Can checking preserve mistaken meanings through repeated retelling?

HERETICAL CANDIDATE — Successful verification becomes an operator on pragmatic scope. Recipients treat a repeatedly and deliberately certified statement as marked evidence of an exception, then transmit the inferred exception as the ordinary rule. Thus genuinely correct, separately checkable cues can increase a source-specific default/exception reversal even while every literal cue is retained. The proposed extra dependency is on a cue having successfully served as a check, rather than on its words, frequency, author identity, or factual reliability: successful checking changes the inferred reason the proposition was worth asserting. The inherited state is an explicit default/exception presupposition in the descendant, not opposition to correction, an ownership claim, or re-encoding the source in a new relational code. This mechanism destabilizes SPV_4; disclosing the noncommunicative generation of checks is predicted to stabilize it. Ordinary redundancy-induced pragmatic enrichment is established; the candidate new claim is a source-specific certification effect beyond a calibrated pragmatic-reconstruction model.

Cross both with deliberate-author versus automatic-rule generation of the same cues; independently randomize a pragmatic-cancellation notice that the repetition conveys no additional typicality information. Keep all subsequent tests and source access fixed. Let Y be a wrong default/exception reversal, V verification, R relational redundancy, I perceived deliberate selection, and C cancellation. The strong prediction is [P(Y|V=1,R=1,I=1)-P(Y|V=0,R=1,I=1)] minus the same difference for automatic checks > delta, with the excess reduced within epsilon by C, even among materials with no detectable one-step literal or cue-validity deficit. Estimate these as randomized contrasts, not by selecting post-treatment correct participants. Source-grounded checking must still show the effect on the prespecified pragmatic proposition; an effect only without source access is weaker evidence. A calibrated ordinary pragmatic model fitted to separate nonchecking utterances and matched certification/attention controls must underpredict the held-out certification contrast. Stable identity tags, changing random switching rate, and forcing additional causal tests should not specifically remove this default/exception error when intention framing is retained. If the pragmatic model already predicts the contrast, or verification has no residual effect within epsilon, remove this as a distinct family and retain ordinary pragmatic reconstruction.

Use the same number and occupancy of formats but randomize their telegraph switching rate nu; equalize trial duration with content-neutral padding and include blocked and very rapid alternation. With parameters fitted on separate fixed-format and transition-probe trials, predict the full first-passage distributions on held-out nu values. The preregistered signature is an interior minimum of mean time T(nu) to the first source-inconsistent committed proposition: T(nu_mid) < min[T(nu_slow),T(nu_fast)]-delta_T, plus a predicted movement of nu_mid when the independently calibrated interpretation-progress timescale changes. There must be acceptable single-format performance and an independently observed first stage on switching, not merely an inverted-U accuracy plot. Private independent reconstruction should retain the switching-rate effect; pragmatic cancellation and identity tagging should not remove it. Fit standard sequential priming, adaptation, serially correlated errors and resource-matched state-dependent transition-kernel composition. If one of these predicts the held-out first-passage curves and timescale shift within epsilon, the stochastic model is a useful representation of established dynamics, not a distinct cultural family. If no barrier separation is achieved, redesign; if achieved separation yields monotonic or correctly baseline-predicted curves, reject the added resonant-activation mechanism.

Checking may carry mistaken identity pairings into later cultural retellings predicts instead: Use narratives with two equally memorable agents and reversible roles, and recipe analogues with two visually distinguishable containers. All source identities and facts remain accessible. Show equivalent rewrite histories with preserved versus disrupted token correspondence, then present identical current drafts for the actual check. Cross this with stable nonsemantic identity tags versus equally salient tags reassigned between rewrites; both arms retain the same explicit identity table, so tags add no new source proposition. Include matched nonchecking rewrite histories to estimate ordinary binding/attention errors. The candidate predicts an excess checking-by-correspondence interaction on complete bijective role-swap errors >delta, little corresponding effect on unary predicate omission or default/exception errors, and selective rescue by stable tags. In the rescue, generic reminders, greater font salience, extra reading time and a second view of the identity table must be separately yoked. The committed swapped mapping must predict the exact next-generation role error beyond source/draft wording and measured initial binding error. A source-grounded audit of identity correspondence should help more than an equally informative extra predicate check. If the fully calibrated one-step binding model composed across rewrites predicts all these errors, or continuity has no effect once current mapping and initial error are fixed, remove the distinct checking-capture family and report ordinary binding errors. Initial failure without a tag manipulation first stage does not falsify the hypothesis.

