The Unlearned Pattern

Imagine teaching a thousand people a rule that did not exist last week.

The rule belongs to an invented world: unfamiliar signs, a few rooms, several routes between them. At one junction, either of two exits could lead to a reward. A concealed random assignment decides which. The learners practice the assigned lesson over a fixed training period.

In another laboratory, people who have never encountered those learners explore the same world. They know enough to recognize the junction, but their instructions say nothing about its rewarded exit. We record their first choices before giving feedback. Do those choices favor the exit learned elsewhere?

Across fresh worlds and fresh groups, assign the opposite lesson. If the distant preference follows the change, something more specific than an improvement at solving puzzles needs explaining: the receiver’s response carries an arbitrary detail of someone else’s experience, despite our efforts to close the ordinary routes by which it could arrive.

This article proposes a program to test that possibility, discover which kinds of similarity an effect would follow, and investigate its cause. Such transfer has not been established. The claim we would put to experiment is that acquired history might alter another system’s dispositions through a relation that our present account of communication leaves out.

The question extends Where Choice Enters the World and What the Universe Remembers. The first developed Conditional Ontological Agency Theory (COAT) to examine agency and the consequences of earlier choices; the second considered physical memory and the possibility that authorship concerns a whole history. Here we ask how far a history can reach. That question permits experiments before we settle whether consciousness has powers beyond ordinary biological control.¹²

The experiment does not test COAT’s claims about authentic authorship or global selection. It tests a distinct claim about acquired history, motivated by questions the earlier essays left open.

Where the past is kept

The results below establish ways that physical systems retain and expose their own histories. They supply methods for investigating a carrier, but no evidence that learning transfers between isolated people. The distant-learning proposal stands or falls on its separate experiment.

A system can lose an obvious record while retaining a less accessible one. Physics offers several ways to investigate the difference: separate persistence from successful readout, disturb a suspected carrier, and measure what survives an attempted erasure. These methods would remain useful to the proposed research even if every effect eventually yielded to an ordinary physical explanation.

A birthmark survives the scrambling

In Anton Graf and colleagues’ Quantum Birthmarks: Ergodicity Breaking Beyond Scarring, a chaotic quantum system retains statistical traces of its preparation. Their billiard calculations follow a wavepacket as it spreads through an enclosure. Long-time statistics nevertheless show enhanced overlap with its starting state and the states along its subsequent orbit through Hilbert space.³

For a time-independent Hamiltonian with nondegenerate energies, the infinite-time-averaged return probability is

$$ \overline{P}_{\mathrm{return}} =\lim_{T\to\infty}\frac{1}{T}\int_0^T \left|\langle\psi(0)|\psi(t)\rangle\right|^2\,dt =\sum_n |c_n|^4, \qquad |\psi(0)\rangle=\sum_n c_n|E_n\rangle. $$

The coefficients describe how the preparation occupies the energy eigenstates. To identify a birthmark, compare overlap with that preparation against overlap with an independently chosen state satisfying the appropriate energy and symmetry constraints. Comparing each state’s return to itself answers a different question. Xiaoya and colleagues find enhancement factors approaching two or three under specified random-matrix comparisons; early returns can strengthen the trace, with Heller’s scars as precedent.⁴⁵

An infinite-time average does not specify a practical readout time. Nor does every environmental disturbance erase the same information. Pure dephasing in the energy basis preserves the populations entering the displayed sum; energy relaxation can change them. Phase information may disappear while a trace of preparation remains. An experiment therefore needs an observable, a description of the disturbance, and a time within which the trace can be read.

Kos, Wild, and Nielsen study another form of persistence: magnetic domains stabilized by symmetry-protected zero modes in interacting spin systems. Under their model’s symmetries and preparation conditions, an initial magnetic pattern can resist the thermalization expected from the remaining dynamics.⁶ Protection of a particular part of the state gives the past a means of surviving. Perturbations that remove that protection offer a way to test why it survived.

