Which order of correction and retelling best preserves meaning?

HERETICAL CANDIDATE — Accurate correction can make an accurately informed person a systematically worse cultural transmitter by prematurely closing the agenda of what still needs to be said. The stored state is an unexpressed communicative plan, not a false factual belief, a right of authorship, or a demand to preserve one's own version. Human reconstruction first opens proposition-specific production slots; a later source correction can repair their contents while leaving them scheduled for expression. Correction first instead marks those relations as already settled before an outgoing message has been planned, so they are omitted even when the speaker privately knows them, knows that the new recipient does not know them, has enough time, and is paid for receiver comprehension. The strong, unconventional claim is correction-specific failure to reopen these slots after an explicit audience reset; ordinary audience design alone is the nearest rival. Let a_j be the independently manipulated presence of an outgoing slot for relation j before C, z_j indicate that j was corrected, and o_j indicate omission. A minimal extension is logit P(o_j=1)=alpha_j+beta z_j(1-a_j)+ordinary measured predictors, beta>0, while private source accuracy does not decrease. R-before-C should preserve meaning best when enough relational identity survives to attach correction to an existing slot; recognition of the entities alone is insufficient. The resulting more complete messages can preserve their advantage after withdrawal in new recipients, but no immaterial agenda is assumed to cross between minds. This mechanism stabilizes SPV_4 through production completeness and can improve semantic SPV_7 without improving verbatim copying.

The agenda lists only which already-present relation must be conveyed; it does not supply its answer. Give an explicit new-recipient ignorance cue in every cell, match overt retrieval and elapsed time, and assess source knowledge on an independent unit set to avoid rehearsal contamination. Define D_A as the C-R minus R-C omission difference. The candidate predicts D_A>delta with no pre-opened agenda, but |D_A|<epsilon when the same agenda is opened before C; opening it only after C fails to rescue within the prespecified interval. Crucially, the correction arm has equal or higher private source accuracy but lower source-specific receiver accuracy. The correction-by-agenda interaction remains after matching trial-level attention, source-memory discrimination, total message length and ordinary given/new status. Uniform correction-time allocation, unchanged source-selection rules and a no-sleep immediate session do not remove it. Subsequent human-only descendants retain the between-lineage semantic advantage over a pilot-defined horizon; exact-artifact yoking abolishes predecessor-history differences. An absence of agenda timing specificity with a selective bundle-allocation effect favors IH_02; a sleep-stage/cue reversal favors IH_03; an explicit sampling-rule reversal favors IH_04. If ordinary audience-design/retrieval models already predict the whole interaction, remove the new-family claim even if the best schedule remains R-C.

Include targets whose experimentally defined functional value is additive and targets whose joint correction is complementary, while keeping the proposition rubric fixed. Estimate each participant's complementarity from incentive-compatible choices on independent calibration units. Let D_B be the R-C minus C-R source-specific functional accuracy gap. Predict D_B>delta under separate-target allocation for complementary relations, with a smaller |D_B| under package allocation; the package interaction is near zero for additive units. Manipulating retained relational identity changes measured bundle complementarity and predicts this interaction before seeing the final outcomes. The decisive replay control randomly yokes the exact per-target attention schedule and correction content from package winners to recipients who never bid: the auction-label effect and any unexplained order advantage should fall within epsilon once this allocation path is fixed. Agenda timing, sleep stage and evidence-sampling labels do not produce independent effects in the allocation-only model. Continued advantage after withdrawal must be predicted from what was actually corrected, not from the presence of old prices. If bids fail to forecast repair allocation, or fixed allocation leaves a reproducible agenda-, sleep-, or selection-rule-specific residual, reject this mechanism as the sufficient explanation in favor of the appropriate rival.

