private. Remove future evaluation or assign liability to an independent editor while preserving authorship rights: strategic omissions should shrink; merely transferring revision ownership should not suffice. Compare with independently calibrated one-shot incentive, audience-design, risk-aversion, self-presentation and accountability kernels composed across rounds. Only a held-out continuation-by-verifiability effect beyond those components supports the proposed extra state. If the effects are fully predicted by ordinary task-conditioned editing, remove the recursive/new-family claim. No contract-specific selectivity, despite verified incentive comprehension, falsifies this mechanism in the task.

PHENOMENON-DOESN'T-EXIST: the apparently distinct recursive human–AI mechanism is an epiphenomenon of composing ordinary content-biased transformations, resource-matched reconstruction and the observation process. The empirical attractors and semantic losses may be completely real. What does not exist in the tested scope is a causally necessary extra chain-level state. Calibrate H(y|x,c) and A(y|x,c) on randomized one-step transformations, allowing sufficiently rich current-artifact features x and declared context c. For row-vector distributions under a stationary discretization, alternating chains predict p0(HA)^k or p0(AH)^k; H and A need not commute. Independently calibrate q(y|theme, task, supplied information, budget), source regeneration, and ordinary repeated-individual learning/reconstruction where a human returns. Coarse coding, unequal budgets and unmeasured lexical cues can otherwise look like recursion. These are a prespecified family of compositional nulls, not a post hoc universal model allowed to absorb every discrepancy.

In human-only, model-only, HA and AH chains, held-out proposition transitions, causal-role reversals, exception loss and private-source retention fall inside the propagated predictive envelope, and any candidate extension improves proper predictive scores or prespecified discrepancies by less than the smallest meaningful margin. Order effects are allowed. At an identical-current-artifact checkpoint, matched context, resources and independently measured ordinary learning explain later differences; no additional lineage-right, local-fatigue, motor-history or joint-contract state earns predictive value. Intervene on ancestry access and source regeneration with explicit resource matching: context-conditioned kernels predict the changes without chain-specific refitting. Under predictive equivalence on independent seeds and genuinely held-out contexts, remove the distinct recursive family's novelty and priority, while retaining the observed attractor phenomenon. This IH loses if a reproducible intervention-specific discrepancy exceeds uncertainty and the meaningful margin after competent state enrichment and absolute model checks; that loss does not automatically identify which extension is right.

In development sessions, establish either participant final approval or neutral editorial approval using matched stories and identical accepted text. Subsequently provide identical verified source-correct model repairs while experimentally retaining or overriding the previously exercised approval right. Include a yoked observer with the same texts, actions, timing and accuracy evidence but no ownership of that lineage; model-versus-human source labels are independently counterbalanced. At an identical current-artifact checkpoint, the rights hypothesis predicts more intentional correct-to-incorrect or correct-to-incompatible transitions after accurate override than after accurate authorized repair, despite equivalent private source-question accuracy. This negative correction-dose slope should transfer with assignment of the lineage's revision right and disappear when the right is prospectively relinquished; an unrelated right on another lineage should not suffice. Binding fatigue predicts dependence on conflicting revision cycles, not legitimate versus illegitimate authority; embodied reinstatement predicts action matching; contract ambiguity predicts audit payoffs. Compare with independently calibrated reactance, algorithm-aversion, endowment, ordinary learning and belief-conditioned H kernels, not only an unconditioned H. If those established components predict the authority-by-history contrast within the meaningful margin on held-out lineages, retire the proposed distinct family. No residual interaction alone identifies a new norm mechanism.

