Can meaning changes in human–machine retellings be predicted?

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.

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.

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, IH_03 wins.

Lost quotation scope may turn story fragments into self-reinforcing model instructions predicts instead: 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. IH_03 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 IH_02.

Mistaken choice summaries may reinforce human preferences through repeated justification predicts instead: 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, IH_03 wins. A receipt-sensitive effect without any own-choice/rationale interaction supports ordinary source monitoring and does not satisfy this candidate.

CROSS-DOMAIN TRANSFER: retained interaction history sets the phase at which an interpretation can be reassigned. A human alternates between an interpretation-updating phase and a committed retelling phase; model-generated event boundaries can reset that cycle, while human boundary placement changes the model's next segmentation. An identical parent can therefore enter a different semantic basin depending on the relative phase established by the preceding lineage, even with identical immediate instructions. The proposed extra dependency is reciprocal phase resetting of semantic competence, not merely ordinary recency, the number of rewrites, or a longer reading interval. A minimal clock-and-wavefront-inspired model has dphi_H/dt=omega_H+k_H sin(phi_M(t-tau_MH)-phi_H), dphi_M/dt=omega_M+k_M sin(phi_H(t-tau_HM)-phi_M); commitment of proposition j occurs when c_j(t) crosses threshold c_star AND phi_H lies in gate W. The model phase is an observable task/context phase, not a biological or spontaneous neural oscillator. c_j is measured interpretation uncertainty/competence for reassignment. A subsequent boundary reopens c_j only if the counterpart's boundary arrives in a reset-sensitive phase. The mechanism stabilizes a lineage-specific pattern of SPV_4 transitions by phase-dependent commitment, without requiring a different steady cognitive prior.

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, IH_03 wins.

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.

Lost quotation scope may turn story fragments into self-reinforcing model instructions predicts instead: 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. IH_03 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 IH_02.

Mistaken choice summaries may reinforce human preferences through repeated justification predicts instead: 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, IH_03 wins. A receipt-sensitive effect without any own-choice/rationale interaction supports ordinary source monitoring and does not satisfy this candidate.

SCOUT 1 — computer security and language-based execution: neutral narrative quotations or editing comments become operative instructions after repeated human-model transformations strip their quotation scope. Retained model context then executes an ancestor's instruction-like fragment when transforming an identical current source. The human's subsequent fluent repair preferentially preserves the now task-directing paraphrase, making the next model context still more susceptible. The state is concrete retained text plus its role/scope interpretation, not unexplained model memory or updated weights. The extra causal loop proposed for testing is descendant-induced change in instruction status: H converts quoted/story-level material into apparently general editing guidance; M follows that guidance and emits a descendant that increases the next H conversion probability. Let a_g be the independently annotated instruction-status indicator of an ancestor fragment and e_g the probability the model treats it as operative. The proposed loop is a_g -> e_g -> scope-loss in descendant g+1 -> a_(g+1); a current-source-only K_M omits a_g. A special hybrid family is retained only if closed-loop scope conversion is not predicted by independently measured scope-loss and instruction-following channels. It changes SPV_4 through task redirection rather than through altered human belief or semantic attraction.

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. IH_03 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 IH_02.

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.

Mutual timing resets may steer meaning in human–model retelling chains predicts instead: 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, IH_03 wins.

Mistaken choice summaries may reinforce human preferences through repeated justification predicts instead: 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, IH_03 wins. A receipt-sensitive effect without any own-choice/rationale interaction supports ordinary source monitoring and does not satisfy this candidate.

SCOUT 2 — experimental decision science and choice-blindness paradigms: a human reconstructs a semantic preference from a model-returned account of their previous editing decision, rather than retrieving a stable preference or merely believing the model's facts. In a retained interaction, an unnoticed model substitution of a previously selected narrative interpretation elicits a human justification; the model then treats that justification as a stable preference and uses it to resolve future ambiguities, causing the human to reconstruct still stronger preference on the next encounter. The proposed additional dependency is repeated preference construction from the partner's representation of one's own decision, followed by partner conditioning on the newly constructed criterion. Let c_g be the actual logged choice, c_tilde_g the choice returned in the model summary, r_g the human's rationale, and theta_g the human's measured interpretation criterion. The candidate loop is c_tilde_g -> r_g -> theta_(g+1) -> model choice-summary c_tilde_(g+1), conditional on known discrepancies c_g != c_tilde_g. It is not normative ownership, a contractual right to resist correction, or simple copying of a false fact. The relevant state is a newly constructed evaluation rule expressed in justifications and revealed in private interpretation choices. It stabilizes a changed SPV_4 transition distribution even when the current narrative is unchanged.

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, IH_03 wins. A receipt-sensitive effect without any own-choice/rationale interaction supports ordinary source monitoring and does not satisfy this candidate.

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.

Mutual timing resets may steer meaning in human–model retelling chains predicts instead: 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, IH_03 wins.

Lost quotation scope may turn story fragments into self-reinforcing model instructions predicts instead: 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. IH_03 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 IH_02.