Sleep may favor correction after reconstruction by stabilizing source-role bindings predicts instead: Independently tag source-relation encoding and the human reconstruction episode with counterbalanced nonsemantic cues. Cross C-R versus R-C with a matched nap/quiet-rest interval, then, in the physiological validation, deliver source-episode versus reconstruction-episode cues during verified slow-wave sleep, with sham and waking-cue controls. Hold clock time, interference, cue count, arousal and source information constant or explicitly counterbalance them. Define D_S as the post-interval R-C minus C-R probability of preserving the arbitrary source relation, subtracting each schedule's immediate baseline on independent items. Predict a stage-by-cue-by-order interaction exceeding delta: stabilizing a correctable source-bound episode benefits R-C, whereas selectively stabilizing a conflicting reconstruction reduces or reverses that advantage. The corresponding effect is weaker than delta in waking/sham controls under this candidate's stated scope. Entity-only retention weakens the interaction despite matched recognition. One-step kernels estimated separately for sleep and wake plus standard retrieval, spacing and targeted-memory-reactivation effects are the nearest strong rival; an extra order-by-binding state must improve held-out prediction beyond them. If only a nonspecific sleep main effect occurs, or a standard replay model already predicts the full contrast, remove the proposed distinct family. If awake agenda timing or package allocation predicts the advantage and verified stage-specific cueing adds no informative effect, prefer IH_01 or IH_02. After withdrawal, test both original individuals and naive human chains; exact-message replay must eliminate inaccessible predecessor-sleep effects.

Correction order may bias later rule meanings through beliefs about why examples were chosen predicts instead: Use invented narrative worlds with a general rule and explicitly indexed exceptions; correction supplies a true, partially diagnostic source example rather than simply spelling out the entire answer. Pilot that the example is understood and supports the prespecified source interpretation. Cross C-R/R-C with a transparent random-example selection rule versus error-triggered selection. Construct worlds where the exact observed corrective example has the same surface form and source truth under both sampling rules but different evidential relevance to the unsupplied generalization. Everyone receives the same number of source examples and equal time. Establish the ordinary sampling-conditioned response on independent one-step items. Then use new, exchangeable source units to test whether previously experienced correction order changes the inferred selection policy despite an explicitly reset current policy. Define D_Q as the order difference in a held-out scope-transfer error, with exact fact recall held equivalent. The candidate predicts D_Q changes sign between general-rule and genuine-exception worlds, and an experimentally verified policy reset reduces the excess order effect below epsilon. The extra-history term is required only if order-linked policy carryover predicts new-source errors beyond the known sampling-conditioned Bayesian baseline; a mere random-versus-selected example effect is established inference, not a new memetic result. Opening a communicative agenda cannot rescue the private scope error while leaving the selection interpretation unchanged; neither a bundle-allocation change nor a sleep cue has a necessary effect when its proposed state is clamped. Falsify in favor of IH_01 if only output omission changes with intact private scope, or in favor of IH_03 if only the physiological cue interaction remains after sampling comprehension is verified.

CROSS-DOMAIN TRANSFER — Reconstruction changes the complementarity of attention invested in correcting different source relations. All source-grounded correction information is already available in both orders, but a human who has attempted reconstruction discovers which two repairs are jointly useful. When attention time is allocated by separate-target bids, the risk of paying for one unusable half of a repair pair discourages investment; package allocation can remove that exposure problem. The relevant history state is a revealed valuation over correction-time bundles, not depletion of a finite memory resource. R-before-C ordinarily helps because it reveals those complements before integration time is assigned. A source-grounded C-before-R schedule can match it if recipients already retain which roles must be repaired together and are allowed package allocation. The prerequisite is retained identification of complementary targets; knowing their names without the relation is insufficient. Correction withdrawal leaves a benefit only insofar as the previously allocated integration produces more usable transmitted content; a continued market or reminders must not be silently retained. This stabilizes SPV_4 by endogenous selection of integration effort and may preserve SPV_7. It is an established economic mechanism proposed as a rival explanation of a cultural interaction, not a discovered cultural law.

Include targets whose experimentally defined functional value is additive and targets whose joint correction is complementary, while keeping the proposition rubric fixed. Estimate each participant's complementarity from incentive-compatible choices on independent calibration units. Let D_B be the R-C minus C-R source-specific functional accuracy gap. Predict D_B>delta under separate-target allocation for complementary relations, with a smaller |D_B| under package allocation; the package interaction is near zero for additive units. Manipulating retained relational identity changes measured bundle complementarity and predicts this interaction before seeing the final outcomes. The decisive replay control randomly yokes the exact per-target attention schedule and correction content from package winners to recipients who never bid: the auction-label effect and any unexplained order advantage should fall within epsilon once this allocation path is fixed. Agenda timing, sleep stage and evidence-sampling labels do not produce independent effects in the allocation-only model. Continued advantage after withdrawal must be predicted from what was actually corrected, not from the presence of old prices. If bids fail to forecast repair allocation, or fixed allocation leaves a reproducible agenda-, sleep-, or selection-rule-specific residual, reject this mechanism as the sufficient explanation in favor of the appropriate rival.