Inherited test exclusions may hide causal errors despite improving check results predicts instead: In a finite toy-recipe simulator, choose source variants with equal familiar-case outcomes but different outcomes under one prespecified counterfactual intervention a*. Before any loss occurs, train recipients to understand that distinction and validate all possible test readouts. Each generation gets the same number of optional tests, the same simulator and the same current recipe; randomize whether it inherits a predecessor's explicit test-exclusion policy, an exposure-matched unordered record of exactly the same previous tests/results, or a policy replaced by a uniform/diagnostic-coverage rule. All available source facts and past outcomes are identical; only the inherited decision rule differs. First measure the probability of selecting a*, then source-contingency error and performance on withheld interventions. The candidate predicts inherited exclusion lowers pi_g(a*) by >delta_pi and increases later error by >delta_Y beyond a frozen Bayesian active-learning/pedagogical-inference-plus-reinforcement model calibrated in isolated learners with the same records. A randomized policy reset must restore selection and future source-specific performance without altering the text; hold subsequent outcome exposure constant in a yoked arm to show that the policy acts through which evidence is sampled, not a general motivational benefit. Once diagnostic test coverage is externally fixed for all groups, the distinctive inheritance effect should fall within epsilon. Stronger evidence requires persistence of the learned exclusion rule into successors rather than only compliance while a checklist is displayed. If standard social imitation, pedagogical inference and active learning composed with the observed records fully predict these contrasts, or swapping/resetting the policy has no independent effect, remove this as a distinct family and retain the established components.

CROSS-DOMAIN TRANSFER — Resonant activation of a wrong interpretation. Equivalent checking formats alternately make it easy or difficult to abandon the source-consistent interpretation. A random switch that arrives during a partially completed reinterpretation can complete an otherwise rare transition. Consequently, the rate of genuine semantic capture can be maximal at an intermediate switching rate despite identical format frequencies, error marginals and contemporaneous cross-check dependence. The candidate cultural dependency is the coupling between ongoing interpretive progress and externally randomized barrier switching within a checking episode; a static majority decoder or independently composed finished-response kernels lacks that dependence. This is an imported first-passage mechanism, not evidence that a meme is a physical particle. It destabilizes SPV_4 by accelerating the first committed source-specific reversal. No hidden oscillatory clock, periodic scheduling optimum or inherited reversal memory is assumed.

Use the same number and occupancy of formats but randomize their telegraph switching rate nu; equalize trial duration with content-neutral padding and include blocked and very rapid alternation. With parameters fitted on separate fixed-format and transition-probe trials, predict the full first-passage distributions on held-out nu values. The preregistered signature is an interior minimum of mean time T(nu) to the first source-inconsistent committed proposition: T(nu_mid) < min[T(nu_slow),T(nu_fast)]-delta_T, plus a predicted movement of nu_mid when the independently calibrated interpretation-progress timescale changes. There must be acceptable single-format performance and an independently observed first stage on switching, not merely an inverted-U accuracy plot. Private independent reconstruction should retain the switching-rate effect; pragmatic cancellation and identity tagging should not remove it. Fit standard sequential priming, adaptation, serially correlated errors and resource-matched state-dependent transition-kernel composition. If one of these predicts the held-out first-passage curves and timescale shift within epsilon, the stochastic model is a useful representation of established dynamics, not a distinct cultural family. If no barrier separation is achieved, redesign; if achieved separation yields monotonic or correctly baseline-predicted curves, reject the added resonant-activation mechanism.