Conflicting revisions may erode connected story memories even after the text is repaired predicts instead: Use graph-matched fictional narratives with experimentally known causal links. Randomize whether an equal number of incompatible intermediate corrections repeatedly touches one connected neighborhood or dispersed unrelated links; restore the identical correct full text at a checkpoint and equate final exposure, total conflicting propositions, task time and output tokens. With the same human returning, the localized condition should show accelerating, spatially adjacent causal-role failures predicted by independently estimated a_t and revision-load amplitude, despite matched checkpoint text. Fresh humans should reset that excess; reinstating a gesture without repairing the affected bindings should not. Estimate load and binding accessibility in separate calibration participants to avoid the diagnostic test becoming retrieval practice. Compare the law against arbitrary flexible item-level learning/interference models, graph-conditioned one-step H and A, repeated-individual reconstruction, and exposure-position controls. A stable power-law relation fitted on one load/graph range must predict another without refitting its exponent. If endpoint text, ordinary interference and causal connectivity explain the trajectories; if growth is unrelated to connected damage; or if a_t merely redescribes the same scored errors, reject fatigue as a distinct mechanism. A good curve fit alone is insufficient.

Reinstating learned gestures may preserve causal roles during human–model retelling predicts instead: At an identical-artifact checkpoint, cross semantically congruent role enactment at encoding with matched or swapped spatial enactment at later human production. Include no-enactment and equal-amplitude meaningless-movement controls, the same verbal generation and source-question practice, matched delays and workload, and fresh-human handoffs. Gesture instructions must not reveal any missing proposition; assign counterbalanced arbitrary locations to already supplied characters. This IH predicts an encoding-by-reinstatement interaction: congruent motor reinstatement selectively preserves the earlier causal roles, whereas swapping the learned locations increases role reversals even with the same current text. The interaction should persist after balancing ordinary verbal generation/retrieval practice and be absent for an unlearned movement mapping. Fatigue predicts localized cycle-dose deficits and fresh-person reset, but not this sign-changing mapping interaction; the rights and contract accounts predict their social manipulations instead. Calibrate an ordinary multimodal encoding-specificity model on nonrecursive tasks and replay controls. If it predicts the entire chain interaction, the motor explanation may be useful but the proposed new recursive family is eliminated. If matched motor perturbations have no meaningful role-specific effect despite a successful action-memory manipulation, reject this scout in favor of other models.

Continuing evaluation with coarse checks may reward strategic omissions in retelling predicts instead: Cross continuation of the same evaluation relationship versus a one-shot handoff with coarse whole-story evaluation versus prespecified proposition-specific audit. Hold expected reward, current source, evidence access, output length/time, task wording and candidate policy fixed as far as feasible; report residual incentive differences. Explicitly distinguish descriptive accuracy from public approval. The contract account predicts more omission of auditable exceptions/attributions under continuing relationships with coarse verification, and a selective reversal under item-level audit. The expected gain from possible later reinterpretation should predict WHICH details disappear, even when those details are causally central, easy to recall and accurately answered in private. Remove future evaluation or assign liability to an independent editor while preserving authorship rights: strategic omissions should shrink; merely transferring revision ownership should not suffice. Compare with independently calibrated one-shot incentive, audience-design, risk-aversion, self-presentation and accountability kernels composed across rounds. Only a held-out continuation-by-verifiability effect beyond those components supports the proposed extra state. If the effects are fully predicted by ordinary task-conditioned editing, remove the recursive/new-family claim. No contract-specific selectivity, despite verified incentive comprehension, falsifies this mechanism in the task.

SCOUT 1 — motor learning and embodied action: semantic continuity is partly carried by a human's event-enactment plan. During retelling, directional gestures or enacted role assignments bind who-did-what-to-whom. Model rewriting can remove the verbal cues that normally reinstate that plan, while the source propositions remain literally available. A returning human then reconstructs the wrong causal roles when the earlier action organization cannot be reinstated. The extra dependency is the compatibility between the person's earlier role-specific action pattern and the current production action, not total rehearsal or text similarity. This is a task-specific extension of established gesture/enactment effects, not a claim that AI has bodily memory. It predicts a cross-modal rescue of source-role retention at a matched current text and verbal exposure, without requiring extra ancestry text.