The agenda lists only which already-present relation must be conveyed; it does not supply its answer. Give an explicit new-recipient ignorance cue in every cell, match overt retrieval and elapsed time, and assess source knowledge on an independent unit set to avoid rehearsal contamination. Define D_A as the C-R minus R-C omission difference. The candidate predicts D_A>delta with no pre-opened agenda, but |D_A|<epsilon when the same agenda is opened before C; opening it only after C fails to rescue within the prespecified interval. Crucially, the correction arm has equal or higher private source accuracy but lower source-specific receiver accuracy. The correction-by-agenda interaction remains after matching trial-level attention, source-memory discrimination, total message length and ordinary given/new status. Uniform correction-time allocation, unchanged source-selection rules and a no-sleep immediate session do not remove it. Subsequent human-only descendants retain the between-lineage semantic advantage over a pilot-defined horizon; exact-artifact yoking abolishes predecessor-history differences. An absence of agenda timing specificity with a selective bundle-allocation effect favors IH_02; a sleep-stage/cue reversal favors IH_03; an explicit sampling-rule reversal favors IH_04. If ordinary audience-design/retrieval models already predict the whole interaction, remove the new-family claim even if the best schedule remains R-C.

Sleep may favor correction after reconstruction by stabilizing source-role bindings predicts instead: Independently tag source-relation encoding and the human reconstruction episode with counterbalanced nonsemantic cues. Cross C-R versus R-C with a matched nap/quiet-rest interval, then, in the physiological validation, deliver source-episode versus reconstruction-episode cues during verified slow-wave sleep, with sham and waking-cue controls. Hold clock time, interference, cue count, arousal and source information constant or explicitly counterbalance them. Define D_S as the post-interval R-C minus C-R probability of preserving the arbitrary source relation, subtracting each schedule's immediate baseline on independent items. Predict a stage-by-cue-by-order interaction exceeding delta: stabilizing a correctable source-bound episode benefits R-C, whereas selectively stabilizing a conflicting reconstruction reduces or reverses that advantage. The corresponding effect is weaker than delta in waking/sham controls under this candidate's stated scope. Entity-only retention weakens the interaction despite matched recognition. One-step kernels estimated separately for sleep and wake plus standard retrieval, spacing and targeted-memory-reactivation effects are the nearest strong rival; an extra order-by-binding state must improve held-out prediction beyond them. If only a nonspecific sleep main effect occurs, or a standard replay model already predicts the full contrast, remove the proposed distinct family. If awake agenda timing or package allocation predicts the advantage and verified stage-specific cueing adds no informative effect, prefer IH_01 or IH_02. After withdrawal, test both original individuals and naive human chains; exact-message replay must eliminate inaccessible predecessor-sleep effects.

Correction order may bias later rule meanings through beliefs about why examples were chosen predicts instead: Use invented narrative worlds with a general rule and explicitly indexed exceptions; correction supplies a true, partially diagnostic source example rather than simply spelling out the entire answer. Pilot that the example is understood and supports the prespecified source interpretation. Cross C-R/R-C with a transparent random-example selection rule versus error-triggered selection. Construct worlds where the exact observed corrective example has the same surface form and source truth under both sampling rules but different evidential relevance to the unsupplied generalization. Everyone receives the same number of source examples and equal time. Establish the ordinary sampling-conditioned response on independent one-step items. Then use new, exchangeable source units to test whether previously experienced correction order changes the inferred selection policy despite an explicitly reset current policy. Define D_Q as the order difference in a held-out scope-transfer error, with exact fact recall held equivalent. The candidate predicts D_Q changes sign between general-rule and genuine-exception worlds, and an experimentally verified policy reset reduces the excess order effect below epsilon. The extra-history term is required only if order-linked policy carryover predicts new-source errors beyond the known sampling-conditioned Bayesian baseline; a mere random-versus-selected example effect is established inference, not a new memetic result. Opening a communicative agenda cannot rescue the private scope error while leaving the selection interpretation unchanged; neither a bundle-allocation change nor a sleep cue has a necessary effect when its proposed state is clamped. Falsify in favor of IH_01 if only output omission changes with intact private scope, or in favor of IH_03 if only the physiological cue interaction remains after sampling comprehension is verified.