Cross both with deliberate-author versus automatic-rule generation of the same cues; independently randomize a pragmatic-cancellation notice that the repetition conveys no additional typicality information. Keep all subsequent tests and source access fixed. Let Y be a wrong default/exception reversal, V verification, R relational redundancy, I perceived deliberate selection, and C cancellation. The strong prediction is [P(Y|V=1,R=1,I=1)-P(Y|V=0,R=1,I=1)] minus the same difference for automatic checks > delta, with the excess reduced within epsilon by C, even among materials with no detectable one-step literal or cue-validity deficit. Estimate these as randomized contrasts, not by selecting post-treatment correct participants. Source-grounded checking must still show the effect on the prespecified pragmatic proposition; an effect only without source access is weaker evidence. A calibrated ordinary pragmatic model fitted to separate nonchecking utterances and matched certification/attention controls must underpredict the held-out certification contrast. Stable identity tags, changing random switching rate, and forcing additional causal tests should not specifically remove this default/exception error when intention framing is retained. If the pragmatic model already predicts the contrast, or verification has no residual effect within epsilon, remove this as a distinct family and retain ordinary pragmatic reconstruction.

Checking may carry mistaken identity pairings into later cultural retellings predicts instead: Use narratives with two equally memorable agents and reversible roles, and recipe analogues with two visually distinguishable containers. All source identities and facts remain accessible. Show equivalent rewrite histories with preserved versus disrupted token correspondence, then present identical current drafts for the actual check. Cross this with stable nonsemantic identity tags versus equally salient tags reassigned between rewrites; both arms retain the same explicit identity table, so tags add no new source proposition. Include matched nonchecking rewrite histories to estimate ordinary binding/attention errors. The candidate predicts an excess checking-by-correspondence interaction on complete bijective role-swap errors >delta, little corresponding effect on unary predicate omission or default/exception errors, and selective rescue by stable tags. In the rescue, generic reminders, greater font salience, extra reading time and a second view of the identity table must be separately yoked. The committed swapped mapping must predict the exact next-generation role error beyond source/draft wording and measured initial binding error. A source-grounded audit of identity correspondence should help more than an equally informative extra predicate check. If the fully calibrated one-step binding model composed across rewrites predicts all these errors, or continuity has no effect once current mapping and initial error are fixed, remove the distinct checking-capture family and report ordinary binding errors. Initial failure without a tag manipulation first stage does not falsify the hypothesis.

Inherited test exclusions may hide causal errors despite improving check results predicts instead: In a finite toy-recipe simulator, choose source variants with equal familiar-case outcomes but different outcomes under one prespecified counterfactual intervention a*. Before any loss occurs, train recipients to understand that distinction and validate all possible test readouts. Each generation gets the same number of optional tests, the same simulator and the same current recipe; randomize whether it inherits a predecessor's explicit test-exclusion policy, an exposure-matched unordered record of exactly the same previous tests/results, or a policy replaced by a uniform/diagnostic-coverage rule. All available source facts and past outcomes are identical; only the inherited decision rule differs. First measure the probability of selecting a*, then source-contingency error and performance on withheld interventions. The candidate predicts inherited exclusion lowers pi_g(a*) by >delta_pi and increases later error by >delta_Y beyond a frozen Bayesian active-learning/pedagogical-inference-plus-reinforcement model calibrated in isolated learners with the same records. A randomized policy reset must restore selection and future source-specific performance without altering the text; hold subsequent outcome exposure constant in a yoked arm to show that the policy acts through which evidence is sampled, not a general motivational benefit. Once diagnostic test coverage is externally fixed for all groups, the distinctive inheritance effect should fall within epsilon. Stronger evidence requires persistence of the learned exclusion rule into successors rather than only compliance while a checklist is displayed. If standard social imitation, pedagogical inference and active learning composed with the observed records fully predict these contrasts, or swapping/resetting the policy has no independent effect, remove this as a distinct family and retain the established components.

