Which explanations of cultural change deserve testing first?

CROSS-DOMAIN TRANSFER — Causal-epoch consensus. Cultural corrections become usable only when recipients reconcile which version of a proposition other contributors have already processed. A repeated majority can be a quorum of stale descendants rather than support for the current meaning. The extra mechanism is an endogenous commitment gate based on perceived causal version and acknowledgment structure: a correction can be understood privately yet excluded from the next public descendant until the receiver believes sufficiently many relevant peers have processed that version. This differs from simple redundancy, error correlation, source credibility and allocation exposure because it depends on correction-before-acknowledgment relations at matched content, dose, marginal source reliability and update opportunity. Let e be a correction event, a_j an acknowledgment by contributor j, and e precede a_j only if the acknowledgment was produced after access to e. The fraction q_e of causally valid acknowledgments, not the fraction of superficially agreeing reports, enters P(commit current meaning)=logit^-1(alpha+beta*q_e+gamma*evidence). beta>0 beyond the matched credibility/common-knowledge baseline is the candidate cultural dependency; the logistic form is an operational model, not a theorem. This stabilizes the semantic transition matrix (SPV_4), potentially with a speed/accuracy cost. It can occur in human-only, mixed human-model and recommendation-mediated relays; neither consensus nor synchronized commitment guarantees truth.

In separate small blocks, randomize whether acknowledgments visibly descend from the correction or from the preceding version, while matching actual display times, report count, total reading time, majority frequency and marginal accuracy. Independently compare verified causal-version labels with equally salient noncausal recency labels; hold the graph's degree and available correction evidence fixed. Primary contrast D_E is the causal-label-by-valid-acknowledgment interaction in corrected proposition reproduction, estimated separately from private endorsement and mere sharing. The conjecture predicts D_E>delta_E plus fewer simultaneous incompatible public descendants when valid correction ancestry is visible; merely moving all wall clocks or adding arbitrary newer-looking labels should not produce the same effect. Measure private source comprehension before an independently randomized public-production opportunity in a subsample, with another subsample spared the comprehension probe to assess rehearsal. Against a fitted recency-weighted Bayesian reliability/common-knowledge coordination model, the proposed commitment gate must predict held-out stale-quorum cases that the rival mispredicts. Remove the extra family if the rival predicts these cases within the declared predictive/equivalence bounds, if verified epoch labels have no residual effect, or if only recency or actual source exposure explains the result. This rejects a proposed new cultural dependency while retaining ordinary consensus effects. The reset-sensitive proofreading rival instead predicts effects of interruptions in a continuously maintained candidate, even without multiple contributors or version ambiguity.

Include intact individual-rule, independent-production and explicitly taught fixed-decoder controls. Pretrain only a common task grammar; do not teach the target constraint or its repair algorithm. Randomize intact versus shuffled role bindings and replace one carrier with a naive participant after amplification withdrawal. The proposed signature is a positive excess recovery contrast D_R=[P(correct ancestry-qualified action after replacement|live,intact)-P(correct action|live,shuffled)]-[the corresponding replay difference], together with recovery of the original constraint under a changed fragment-to-role map. Require D_R>delta_R and an independently replicated increase in conditional joint information; delta_R is the smallest practically meaningful contrast chosen from pilot coding and functional-error costs, not an invented effect estimate. Crucially, a preregistered fixed-decoder/complementary-channel model and a resource-matched individual learning model must underpredict this recovery in held-out seeds, whereas the estimated Psi model predicts it. A raw synergy statistic, a nonsignificant singleton-information test, or a live-versus-replay difference alone is insufficient. Remove the family if fixed joint decoding plus measured individual learning predicts the full intervention, if any intact single carrier or external decoder explains recovery, or if D_R is bounded inside the equivalence region. Epoch-consensus predicts sensitivity to causal update labels; this hypothesis instead requires a selective role-binding effect even when all updates share an unambiguous epoch.