SCOUT 1 — From experimental sleep physiology: an order advantage is set by which source-relation episode is stabilized during offline reactivation, not solely by the final waking interpretation. Reconstruction makes a particular role binding retrievable. A source-grounded correction immediately attached to that episode can replace its usable relation before sleep-dependent stabilization; correction presented before an incompatible reconstruction can instead leave the last reconstructed wrong relation preferentially stabilized. The required retained information is an episodic binding between the source's entities and their roles, not mere word recognition. The hypothesis therefore favors R-C before an informative offline interval, but predicts that selective reactivation of the competing episode can reverse the advantage. The candidate extra state is a experimentally tagged, physiologically gated episode available for stabilization; its tag is not a semantic correction and its effect cannot be inferred from a generic sleep benefit. It stabilizes SPV_4 in the original learners; source-free new recipients can benefit only through the messages those learners later produce. The new cultural conjecture is the order-by-episode-specific stabilization effect on source-specific descendant trajectories, not the existence of sleep consolidation, reactivation or reconsolidation.

Cross C-R versus R-C with a matched nap/quiet-rest interval, then, in the physiological validation, deliver source-episode versus reconstruction-episode cues during verified slow-wave sleep, with sham and waking-cue controls. Hold clock time, interference, cue count, arousal and source information constant or explicitly counterbalance them. Define D_S as the post-interval R-C minus C-R probability of preserving the arbitrary source relation, subtracting each schedule's immediate baseline on independent items. Predict a stage-by-cue-by-order interaction exceeding delta: stabilizing a correctable source-bound episode benefits R-C, whereas selectively stabilizing a conflicting reconstruction reduces or reverses that advantage. The corresponding effect is weaker than delta in waking/sham controls under this candidate's stated scope. Entity-only retention weakens the interaction despite matched recognition. One-step kernels estimated separately for sleep and wake plus standard retrieval, spacing and targeted-memory-reactivation effects are the nearest strong rival; an extra order-by-binding state must improve held-out prediction beyond them. If only a nonspecific sleep main effect occurs, or a standard replay model already predicts the full contrast, remove the proposed distinct family. If awake agenda timing or package allocation predicts the advantage and verified stage-specific cueing adds no informative effect, prefer IH_01 or IH_02. After withdrawal, test both original individuals and naive human chains; exact-message replay must eliminate inaccessible predecessor-sleep effects.

The agenda lists only which already-present relation must be conveyed; it does not supply its answer. Give an explicit new-recipient ignorance cue in every cell, match overt retrieval and elapsed time, and assess source knowledge on an independent unit set to avoid rehearsal contamination. Define D_A as the C-R minus R-C omission difference. The candidate predicts D_A>delta with no pre-opened agenda, but |D_A|<epsilon when the same agenda is opened before C; opening it only after C fails to rescue within the prespecified interval. Crucially, the correction arm has equal or higher private source accuracy but lower source-specific receiver accuracy. The correction-by-agenda interaction remains after matching trial-level attention, source-memory discrimination, total message length and ordinary given/new status. Uniform correction-time allocation, unchanged source-selection rules and a no-sleep immediate session do not remove it. Subsequent human-only descendants retain the between-lineage semantic advantage over a pilot-defined horizon; exact-artifact yoking abolishes predecessor-history differences. An absence of agenda timing specificity with a selective bundle-allocation effect favors IH_02; a sleep-stage/cue reversal favors IH_03; an explicit sampling-rule reversal favors IH_04. If ordinary audience-design/retrieval models already predict the whole interaction, remove the new-family claim even if the best schedule remains R-C.