Include no-enactment and equal-amplitude meaningless-movement controls, the same verbal generation and source-question practice, matched delays and workload, and fresh-human handoffs. Gesture instructions must not reveal any missing proposition; assign counterbalanced arbitrary locations to already supplied characters. This IH predicts an encoding-by-reinstatement interaction: congruent motor reinstatement selectively preserves the earlier causal roles, whereas swapping the learned locations increases role reversals even with the same current text. The interaction should persist after balancing ordinary verbal generation/retrieval practice and be absent for an unlearned movement mapping. Fatigue predicts localized cycle-dose deficits and fresh-person reset, but not this sign-changing mapping interaction; the rights and contract accounts predict their social manipulations instead. Calibrate an ordinary multimodal encoding-specificity model on nonrecursive tasks and replay controls. If it predicts the entire chain interaction, the motor explanation may be useful but the proposed new recursive family is eliminated. If matched motor perturbations have no meaningful role-specific effect despite a successful action-memory manipulation, reject this scout in favor of other models.

In development sessions, establish either participant final approval or neutral editorial approval using matched stories and identical accepted text. Subsequently provide identical verified source-correct model repairs while experimentally retaining or overriding the previously exercised approval right. Include a yoked observer with the same texts, actions, timing and accuracy evidence but no ownership of that lineage; model-versus-human source labels are independently counterbalanced. At an identical current-artifact checkpoint, the rights hypothesis predicts more intentional correct-to-incorrect or correct-to-incompatible transitions after accurate override than after accurate authorized repair, despite equivalent private source-question accuracy. This negative correction-dose slope should transfer with assignment of the lineage's revision right and disappear when the right is prospectively relinquished; an unrelated right on another lineage should not suffice. Binding fatigue predicts dependence on conflicting revision cycles, not legitimate versus illegitimate authority; embodied reinstatement predicts action matching; contract ambiguity predicts audit payoffs. Compare with independently calibrated reactance, algorithm-aversion, endowment, ordinary learning and belief-conditioned H kernels, not only an unconditioned H. If those established components predict the authority-by-history contrast within the meaningful margin on held-out lineages, retire the proposed distinct family. No residual interaction alone identifies a new norm mechanism.

Conflicting revisions may erode connected story memories even after the text is repaired predicts instead: Use graph-matched fictional narratives with experimentally known causal links. Randomize whether an equal number of incompatible intermediate corrections repeatedly touches one connected neighborhood or dispersed unrelated links; restore the identical correct full text at a checkpoint and equate final exposure, total conflicting propositions, task time and output tokens. With the same human returning, the localized condition should show accelerating, spatially adjacent causal-role failures predicted by independently estimated a_t and revision-load amplitude, despite matched checkpoint text. Fresh humans should reset that excess; reinstating a gesture without repairing the affected bindings should not. Estimate load and binding accessibility in separate calibration participants to avoid the diagnostic test becoming retrieval practice. Compare the law against arbitrary flexible item-level learning/interference models, graph-conditioned one-step H and A, repeated-individual reconstruction, and exposure-position controls. A stable power-law relation fitted on one load/graph range must predict another without refitting its exponent. If endpoint text, ordinary interference and causal connectivity explain the trajectories; if growth is unrelated to connected damage; or if a_t merely redescribes the same scored errors, reject fatigue as a distinct mechanism. A good curve fit alone is insufficient.

Ordinary transformations may explain retelling chains without an extra recursive state predicts instead: Freeze nulls and their uncertainty before seeing evaluation chains. In human-only, model-only, HA and AH chains, held-out proposition transitions, causal-role reversals, exception loss and private-source retention fall inside the propagated predictive envelope, and any candidate extension improves proper predictive scores or prespecified discrepancies by less than the smallest meaningful margin. Order effects are allowed. At an identical-current-artifact checkpoint, matched context, resources and independently measured ordinary learning explain later differences; no additional lineage-right, local-fatigue, motor-history or joint-contract state earns predictive value. Intervene on ancestry access and source regeneration with explicit resource matching: context-conditioned kernels predict the changes without chain-specific refitting. Under predictive equivalence on independent seeds and genuinely held-out contexts, remove the distinct recursive family's novelty and priority, while retaining the observed attractor phenomenon. This IH loses if a reproducible intervention-specific discrepancy exceeds uncertainty and the meaningful margin after competent state enrichment and absolute model checks; that loss does not automatically identify which extension is right.