SCOUT 1 — Visual object correspondence transferred to semantic checking. During successive rewrites, a checker may keep tracking the same displayed entity token while its source referent has changed through aliasing, reordering or role-preserving paraphrase. A wrong bijection between two referents then makes several otherwise independent proposition checks endorse the same role swap. Reusing that correspondence during checking, rather than merely misremembering a word, commits the swapped binding to the next descendant. The state is a concrete assignment pi between source entities and descendant entities; attributes and relational predicates can be correctly retained conditional on the wrong pi. The proposed extra dependence is on continuity of a correspondence established during checking after the current source, draft and all identities are made equally available. This is an assignment error, not an inference about event typicality, a causal-test omission or random barrier crossing. Stable correspondence anchors should stabilize SPV_4 specifically against role permutations.

All source identities and facts remain accessible. Show equivalent rewrite histories with preserved versus disrupted token correspondence, then present identical current drafts for the actual check. Cross this with stable nonsemantic identity tags versus equally salient tags reassigned between rewrites; both arms retain the same explicit identity table, so tags add no new source proposition. Include matched nonchecking rewrite histories to estimate ordinary binding/attention errors. The candidate predicts an excess checking-by-correspondence interaction on complete bijective role-swap errors >delta, little corresponding effect on unary predicate omission or default/exception errors, and selective rescue by stable tags. In the rescue, generic reminders, greater font salience, extra reading time and a second view of the identity table must be separately yoked. The committed swapped mapping must predict the exact next-generation role error beyond source/draft wording and measured initial binding error. A source-grounded audit of identity correspondence should help more than an equally informative extra predicate check. If the fully calibrated one-step binding model composed across rewrites predicts all these errors, or continuity has no effect once current mapping and initial error are fixed, remove the distinct checking-capture family and report ordinary binding errors. Initial failure without a tag manipulation first stage does not falsify the hypothesis.

Cross both with deliberate-author versus automatic-rule generation of the same cues; independently randomize a pragmatic-cancellation notice that the repetition conveys no additional typicality information. Keep all subsequent tests and source access fixed. Let Y be a wrong default/exception reversal, V verification, R relational redundancy, I perceived deliberate selection, and C cancellation. The strong prediction is [P(Y|V=1,R=1,I=1)-P(Y|V=0,R=1,I=1)] minus the same difference for automatic checks > delta, with the excess reduced within epsilon by C, even among materials with no detectable one-step literal or cue-validity deficit. Estimate these as randomized contrasts, not by selecting post-treatment correct participants. Source-grounded checking must still show the effect on the prespecified pragmatic proposition; an effect only without source access is weaker evidence. A calibrated ordinary pragmatic model fitted to separate nonchecking utterances and matched certification/attention controls must underpredict the held-out certification contrast. Stable identity tags, changing random switching rate, and forcing additional causal tests should not specifically remove this default/exception error when intention framing is retained. If the pragmatic model already predicts the contrast, or verification has no residual effect within epsilon, remove this as a distinct family and retain ordinary pragmatic reconstruction.

Random changes in checking format may speed commitment to a wrong interpretation predicts instead: Calibrate two truth-condition-equivalent check formats that produce distinct interpretation-switching barriers while matching full cue information, reading duration and source access. Use the same number and occupancy of formats but randomize their telegraph switching rate nu; equalize trial duration with content-neutral padding and include blocked and very rapid alternation. With parameters fitted on separate fixed-format and transition-probe trials, predict the full first-passage distributions on held-out nu values. The preregistered signature is an interior minimum of mean time T(nu) to the first source-inconsistent committed proposition: T(nu_mid) < min[T(nu_slow),T(nu_fast)]-delta_T, plus a predicted movement of nu_mid when the independently calibrated interpretation-progress timescale changes. There must be acceptable single-format performance and an independently observed first stage on switching, not merely an inverted-U accuracy plot. Private independent reconstruction should retain the switching-rate effect; pragmatic cancellation and identity tagging should not remove it. Fit standard sequential priming, adaptation, serially correlated errors and resource-matched state-dependent transition-kernel composition. If one of these predicts the held-out first-passage curves and timescale shift within epsilon, the stochastic model is a useful representation of established dynamics, not a distinct cultural family. If no barrier separation is achieved, redesign; if achieved separation yields monotonic or correctly baseline-predicted curves, reject the added resonant-activation mechanism.