What a reset leaves behind

The record may also reside outside the system we think we have reset. Gosling and colleagues’ superconducting-qubit work, accepted by Physical Review Letters in October 2026, examines long-lived two-level defects that retain information about earlier preparation. Resetting the qubit can leave those defects changed. They can then affect a later measurement, making the qubit’s present state an incomplete account of its future behavior.⁷

Adriazola and Roszak reconstruct memory kernels for non-Markovian qubit dynamics computationally. A kernel expresses how earlier evolution contributes to later evolution in a reduced description. Recovering it from observations can be severely ill-conditioned: the available trajectories may leave some components undetermined.⁸ Several physical accounts can therefore remain compatible with a good fit. Describing the missing history and locating its carrier are separate experimental tasks.

Pollock and colleagues approach memory through sequences of interventions. In their process-tensor framework, a suitable measurement followed by a fixed fresh preparation can interrupt the system’s direct continuity. Later statistics may still depend on earlier interventions or measurement outcomes, revealing an influence that survived the new preparation. The full operational Markov condition concerns an adequate set of instruments; a single reset test that finds no dependence cannot certify the disappearance of all memory.⁹¹⁰

For our proposal, the useful question is what happens when a suspected carrier is replaced, reset, or disconnected. If an effect survives, the intervention left a relevant cause intact. Locating that cause requires another change to the apparatus and another measurement.

A body can remember its shape

A regenerating organism rebuilds a form after losing some of the structures that made it recognizable. Genes, biochemical signals, mechanical conditions, and electrical organization contribute to the outcome. Turing’s reaction–diffusion account showed how interacting components could generate form without consulting a stored picture of the finished animal.¹¹

Bioelectric experiments show that a tissue’s anatomical dispositions can also change and persist. Durant and colleagues briefly disrupted gap-junction signaling in planaria and observed altered regeneration after the treatment ended. Some animals regenerated two heads. Others looked normal but revealed a changed propensity when cut again. Manipulating their bioelectric state could restore ordinary regenerative outcomes.¹²

Surviving biological material carries the history in these experiments. They provide no demonstration that an untreated worm elsewhere acquires the altered form. What they do reveal is a disposition that a photograph of the animal may miss: its current appearance and its response to a later injury can tell different stories about its state.

Cervera, Levin, and Mafe model stable and oscillating patterns supported by ion channels, cellular voltages, and connections between cells.¹³ Manicka and Levin model how electrostatic feedback and transient inputs steer bioelectric patterns.¹⁴ Connecting those electrical configurations to particular anatomical outcomes remains an experimental task.

McMillen and Levin used fluorescence-lifetime measurements to record bioelectric patterns across cells in frog neural-crest explants over seconds to hours.¹⁵ The study recorded activity within tissue; it did not test transmission between disconnected cultures.

This work leaves us with two questions: where does an acquired disposition reside, and what would another system have to share to acquire it?

Two bodies may reach similar forms because they inherit similar components and obey similar dynamics. Corresponding attractors can explain resemblance without an exchange of experience. Yet shared equations do not guarantee a shared outcome: systems can enter different attractors, and small changes can alter the available forms. The resemblance itself establishes neither convergence as a general law nor communication between the organisms.

The invented world adds a fact that shared anatomy and shared dynamics cannot predict. Its rewarded exit is selected after the task has been built. If another person’s preference follows that selection, the explanation has to account for the assignment itself.

What morphic resonance would have to contribute

Rupert Sheldrake’s hypothesis of formative causation proposes that past systems influence later similar systems through morphic resonance. Repetition would make forms and behaviors more probable, with similarity connecting a present system to relevant predecessors. This influence is meant to extend beyond the inheritance of a genome or tissue, or the receipt of a lesson.¹⁶ Sheldrake gives it a temporal direction: it comes from the past, with no assumed weakening solely because of distance in space or time.¹⁷

To test that proposal, we need to know which past systems count. They might resemble the receiver in appearance, microscopic organization, learned function, or some combination. We also need to know what they must have done. Successful learning, repeated performance, and repeated physical events describe different histories and lead to different comparisons.

Gomez-Marin argues that unresolved biological questions warrant investigation rather than dismissal by association.¹⁸ An investigation also needs a version of the hypothesis that could fail. If the relevant kind of similarity can be selected after every outcome, the experiment has supplied no such test.