Interruptions may erase source-bound checks and change which meanings people transmit predicts instead: In Module B compare contiguous checks, interleaved checks that clear the candidate/source binding, and interleaved checks with an external progress record that preserves the same binding. Match the exact evidence set and its reliability, total reading/decision time, sequence of check types, reading difficulty and number of generative passes. A neutral-interruption control matches general attention loss without clearing check state. Calibrate checking and reset rates in separate tasks; predict final source-supported transition rates without refitting n to each condition. The mechanism predicts a larger wrong/right commitment ratio after reset-prone fragmentation than after contiguous checking, and selective rescue by a binding-preserving checkpoint. In the idealized per-attempt model, p_wrong/p_right=[(k+r_right)/(k+r_wrong)]^n; this is a conditional quantitative prediction with estimated rates, not a universal numerical cultural law. Compare against a drift-diffusion/sequential-probability-ratio model with measured leakage, ordinary working-memory interference, and the composition of independently fitted human and model editing kernels. Fragmentation effects explained by those rivals do not establish a new family. The decisive extra test asks whether manipulating binding reset changes the commitment slope after evidence and fitted leakage are fixed. Remove this family if a standard accumulator/channel composition predicts both fragmentation and checkpoint rescue in held-out seeds, if no reset-dependent contrast remains within equivalence bounds, or if only added information improves correction. Under IH_02, when provenance is unambiguous and no peer quorum exists, a mere check-binding reset has no specific causal-epoch effect. Under IH_01, destruction of role relations matters even when checks are uninterrupted.

Linked conventions may store reversal history in how they replace one another predicts instead: In the miniature feed/convention alternative, first estimate individual switching thresholds under independent choice. Then randomize convention compatibility links and conduct small nested pressure cycles: rise to h_A, fall to h_B, rise to h_C<h_A, fall to h_B, then revisit h_A. Construct comparison histories with the same final h, total exposures, availability counts, recency distribution and endpoint variant shares but different reversal extrema; record actual differences that cannot be matched rather than adjusting them away after treatment. Reset recommender state and use a frozen prospective exposure replay at the endpoint. Compare intact compatibility links with randomized link reassignment that preserves individual threshold estimates, degree and immediate payoffs; include zero-interaction choice and archive-access controls. Primary outcomes are the multivariate repertoire recurrence distance, the order of individual switches, replacement probabilities at each standardized opportunity, and their response to an overshoot beyond h_A. The family predicts recurrence within a pilot-defined tolerance on closing an admissible loop, selective loss of the inner-loop trace on overshoot, and a link-reassignment effect on switching sequences that an independently fitted Preisach mixture cannot predict. Return-point memory alone is explicitly nondiscriminating because the independent mixture also predicts it. Ordinary reinforcement learning with matched finite memory, source availability, repeated reward and individual heterogeneity is the stronger compositional rival. Remove the family if those calibrated rivals predict the full nested-loop/overshoot/reassignment response, if only unreinitialized allocator memory explains it, or if the claimed interaction contrast is bounded near zero. IH_02 instead depends on causal correction ancestry and should not generate this repertoire-specific nested-loop law once version ambiguity is absent; IH_03 predicts source-bound reset effects without requiring an earlier extremum.

SCOUT 1 — Reset-sensitive semantic proofreading, from biochemical reaction kinetics. A candidate proposition must survive a sequence of evidence checks while it remains bound to the same source interpretation before it is committed to public transmission. An interruption can reset completed checks, so the distribution of uninterrupted dwell times changes which meanings survive even when total evidence, total time, number of checks and model calls are matched. Let k be the check-advance rate, r_y the reset/rejection rate for a candidate semantic state y, and n the number of required stages. A deliberately simplified kinetic model gives per-attempt commitment p_y=(k/(k+r_y))^n; overall yield also depends on the number and duration of attempts. Correct and incorrect propositions may differ in r_y because independent contradiction evidence destabilizes one interpretation. The novel cultural dependency is a reset of intermediate source-bound checking state, separately intervenable from evidence availability and elapsed delay. It is not the general observation that extra criticism can help. The transmitted unit is a single prespecified proposition in an unfamiliar narrative plus its identifiable descendants. The mechanism could lower selected semantic error transitions (SPV_4) while slowing reproduction; it makes no automatic claim about reach or durable adoption.