Continuing evaluation with coarse checks may reward strategic omissions in retelling predicts instead: Cross continuation of the same evaluation relationship versus a one-shot handoff with coarse whole-story evaluation versus prespecified proposition-specific audit. Hold expected reward, current source, evidence access, output length/time, task wording and candidate policy fixed as far as feasible; report residual incentive differences. Explicitly distinguish descriptive accuracy from public approval. The contract account predicts more omission of auditable exceptions/attributions under continuing relationships with coarse verification, and a selective reversal under item-level audit. The expected gain from possible later reinterpretation should predict WHICH details disappear, even when those details are causally central, easy to recall and accurately answered in private. Remove future evaluation or assign liability to an independent editor while preserving authorship rights: strategic omissions should shrink; merely transferring revision ownership should not suffice. Compare with independently calibrated one-shot incentive, audience-design, risk-aversion, self-presentation and accountability kernels composed across rounds. Only a held-out continuation-by-verifiability effect beyond those components supports the proposed extra state. If the effects are fully predicted by ordinary task-conditioned editing, remove the recursive/new-family claim. No contract-specific selectivity, despite verified incentive comprehension, falsifies this mechanism in the task.

SCOUT 2 — relational contract economics: some semantic erosion is an intentional investment in interpretive discretion. A human who expects to be evaluated later for a collaboratively produced narrative can benefit from keeping evidence, exception clauses and attribution underspecified, because the eventual reader or successor may apply a different criterion. Successive human–AI edits provide chances to omit verifiable commitments while maintaining a fluent central claim. The causal state is the allocation of downstream accountability and the expected continuation of the same evaluative relationship, not an internal semantic attractor. The extra dependency is the complementarity between future revision opportunities and coarse versus proposition-specific verification. It predicts preservation of broad meaning alongside selective deletion of details that would constrain later reinterpretation. The model is a strategic production account; there is no presumption that participants are consciously deceptive or that ordinary real-world writers have the experimental incentives.

Hold expected reward, current source, evidence access, output length/time, task wording and candidate policy fixed as far as feasible; report residual incentive differences. Explicitly distinguish descriptive accuracy from public approval. The contract account predicts more omission of auditable exceptions/attributions under continuing relationships with coarse verification, and a selective reversal under item-level audit. The expected gain from possible later reinterpretation should predict WHICH details disappear, even when those details are causally central, easy to recall and accurately answered in private. Remove future evaluation or assign liability to an independent editor while preserving authorship rights: strategic omissions should shrink; merely transferring revision ownership should not suffice. Compare with independently calibrated one-shot incentive, audience-design, risk-aversion, self-presentation and accountability kernels composed across rounds. Only a held-out continuation-by-verifiability effect beyond those components supports the proposed extra state. If the effects are fully predicted by ordinary task-conditioned editing, remove the recursive/new-family claim. No contract-specific selectivity, despite verified incentive comprehension, falsifies this mechanism in the task.

In development sessions, establish either participant final approval or neutral editorial approval using matched stories and identical accepted text. Subsequently provide identical verified source-correct model repairs while experimentally retaining or overriding the previously exercised approval right. Include a yoked observer with the same texts, actions, timing and accuracy evidence but no ownership of that lineage; model-versus-human source labels are independently counterbalanced. At an identical current-artifact checkpoint, the rights hypothesis predicts more intentional correct-to-incorrect or correct-to-incompatible transitions after accurate override than after accurate authorized repair, despite equivalent private source-question accuracy. This negative correction-dose slope should transfer with assignment of the lineage's revision right and disappear when the right is prospectively relinquished; an unrelated right on another lineage should not suffice. Binding fatigue predicts dependence on conflicting revision cycles, not legitimate versus illegitimate authority; embodied reinstatement predicts action matching; contract ambiguity predicts audit payoffs. Compare with independently calibrated reactance, algorithm-aversion, endowment, ordinary learning and belief-conditioned H kernels, not only an unconditioned H. If those established components predict the authority-by-history contrast within the meaningful margin on held-out lineages, retire the proposed distinct family. No residual interaction alone identifies a new norm mechanism.