The chick studies illustrate the difficulty. Sheldrake and Steven Rose collaborated on an experiment, then published conflicting interpretations. Sheldrake saw support for formative causation in changes across successive cohorts; Rose challenged the comparisons and analysis.¹⁹²⁰ Calendar progression leaves accumulated learning entangled with cohort differences, baseline preferences, handling, and testing conditions. It does not assign a learning dose at random.

Black, Butzer, and Sheldrake’s 2026 Wordle paper follows a different task through three studies. The first found no predicted improvement later in the day. In the second, the evening sample had very small advantages at the earliest cumulative success thresholds and disadvantages at later thresholds, rather than a consistent improvement across the guess distribution. The third failed to replicate the early advantage.²¹ Answer difficulty, participant selection, timing, shared language, and information circulating online further complicate interpretation. An invented task would allow the crucial information to be assigned and concealed from the outset.

We propose to give the source history an arbitrary sign. In one world the source cohort learns one policy; in another, a hidden random choice assigns its opposite. An account of transferred content then has to predict which policy the receiver follows. Alertness, a fortunate cohort, or an ordinary preference for one exit cannot systematically track that fresh concealed assignment.

This content claim goes beyond the idea that repeated activity makes similar activity easier. Facilitation might emerge only while someone learns, without supplying an answer beforehand. We would test that possibility in its own experiment. The principal wager here concerns a choice for which the receiver has received no local evidence about the rewarded answer.

The lesson the receiver has not had

Researchers have used invented regularities to study learning since Reber’s artificial-grammar experiments. Mark and colleagues’ graph-learning work shows how people can transfer structural knowledge between tasks with different appearances.²²²³ In both cases, the learner has experience through which information can arrive. Their methods help separate a rule from its presentation; the proposed influence from another person’s learning remains to be established.

The arbitrary fact

We would construct families of small navigable worlds, each with identifiable relations and diagnostic choices. Source participants explore, receive rewards, and learn a policy. A common scaffold teaches receivers the controls and relevant locations while withholding the outcomes of the diagnostic choices.

Each family has two policies that give opposite answers at those choices. Reward, complexity, and amount of training are matched. A concealed assignment determines which policy the source cohort receives; receiver materials stay identical under both assignments. The first diagnostic response supplies the principal observation.

An eight-choice policy could test whether the response follows a larger pattern. Selecting between two complementary eight-choice policies still assigns only one bit, so the eight answers would not demonstrate eight independently transmitted bits. A later experiment could draw each answer independently and ask whether the receiver recovers the resulting larger message. In either design, responses remain clustered within an assigned family. A designated first choice would remain the initial primary endpoint.

Task construction itself can leak an answer. Path lengths, symbol frequencies, an animation, a filename, or a pseudorandom seed might reveal the favored policy. We would freeze receiver packages before assigning source policies and keep the keys out of receiver software. A separate custodian would later join the assignments to sealed receiver records. Both the visible world and its implementation would be inspected for clues the instructions had withheld.

Novelty reduces prior exposure; it cannot remove the learner’s expectations about space, rewards, symmetry, and other people’s intentions. Those expectations help make an invented world learnable. The critical fact is randomized precisely because they cannot supply it.

A new appearance, the same relation

The first comparison uses the closest available match. Further versions separate a task’s relations from its appearance:

Receiver version	What is preserved	What the comparison asks
Original presentation	The source task’s structure and symbols	Does the assigned content appear under the closest match?
New presentation, same structure	Relations among states and choices; all superficial labels change	Does the effect follow the learned organization?
Similar structure with selected relations changed	A preregistered subset of the source’s relations	Which parts of the organization are necessary?
Familiar presentation, different structure	Symbols and visual style; diagnostic relations change	Can surface resemblance account for the effect?
Untrained family	A matched task whose source lesson has not occurred	Does the effect require the assigned history?