Match the exact evidence set and its reliability, total reading/decision time, sequence of check types, reading difficulty and number of generative passes. A neutral-interruption control matches general attention loss without clearing check state. Calibrate checking and reset rates in separate tasks; predict final source-supported transition rates without refitting n to each condition. The mechanism predicts a larger wrong/right commitment ratio after reset-prone fragmentation than after contiguous checking, and selective rescue by a binding-preserving checkpoint. In the idealized per-attempt model, p_wrong/p_right=[(k+r_right)/(k+r_wrong)]^n; this is a conditional quantitative prediction with estimated rates, not a universal numerical cultural law. Compare against a drift-diffusion/sequential-probability-ratio model with measured leakage, ordinary working-memory interference, and the composition of independently fitted human and model editing kernels. Fragmentation effects explained by those rivals do not establish a new family. The decisive extra test asks whether manipulating binding reset changes the commitment slope after evidence and fitted leakage are fixed. Remove this family if a standard accumulator/channel composition predicts both fragmentation and checkpoint rescue in held-out seeds, if no reset-dependent contrast remains within equivalence bounds, or if only added information improves correction. Under IH_02, when provenance is unambiguous and no peer quorum exists, a mere check-binding reset has no specific causal-epoch effect. Under IH_01, destruction of role relations matters even when checks are uninterrupted.

Include intact individual-rule, independent-production and explicitly taught fixed-decoder controls. Pretrain only a common task grammar; do not teach the target constraint or its repair algorithm. Randomize intact versus shuffled role bindings and replace one carrier with a naive participant after amplification withdrawal. The proposed signature is a positive excess recovery contrast D_R=[P(correct ancestry-qualified action after replacement|live,intact)-P(correct action|live,shuffled)]-[the corresponding replay difference], together with recovery of the original constraint under a changed fragment-to-role map. Require D_R>delta_R and an independently replicated increase in conditional joint information; delta_R is the smallest practically meaningful contrast chosen from pilot coding and functional-error costs, not an invented effect estimate. Crucially, a preregistered fixed-decoder/complementary-channel model and a resource-matched individual learning model must underpredict this recovery in held-out seeds, whereas the estimated Psi model predicts it. A raw synergy statistic, a nonsignificant singleton-information test, or a live-versus-replay difference alone is insufficient. Remove the family if fixed joint decoding plus measured individual learning predicts the full intervention, if any intact single carrier or external decoder explains recovery, or if D_R is bounded inside the equivalence region. Epoch-consensus predicts sensitivity to causal update labels; this hypothesis instead requires a selective role-binding effect even when all updates share an unambiguous epoch.

Knowing which peers processed a correction may gate its public retelling predicts instead: Use unfamiliar narratives with one experimentally corrected causal proposition. In separate small blocks, randomize whether acknowledgments visibly descend from the correction or from the preceding version, while matching actual display times, report count, total reading time, majority frequency and marginal accuracy. Independently compare verified causal-version labels with equally salient noncausal recency labels; hold the graph's degree and available correction evidence fixed. Primary contrast D_E is the causal-label-by-valid-acknowledgment interaction in corrected proposition reproduction, estimated separately from private endorsement and mere sharing. The conjecture predicts D_E>delta_E plus fewer simultaneous incompatible public descendants when valid correction ancestry is visible; merely moving all wall clocks or adding arbitrary newer-looking labels should not produce the same effect. Measure private source comprehension before an independently randomized public-production opportunity in a subsample, with another subsample spared the comprehension probe to assess rehearsal. Against a fitted recency-weighted Bayesian reliability/common-knowledge coordination model, the proposed commitment gate must predict held-out stale-quorum cases that the rival mispredicts. Remove the extra family if the rival predicts these cases within the declared predictive/equivalence bounds, if verified epoch labels have no residual effect, or if only recency or actual source exposure explains the result. This rejects a proposed new cultural dependency while retaining ordinary consensus effects. The reset-sensitive proofreading rival instead predicts effects of interruptions in a continuously maintained candidate, even without multiple contributors or version ambiguity.