Inherited test exclusions may hide causal errors despite improving check results predicts instead: In a finite toy-recipe simulator, choose source variants with equal familiar-case outcomes but different outcomes under one prespecified counterfactual intervention a*. Before any loss occurs, train recipients to understand that distinction and validate all possible test readouts. Each generation gets the same number of optional tests, the same simulator and the same current recipe; randomize whether it inherits a predecessor's explicit test-exclusion policy, an exposure-matched unordered record of exactly the same previous tests/results, or a policy replaced by a uniform/diagnostic-coverage rule. All available source facts and past outcomes are identical; only the inherited decision rule differs. First measure the probability of selecting a*, then source-contingency error and performance on withheld interventions. The candidate predicts inherited exclusion lowers pi_g(a*) by >delta_pi and increases later error by >delta_Y beyond a frozen Bayesian active-learning/pedagogical-inference-plus-reinforcement model calibrated in isolated learners with the same records. A randomized policy reset must restore selection and future source-specific performance without altering the text; hold subsequent outcome exposure constant in a yoked arm to show that the policy acts through which evidence is sampled, not a general motivational benefit. Once diagnostic test coverage is externally fixed for all groups, the distinctive inheritance effect should fall within epsilon. Stronger evidence requires persistence of the learned exclusion rule into successors rather than only compliance while a checklist is displayed. If standard social imitation, pedagogical inference and active learning composed with the observed records fully predict these contrasts, or swapping/resetting the policy has no independent effect, remove this as a distinct family and retain the established components.

SCOUT 2 — Active causal discovery and pedagogical sampling: a successful correction procedure is itself transmitted as a rule about which counterfactuals deserve testing. A successor inherits an increasingly narrow test-selection policy alongside an otherwise readable recipe. Redundant confirmations of familiar cases are interpreted as coverage, so checks that distinguish the source-specific causal contingency are increasingly excluded. Later generations can improve their selected-test performance while their actual ability to handle that contingency deteriorates. The extra causal dependency is inheritance of the predecessor's exclusion policy after matching available evidence, current text, check validity, reward, total tests and learning opportunities; it is not simply fewer resources or more correlated answers to fixed questions. The state is a logged distribution pi_g(a) over possible benign interventions a, including active suppression of a discriminating intervention. Restoring policy exploration without changing narrative facts should stabilize SPV_4 and independently measured practice function.

Before any loss occurs, train recipients to understand that distinction and validate all possible test readouts. Each generation gets the same number of optional tests, the same simulator and the same current recipe; randomize whether it inherits a predecessor's explicit test-exclusion policy, an exposure-matched unordered record of exactly the same previous tests/results, or a policy replaced by a uniform/diagnostic-coverage rule. All available source facts and past outcomes are identical; only the inherited decision rule differs. First measure the probability of selecting a*, then source-contingency error and performance on withheld interventions. The candidate predicts inherited exclusion lowers pi_g(a*) by >delta_pi and increases later error by >delta_Y beyond a frozen Bayesian active-learning/pedagogical-inference-plus-reinforcement model calibrated in isolated learners with the same records. A randomized policy reset must restore selection and future source-specific performance without altering the text; hold subsequent outcome exposure constant in a yoked arm to show that the policy acts through which evidence is sampled, not a general motivational benefit. Once diagnostic test coverage is externally fixed for all groups, the distinctive inheritance effect should fall within epsilon. Stronger evidence requires persistence of the learned exclusion rule into successors rather than only compliance while a checklist is displayed. If standard social imitation, pedagogical inference and active learning composed with the observed records fully predict these contrasts, or swapping/resetting the policy has no independent effect, remove this as a distinct family and retain the established components.