A symmetry can defeat this comparison before any participant arrives. If exchanging the two exits leaves the entire world unchanged, opposite policies may be indistinguishable under the structural relation we intend to test. Anchored, asymmetric scaffolds avoid that problem: starting points, goals, and relational roles identify the alternatives without disclosing their rewards. We would register the source-to-receiver correspondence before collecting data and exclude any allowed symmetry that exchanges the critical answers.

Under this structural hypothesis, a relabeling that preserves all those roles is fully equivalent. Partial similarity has to come from specified changes to the graph. Candidate measures can be compared during development, then fixed for confirmation. Proposed families that are duplicates under the exact role-preserving correspondence would be merged or one rejected before randomization; a new set of labels cannot create an independently acquired structural history.

The outcomes would locate different features of the proposed influence. Survival under a new visual vocabulary, followed by loss when relations change, would point toward task structure. Dependence on familiar images despite structural disruption would direct attention toward sensory resemblance. Each result would say more than a general increase in scores.

When the distant lesson would hinder learning

Distant experience could also lead a receiver astray.

In a separately scheduled learning experiment, receivers would undergo ordinary local training. Some would learn the source policy; others would learn its opposite. A content-specific influence predicts assistance in the first group and systematic interference in the second. A receiver might initially choose the source-favored exit while local feedback rewards the other one.

This experiment needs dedicated fresh families. In the first-choice experiment, every primary receiver response for a family would be sealed before diagnostic feedback was given to anyone. Otherwise the earliest receivers to learn an answer could themselves become sources for those tested later. The feedback embargo would also cover any proposed multi-choice message.

An influence too weak to affect an uninstructed first choice might emerge as the receiver acquires a relevant representation. Independently randomizing local and source policies makes that a testable possibility: does congruence help, and does conflict produce errors in the source’s particular direction? A general change in motivation does not predict that error pattern.

The interpretation would follow the endpoint. A positive learning-curve result would support the declared learning claim while leaving a failed test of uninstructed content failed.

An isolation result and a candidate response model

The experiment asks mathematics to do two jobs: state what ordinary isolation predicts, and specify an additional historical influence precisely enough to test. The isolation proposition applies a standard result. The response law that follows is a conjecture introduced here.

Proposition: changing a source cannot alter an isolated receiver’s marginal

Let $Z$ be the randomly selected source policy. Assume that the joint initial preparation is independent of $Z$, that changing $Z$ acts only on the source, and that no source-to-receiver interaction occurs before readout. Hold the receiver’s preparation, controls, and scoring rule fixed. Do not select trials according to outcomes measured at the source.

Under ordinary classical causal dynamics or quantum dynamics, the receiver’s outcome distribution is then invariant under changing $Z$.²⁴²⁵

For the quantum case, let $\rho_{AB}$ be the initial joint state, $\mathcal E_A^{(z)}$ the source operation, and $\mathcal F_B$ the receiver’s evolution. Both maps are completely positive and trace-preserving. Include each side’s relevant local environment in $A$ or $B$, so that local memory remains compatible with the argument. Exclude a mediating environment that interacts with both sides during the experiment. Then

$$ \rho_B^{(z)} =\operatorname{Tr}_A\!\left[ (\mathcal E_A^{(z)}\otimes\mathcal F_B)(\rho_{AB}) \right] =\mathcal F_B\!\left(\operatorname{Tr}_A\rho_{AB}\right), $$

which contains no $z$.

Proof. Write the source map with operators $M_r^{(z)}$ satisfying $\sum_r M_r^{(z)\dagger}M_r^{(z)}=I_A$. Taking the partial trace over the source and summing all outcomes removes that operation by the completeness relation. The fixed receiver map therefore acts on the same reduced state for every source choice. In a classical causal model, the receiver has no dependence on the intervention variable through a causal path, while the distribution of its other inputs stays fixed.

Initial entanglement can correlate joint outcomes without allowing a local source operation to change the receiver’s unconditioned statistics. The proposition accommodates those correlations. It uses invariance under intervention, which does not require separated quantum outcomes to factor into classical independent probabilities.²⁶

Applying the result to a laboratory requires evidence about its assumptions. People separated by a continent and a day remain connected by many possible ordinary routes. A repeatable change in the receiver’s marginal would call for an account of which assumption failed; the mathematical proof cannot establish that the apparatus excluded every interaction.