Linked conventions may store reversal history in how they replace one another predicts instead: In the miniature feed/convention alternative, first estimate individual switching thresholds under independent choice. Then randomize convention compatibility links and conduct small nested pressure cycles: rise to h_A, fall to h_B, rise to h_C<h_A, fall to h_B, then revisit h_A. Construct comparison histories with the same final h, total exposures, availability counts, recency distribution and endpoint variant shares but different reversal extrema; record actual differences that cannot be matched rather than adjusting them away after treatment. Reset recommender state and use a frozen prospective exposure replay at the endpoint. Compare intact compatibility links with randomized link reassignment that preserves individual threshold estimates, degree and immediate payoffs; include zero-interaction choice and archive-access controls. Primary outcomes are the multivariate repertoire recurrence distance, the order of individual switches, replacement probabilities at each standardized opportunity, and their response to an overshoot beyond h_A. The family predicts recurrence within a pilot-defined tolerance on closing an admissible loop, selective loss of the inner-loop trace on overshoot, and a link-reassignment effect on switching sequences that an independently fitted Preisach mixture cannot predict. Return-point memory alone is explicitly nondiscriminating because the independent mixture also predicts it. Ordinary reinforcement learning with matched finite memory, source availability, repeated reward and individual heterogeneity is the stronger compositional rival. Remove the family if those calibrated rivals predict the full nested-loop/overshoot/reassignment response, if only unreinitialized allocator memory explains it, or if the claimed interaction contrast is bounded near zero. IH_02 instead depends on causal correction ancestry and should not generate this repertoire-specific nested-loop law once version ambiguity is absent; IH_03 predicts source-bound reset effects without requiring an earlier extremum.

SCOUT 2 — Interaction-stored reversal memory, from disordered materials. Cultural persistence after feed withdrawal can reside in the sequence-dependent compatibility of choices within a repertoire, rather than in a slowly decaying exposure total, a reservoir of forgotten content, or the recommender's retained parameters. Let s_i in {-1,+1} denote the currently enacted member of a paired harmless convention at node i; h the experimentally controlled relative payoff/availability pressure; u_i and l_i the switching thresholds; and J_ij a measured compatibility effect of convention j on i. A minimal interacting hysteron model switches upward when h+sum_j J_ij*s_j>u_i and downward when it falls below l_i. The new cultural claim is that these interactions carry nested reversal history into replacement risk after measured individual propensities and current variant shares have been matched. Mere hysteresis is established and independent hysterons can already show return-point memory. Therefore the candidate extra dependency is history-sensitive switching of linked conventions that exceeds a fitted independent-threshold mixture and resource-matched reinforcement learning. Under a restricted cooperative, quasi-static regime it predicts approximate return to a previous joint repertoire when a nested excursion closes, and erasure of an inner memory when the earlier extremum is exceeded. These are conditional material-model predictions to test, not assumed laws of memes. This mechanism stabilizes competitor-conditioned replacement patterns (SPV_10), which can support ancestry-qualified persistence only if recorded descendants retain the relevant convention.

Then randomize convention compatibility links and conduct small nested pressure cycles: rise to h_A, fall to h_B, rise to h_C<h_A, fall to h_B, then revisit h_A. Construct comparison histories with the same final h, total exposures, availability counts, recency distribution and endpoint variant shares but different reversal extrema; record actual differences that cannot be matched rather than adjusting them away after treatment. Reset recommender state and use a frozen prospective exposure replay at the endpoint. Compare intact compatibility links with randomized link reassignment that preserves individual threshold estimates, degree and immediate payoffs; include zero-interaction choice and archive-access controls. Primary outcomes are the multivariate repertoire recurrence distance, the order of individual switches, replacement probabilities at each standardized opportunity, and their response to an overshoot beyond h_A. The family predicts recurrence within a pilot-defined tolerance on closing an admissible loop, selective loss of the inner-loop trace on overshoot, and a link-reassignment effect on switching sequences that an independently fitted Preisach mixture cannot predict. Return-point memory alone is explicitly nondiscriminating because the independent mixture also predicts it. Ordinary reinforcement learning with matched finite memory, source availability, repeated reward and individual heterogeneity is the stronger compositional rival. Remove the family if those calibrated rivals predict the full nested-loop/overshoot/reassignment response, if only unreinitialized allocator memory explains it, or if the claimed interaction contrast is bounded near zero. IH_02 instead depends on causal correction ancestry and should not generate this repertoire-specific nested-loop law once version ambiguity is absent; IH_03 predicts source-bound reset effects without requiring an earlier extremum.