Reconstruction may improve correction by revealing which repairs are useful together predicts instead: Cross C-R versus R-C with package-versus-separate-target allocation of an equal integration-time budget; everyone sees the same source cards for the same total time and completes the same number of model calls. Include targets whose experimentally defined functional value is additive and targets whose joint correction is complementary, while keeping the proposition rubric fixed. Estimate each participant's complementarity from incentive-compatible choices on independent calibration units. Let D_B be the R-C minus C-R source-specific functional accuracy gap. Predict D_B>delta under separate-target allocation for complementary relations, with a smaller |D_B| under package allocation; the package interaction is near zero for additive units. Manipulating retained relational identity changes measured bundle complementarity and predicts this interaction before seeing the final outcomes. The decisive replay control randomly yokes the exact per-target attention schedule and correction content from package winners to recipients who never bid: the auction-label effect and any unexplained order advantage should fall within epsilon once this allocation path is fixed. Agenda timing, sleep stage and evidence-sampling labels do not produce independent effects in the allocation-only model. Continued advantage after withdrawal must be predicted from what was actually corrected, not from the presence of old prices. If bids fail to forecast repair allocation, or fixed allocation leaves a reproducible agenda-, sleep-, or selection-rule-specific residual, reject this mechanism as the sufficient explanation in favor of the appropriate rival.

Correction order may bias later rule meanings through beliefs about why examples were chosen predicts instead: Use invented narrative worlds with a general rule and explicitly indexed exceptions; correction supplies a true, partially diagnostic source example rather than simply spelling out the entire answer. Pilot that the example is understood and supports the prespecified source interpretation. Cross C-R/R-C with a transparent random-example selection rule versus error-triggered selection. Construct worlds where the exact observed corrective example has the same surface form and source truth under both sampling rules but different evidential relevance to the unsupplied generalization. Everyone receives the same number of source examples and equal time. Establish the ordinary sampling-conditioned response on independent one-step items. Then use new, exchangeable source units to test whether previously experienced correction order changes the inferred selection policy despite an explicitly reset current policy. Define D_Q as the order difference in a held-out scope-transfer error, with exact fact recall held equivalent. The candidate predicts D_Q changes sign between general-rule and genuine-exception worlds, and an experimentally verified policy reset reduces the excess order effect below epsilon. The extra-history term is required only if order-linked policy carryover predicts new-source errors beyond the known sampling-conditioned Bayesian baseline; a mere random-versus-selected example effect is established inference, not a new memetic result. Opening a communicative agenda cannot rescue the private scope error while leaving the selection interpretation unchanged; neither a bundle-allocation change nor a sleep cue has a necessary effect when its proposed state is clamped. Falsify in favor of IH_01 if only output omission changes with intact private scope, or in favor of IH_03 if only the physiological cue interaction remains after sampling comprehension is verified.

SCOUT 2 — From statistical evidence evaluation and ascertainment: people preserve the literal corrective fact but learn the wrong scope of the cultural rule because correction is sampled conditional on their preceding reconstruction. R-C makes a supplied corrective example look like an error-selected exception; C-R can make the same example look like an ordinary draw from the source's rule. Reusing that inferred sampling policy across later propositions yields history dependence even when the current sentence and its remembered origin are identical. The retained prerequisite is the rule describing why this particular source-grounded example was selected and the contextual relation that defines its scope. It is not sufficient to remember who said it. The mechanism is inferential likelihood reweighting, not changing a network, information-routing barrier, proofreader or transmitter's agenda. It predicts that C-R is superior when the source intends a general relation but R-C induces an exceptional-case interpretation; the direction reverses when the original source genuinely describes a rare exception. Thus there is no universal best order. SPV_4 is stabilized by appropriate scope inference, with private functional application potentially changing despite equal literal recall. After correction withdrawal, a scope interpretation can continue through descendants' explicit conditional statements and choices; the hypothesis does not allow recovery of an omitted arbitrary rule without a documented information path.

Pilot that the example is understood and supports the prespecified source interpretation. Cross C-R/R-C with a transparent random-example selection rule versus error-triggered selection. Construct worlds where the exact observed corrective example has the same surface form and source truth under both sampling rules but different evidential relevance to the unsupplied generalization. Everyone receives the same number of source examples and equal time. Establish the ordinary sampling-conditioned response on independent one-step items. Then use new, exchangeable source units to test whether previously experienced correction order changes the inferred selection policy despite an explicitly reset current policy. Define D_Q as the order difference in a held-out scope-transfer error, with exact fact recall held equivalent. The candidate predicts D_Q changes sign between general-rule and genuine-exception worlds, and an experimentally verified policy reset reduces the excess order effect below epsilon. The extra-history term is required only if order-linked policy carryover predicts new-source errors beyond the known sampling-conditioned Bayesian baseline; a mere random-versus-selected example effect is established inference, not a new memetic result. Opening a communicative agenda cannot rescue the private scope error while leaving the selection interpretation unchanged; neither a bundle-allocation change nor a sleep cue has a necessary effect when its proposed state is clamped. Falsify in favor of IH_01 if only output omission changes with intact private scope, or in favor of IH_03 if only the physiological cue interaction remains after sampling comprehension is verified.