Postulate: learned histories contribute a signed influence

Suppose a receiver encounters a diagnostic choice $j$ at time $t$. Let $p_j(t)$ be its probability of choosing the option designated positive in the registered correspondence. We propose the historical contribution

$$ \begin{aligned} H_j(t)&=\sum_{e:\,t_e\lt t} z_{ej}\,w(G_e,G_j)\,k(t-t_e),\\ \operatorname{logit}p_j(t)&=a_j(t)+\lambda H_j(t). \end{aligned} $$

Here $e$ indexes completed source sessions. The sign $z_{ej}=+1$ or $-1$ expresses the randomly assigned source policy in receiver $j$’s fixed correspondence, whether or not the source participant mastered it. The representations $G_e$ and $G_j$ describe the registered task structures. Their similarity weight $w$ lies between zero and one, equals one for an exact role-preserving isomorphism, and is zero for unrelated families. Transporting the sign through the correspondence prevents a changed label from silently reversing the content being tested.

The nonnegative temporal weight $k$ is normalized at a preregistered reference lag $\tau_*$ where the candidate model has a nonzero response: $k(\tau_*)=1$. This allows a delayed onset. We define the model here at a fixed source–receiver separation, over the timelike lag range tested. The term $a_j(t)$ represents local response tendencies, including local training, and is specified independently of source assignment. The coefficient $\lambda$ measures the proposed additional influence per completed session under these weight conventions. The sum describes controlled incremental history in the experiment; unrecorded background history is assumed independent of the source assignment.

This added term is a phenomenological assumption. The quantum results reviewed above supply no derivation of it. Its commitments are that opposite assigned contents contribute with opposite signs and that their influence follows the registered relation between source and receiver. Ordinary isolation gives $\lambda=0$; the proposed aligned-content influence has $\lambda>0$.

Switching source policy across otherwise matched histories reverses the added contribution to the receiver’s log-odds. Adding aligned sessions increases it when the other history and readout time are held fixed, while the logistic response makes probabilities saturate. Opposing histories cancel when their temporal and similarity weights balance. All three statements concern a change relative to the local baseline, so an ordinary preference can remain even when the added contributions cancel.

A lasting accumulation model uses $k(\Delta t)=1$. A fading-memory model weights recent sessions more heavily; that is an alternative proposed here, distinct from Sheldrake’s stated assumption about temporal distance. Teach one policy, wait, then teach its opposite. Responses from fresh receiver cohorts can reveal whether accumulated exposure or recent learning dominates. This temporal reversal combines two lessons in one recorded history. The opening experiment instead assigns opposite policies independently across fresh families. A large opposite lesson that leaves older learning dominant would challenge a model concerned only with recent activity.

Randomization can identify the primary assignment contrast without fitting this kernel. Identifying unique values of $\lambda$, $w$, and $k$ asks more: a restricted baseline, separately varied dose, lag, and structure, and predictions on held-out observations. Normalization fixes arbitrary scale but does not solve that problem. An unrestricted $a_j(t)$ could absorb the entire apparent historical contribution. Nor can the analysis treat earlier receiver responses as harmless adjustments if those responses may already carry the influence under study.

Ordinary task pilots would establish learnability and measurement, not estimate an unestablished distant effect. Before source assignments were disclosed, we would fix the baseline specification and a declared set of candidate kernels and similarity weights. Only $\lambda$ would be estimated for each candidate from confirmatory outcomes, with multiplicity handled in the registered analysis. If discovery-stage data revealed a signal, response mapping could estimate its properties; those choices would then be frozen for held-out tests and independent replication with new families. Transformations and training sequences excluded from the fit would test whether the model predicts more than the data used to construct it.