Conflicting revisions may erode connected story memories even after the text is repaired predicts instead: Use graph-matched fictional narratives with experimentally known causal links. Randomize whether an equal number of incompatible intermediate corrections repeatedly touches one connected neighborhood or dispersed unrelated links; restore the identical correct full text at a checkpoint and equate final exposure, total conflicting propositions, task time and output tokens. With the same human returning, the localized condition should show accelerating, spatially adjacent causal-role failures predicted by independently estimated a_t and revision-load amplitude, despite matched checkpoint text. Fresh humans should reset that excess; reinstating a gesture without repairing the affected bindings should not. Estimate load and binding accessibility in separate calibration participants to avoid the diagnostic test becoming retrieval practice. Compare the law against arbitrary flexible item-level learning/interference models, graph-conditioned one-step H and A, repeated-individual reconstruction, and exposure-position controls. A stable power-law relation fitted on one load/graph range must predict another without refitting its exponent. If endpoint text, ordinary interference and causal connectivity explain the trajectories; if growth is unrelated to connected damage; or if a_t merely redescribes the same scored errors, reject fatigue as a distinct mechanism. A good curve fit alone is insufficient.

Ordinary transformations may explain retelling chains without an extra recursive state predicts instead: Freeze nulls and their uncertainty before seeing evaluation chains. In human-only, model-only, HA and AH chains, held-out proposition transitions, causal-role reversals, exception loss and private-source retention fall inside the propagated predictive envelope, and any candidate extension improves proper predictive scores or prespecified discrepancies by less than the smallest meaningful margin. Order effects are allowed. At an identical-current-artifact checkpoint, matched context, resources and independently measured ordinary learning explain later differences; no additional lineage-right, local-fatigue, motor-history or joint-contract state earns predictive value. Intervene on ancestry access and source regeneration with explicit resource matching: context-conditioned kernels predict the changes without chain-specific refitting. Under predictive equivalence on independent seeds and genuinely held-out contexts, remove the distinct recursive family's novelty and priority, while retaining the observed attractor phenomenon. This IH loses if a reproducible intervention-specific discrepancy exceeds uncertainty and the meaningful margin after competent state enrichment and absolute model checks; that loss does not automatically identify which extension is right.

Reinstating learned gestures may preserve causal roles during human–model retelling predicts instead: At an identical-artifact checkpoint, cross semantically congruent role enactment at encoding with matched or swapped spatial enactment at later human production. Include no-enactment and equal-amplitude meaningless-movement controls, the same verbal generation and source-question practice, matched delays and workload, and fresh-human handoffs. Gesture instructions must not reveal any missing proposition; assign counterbalanced arbitrary locations to already supplied characters. This IH predicts an encoding-by-reinstatement interaction: congruent motor reinstatement selectively preserves the earlier causal roles, whereas swapping the learned locations increases role reversals even with the same current text. The interaction should persist after balancing ordinary verbal generation/retrieval practice and be absent for an unlearned movement mapping. Fatigue predicts localized cycle-dose deficits and fresh-person reset, but not this sign-changing mapping interaction; the rights and contract accounts predict their social manipulations instead. Calibrate an ordinary multimodal encoding-specificity model on nonrecursive tasks and replay controls. If it predicts the entire chain interaction, the motor explanation may be useful but the proposed new recursive family is eliminated. If matched motor perturbations have no meaningful role-specific effect despite a successful action-memory manipulation, reject this scout in favor of other models.