Cross both with deliberate-author versus automatic-rule generation of the same cues; independently randomize a pragmatic-cancellation notice that the repetition conveys no additional typicality information. Keep all subsequent tests and source access fixed. Let Y be a wrong default/exception reversal, V verification, R relational redundancy, I perceived deliberate selection, and C cancellation. The strong prediction is [P(Y|V=1,R=1,I=1)-P(Y|V=0,R=1,I=1)] minus the same difference for automatic checks > delta, with the excess reduced within epsilon by C, even among materials with no detectable one-step literal or cue-validity deficit. Estimate these as randomized contrasts, not by selecting post-treatment correct participants. Source-grounded checking must still show the effect on the prespecified pragmatic proposition; an effect only without source access is weaker evidence. A calibrated ordinary pragmatic model fitted to separate nonchecking utterances and matched certification/attention controls must underpredict the held-out certification contrast. Stable identity tags, changing random switching rate, and forcing additional causal tests should not specifically remove this default/exception error when intention framing is retained. If the pragmatic model already predicts the contrast, or verification has no residual effect within epsilon, remove this as a distinct family and retain ordinary pragmatic reconstruction.

Random changes in checking format may speed commitment to a wrong interpretation predicts instead: Calibrate two truth-condition-equivalent check formats that produce distinct interpretation-switching barriers while matching full cue information, reading duration and source access. Use the same number and occupancy of formats but randomize their telegraph switching rate nu; equalize trial duration with content-neutral padding and include blocked and very rapid alternation. With parameters fitted on separate fixed-format and transition-probe trials, predict the full first-passage distributions on held-out nu values. The preregistered signature is an interior minimum of mean time T(nu) to the first source-inconsistent committed proposition: T(nu_mid) < min[T(nu_slow),T(nu_fast)]-delta_T, plus a predicted movement of nu_mid when the independently calibrated interpretation-progress timescale changes. There must be acceptable single-format performance and an independently observed first stage on switching, not merely an inverted-U accuracy plot. Private independent reconstruction should retain the switching-rate effect; pragmatic cancellation and identity tagging should not remove it. Fit standard sequential priming, adaptation, serially correlated errors and resource-matched state-dependent transition-kernel composition. If one of these predicts the held-out first-passage curves and timescale shift within epsilon, the stochastic model is a useful representation of established dynamics, not a distinct cultural family. If no barrier separation is achieved, redesign; if achieved separation yields monotonic or correctly baseline-predicted curves, reject the added resonant-activation mechanism.

Checking may carry mistaken identity pairings into later cultural retellings predicts instead: Use narratives with two equally memorable agents and reversible roles, and recipe analogues with two visually distinguishable containers. All source identities and facts remain accessible. Show equivalent rewrite histories with preserved versus disrupted token correspondence, then present identical current drafts for the actual check. Cross this with stable nonsemantic identity tags versus equally salient tags reassigned between rewrites; both arms retain the same explicit identity table, so tags add no new source proposition. Include matched nonchecking rewrite histories to estimate ordinary binding/attention errors. The candidate predicts an excess checking-by-correspondence interaction on complete bijective role-swap errors >delta, little corresponding effect on unary predicate omission or default/exception errors, and selective rescue by stable tags. In the rescue, generic reminders, greater font salience, extra reading time and a second view of the identity table must be separately yoked. The committed swapped mapping must predict the exact next-generation role error beyond source/draft wording and measured initial binding error. A source-grounded audit of identity correspondence should help more than an equally informative extra predicate check. If the fully calibrated one-step binding model composed across rewrites predicts all these errors, or continuity has no effect once current mapping and initial error are fixed, remove the distinct checking-capture family and report ordinary binding errors. Initial failure without a tag manipulation first stage does not falsify the hypothesis.