Include intact individual-rule, independent-production and explicitly taught fixed-decoder controls. Pretrain only a common task grammar; do not teach the target constraint or its repair algorithm. Randomize intact versus shuffled role bindings and replace one carrier with a naive participant after amplification withdrawal. The proposed signature is a positive excess recovery contrast D_R=[P(correct ancestry-qualified action after replacement|live,intact)-P(correct action|live,shuffled)]-[the corresponding replay difference], together with recovery of the original constraint under a changed fragment-to-role map. Require D_R>delta_R and an independently replicated increase in conditional joint information; delta_R is the smallest practically meaningful contrast chosen from pilot coding and functional-error costs, not an invented effect estimate. Crucially, a preregistered fixed-decoder/complementary-channel model and a resource-matched individual learning model must underpredict this recovery in held-out seeds, whereas the estimated Psi model predicts it. A raw synergy statistic, a nonsignificant singleton-information test, or a live-versus-replay difference alone is insufficient. Remove the family if fixed joint decoding plus measured individual learning predicts the full intervention, if any intact single carrier or external decoder explains recovery, or if D_R is bounded inside the equivalence region. Epoch-consensus predicts sensitivity to causal update labels; this hypothesis instead requires a selective role-binding effect even when all updates share an unambiguous epoch.

Knowing which peers processed a correction may gate its public retelling predicts instead: Use unfamiliar narratives with one experimentally corrected causal proposition. In separate small blocks, randomize whether acknowledgments visibly descend from the correction or from the preceding version, while matching actual display times, report count, total reading time, majority frequency and marginal accuracy. Independently compare verified causal-version labels with equally salient noncausal recency labels; hold the graph's degree and available correction evidence fixed. Primary contrast D_E is the causal-label-by-valid-acknowledgment interaction in corrected proposition reproduction, estimated separately from private endorsement and mere sharing. The conjecture predicts D_E>delta_E plus fewer simultaneous incompatible public descendants when valid correction ancestry is visible; merely moving all wall clocks or adding arbitrary newer-looking labels should not produce the same effect. Measure private source comprehension before an independently randomized public-production opportunity in a subsample, with another subsample spared the comprehension probe to assess rehearsal. Against a fitted recency-weighted Bayesian reliability/common-knowledge coordination model, the proposed commitment gate must predict held-out stale-quorum cases that the rival mispredicts. Remove the extra family if the rival predicts these cases within the declared predictive/equivalence bounds, if verified epoch labels have no residual effect, or if only recency or actual source exposure explains the result. This rejects a proposed new cultural dependency while retaining ordinary consensus effects. The reset-sensitive proofreading rival instead predicts effects of interruptions in a continuously maintained candidate, even without multiple contributors or version ambiguity.

Interruptions may erase source-bound checks and change which meanings people transmit predicts instead: In Module B compare contiguous checks, interleaved checks that clear the candidate/source binding, and interleaved checks with an external progress record that preserves the same binding. Match the exact evidence set and its reliability, total reading/decision time, sequence of check types, reading difficulty and number of generative passes. A neutral-interruption control matches general attention loss without clearing check state. Calibrate checking and reset rates in separate tasks; predict final source-supported transition rates without refitting n to each condition. The mechanism predicts a larger wrong/right commitment ratio after reset-prone fragmentation than after contiguous checking, and selective rescue by a binding-preserving checkpoint. In the idealized per-attempt model, p_wrong/p_right=[(k+r_right)/(k+r_wrong)]^n; this is a conditional quantitative prediction with estimated rates, not a universal numerical cultural law. Compare against a drift-diffusion/sequential-probability-ratio model with measured leakage, ordinary working-memory interference, and the composition of independently fitted human and model editing kernels. Fragmentation effects explained by those rivals do not establish a new family. The decisive extra test asks whether manipulating binding reset changes the commitment slope after evidence and fitted leakage are fixed. Remove this family if a standard accumulator/channel composition predicts both fragmentation and checkpoint rescue in held-out seeds, if no reset-dependent contrast remains within equivalence bounds, or if only added information improves correction. Under IH_02, when provenance is unambiguous and no peer quorum exists, a mere check-binding reset has no specific causal-epoch effect. Under IH_01, destruction of role relations matters even when checks are uninterrupted.