The agenda lists only which already-present relation must be conveyed; it does not supply its answer. Give an explicit new-recipient ignorance cue in every cell, match overt retrieval and elapsed time, and assess source knowledge on an independent unit set to avoid rehearsal contamination. Define D_A as the C-R minus R-C omission difference. The candidate predicts D_A>delta with no pre-opened agenda, but |D_A|<epsilon when the same agenda is opened before C; opening it only after C fails to rescue within the prespecified interval. Crucially, the correction arm has equal or higher private source accuracy but lower source-specific receiver accuracy. The correction-by-agenda interaction remains after matching trial-level attention, source-memory discrimination, total message length and ordinary given/new status. Uniform correction-time allocation, unchanged source-selection rules and a no-sleep immediate session do not remove it. Subsequent human-only descendants retain the between-lineage semantic advantage over a pilot-defined horizon; exact-artifact yoking abolishes predecessor-history differences. An absence of agenda timing specificity with a selective bundle-allocation effect favors IH_02; a sleep-stage/cue reversal favors IH_03; an explicit sampling-rule reversal favors IH_04. If ordinary audience-design/retrieval models already predict the whole interaction, remove the new-family claim even if the best schedule remains R-C.

Reconstruction may improve correction by revealing which repairs are useful together predicts instead: Cross C-R versus R-C with package-versus-separate-target allocation of an equal integration-time budget; everyone sees the same source cards for the same total time and completes the same number of model calls. Include targets whose experimentally defined functional value is additive and targets whose joint correction is complementary, while keeping the proposition rubric fixed. Estimate each participant's complementarity from incentive-compatible choices on independent calibration units. Let D_B be the R-C minus C-R source-specific functional accuracy gap. Predict D_B>delta under separate-target allocation for complementary relations, with a smaller |D_B| under package allocation; the package interaction is near zero for additive units. Manipulating retained relational identity changes measured bundle complementarity and predicts this interaction before seeing the final outcomes. The decisive replay control randomly yokes the exact per-target attention schedule and correction content from package winners to recipients who never bid: the auction-label effect and any unexplained order advantage should fall within epsilon once this allocation path is fixed. Agenda timing, sleep stage and evidence-sampling labels do not produce independent effects in the allocation-only model. Continued advantage after withdrawal must be predicted from what was actually corrected, not from the presence of old prices. If bids fail to forecast repair allocation, or fixed allocation leaves a reproducible agenda-, sleep-, or selection-rule-specific residual, reject this mechanism as the sufficient explanation in favor of the appropriate rival.

Sleep may favor correction after reconstruction by stabilizing source-role bindings predicts instead: Independently tag source-relation encoding and the human reconstruction episode with counterbalanced nonsemantic cues. Cross C-R versus R-C with a matched nap/quiet-rest interval, then, in the physiological validation, deliver source-episode versus reconstruction-episode cues during verified slow-wave sleep, with sham and waking-cue controls. Hold clock time, interference, cue count, arousal and source information constant or explicitly counterbalance them. Define D_S as the post-interval R-C minus C-R probability of preserving the arbitrary source relation, subtracting each schedule's immediate baseline on independent items. Predict a stage-by-cue-by-order interaction exceeding delta: stabilizing a correctable source-bound episode benefits R-C, whereas selectively stabilizing a conflicting reconstruction reduces or reverses that advantage. The corresponding effect is weaker than delta in waking/sham controls under this candidate's stated scope. Entity-only retention weakens the interaction despite matched recognition. One-step kernels estimated separately for sleep and wake plus standard retrieval, spacing and targeted-memory-reactivation effects are the nearest strong rival; an extra order-by-binding state must improve held-out prediction beyond them. If only a nonspecific sleep main effect occurs, or a standard replay model already predicts the full contrast, remove the proposed distinct family. If awake agenda timing or package allocation predicts the advantage and verified stage-specific cueing adds no informative effect, prefer IH_01 or IH_02. After withdrawal, test both original individuals and naive human chains; exact-message replay must eliminate inaccessible predecessor-sleep effects.