A completed session is also different from an assigned one. The protocol would randomize the number and kind of sessions offered, record completion and acquisition separately, and retain noncompliance in the primary comparison of assigned programs. A kernel fitted to completed sessions describes realized history and would be secondary unless a fixed, preregistered replacement rule ensured the assigned completed dose. Watching a stimulus, solving a problem, reaching a criterion, and executing a familiar response remain distinct candidate histories. Success-only analyses would be secondary too, because mastery itself is not randomized.

If it carries an arbitrary answer, it carries information

Let $P_+$ and $P_-$ describe the distributions of the decoder’s entire input under the two equally probable source assignments: the raw receiver record together with any supplied, source-independent task identity and registered correspondence. These inputs are coded without the source key. Source-agreement scores cannot serve as decoder inputs because constructing them already uses the assignment we are trying to recover. Standard binary decision theory gives the best possible decoding accuracy as

$$ p_{\mathrm{decode}}^* =\frac{1+D_{\mathrm{TV}}(P_+,P_-)}{2}, $$

where total-variation distance $D_{\mathrm{TV}}$ measures the distributions’ distinguishability. For each output, the optimal rule chooses whichever assignment gives it the greater probability. This is the classical counterpart of the binary discrimination problem in quantum detection theory.²⁷

A reproducible difference between those distributions therefore carries information about source assignment. Any fitted decoder would be trained on separate data, and every decoding rule would be tested on fresh assignments. The theoretical optimum is a limit on performance, not a measurement of what the laboratory achieved. A one-shot advantage also leaves open whether a system with memory can sustain a reliable rate across repeated uses.

Recovering arbitrary content would give a proposed historical influence obligations familiar from communication theory. We would need to explain its carrier or revise the causal account, measure its timing and distance dependence, and reconcile it with observations already explained by existing physics. The name resonance supplies none of those details.

Building the experiment

We propose a staged program whose main round would begin only after feasibility review. Before that round, participating institutions would need a costed recruitment plan, demonstrated source-session throughput, task validation, and simulation-based sample sizes, with resources committed to independent replication before the first main outcome. Its source populations would need enough exposure to test a claim about accumulated experience, while its receiver populations would need enough independent assigned histories to distinguish a small effect from task variation.

Provisional planning ranges are 240–480 independently generated and randomized families per main round, tens of thousands of receiver first responses, and tens of thousands to several hundred thousand source sessions. A high-dose family might receive a thousand sessions distributed across many people. Separate comparisons would distinguish more learners from more repetitions by each learner.

These ranges remain unvalidated for the two-sided test proposed below. Pilots on separate families would estimate how widely baseline preference varies across families, including preferences created by the asymmetric scaffolds needed to identify corresponding choices. Final sample-size simulations would preserve that variation, include participant and site variation, and check sensitivity to stronger dependence. Additional responses within a family cannot substitute for independently assigned histories. At least 90% power for the primary comparison would be a design requirement, not a claim about the current ranges. If that target proved out of reach, recruitment or the number of independently assigned families would have to increase before the main study began.²⁸

The primary estimate would use source-trained families at the prespecified high assigned dose: their equally weighted mean source-policy agreement on the first diagnostic choice in the original presentation, minus one half. With balanced source signs, isolation gives expected agreement of one half even if receivers ordinarily prefer one exit. We would provisionally design for a two-percentage-point increase in agreement. Under those balanced signs, that corresponds to a four-percentage-point difference in raw positive-option choice between the source-positive and source-negative assignments. These numbers specify sensitivity, not an expected effect or an estimate borrowed from resonance studies. The task family carries the randomized assignment; ten thousand responses to a single assigned rule remain one assigned history.

Matched untrained families would receive concealed shadow policies for scoring, with training versus no training also randomized. A claim that source training matters would require a direct trained-versus-untrained contrast. Significance in the trained group and an inconclusive result in the control group would not establish that contrast. The primary causal claim concerns assignment to a source program, without identifying the source learner as its carrier. Which part of that program matters, including the necessity of human learning or later access to the assignment, would require subsequent comparisons, including the post-readout assignment control described below.

The first main round would concentrate resources on substantial source exposure and the distinction between structural and visual resemblance. The latter needs its own powered interaction test: does the source-sign effect differ between a new presentation that preserves the graph and a familiar presentation with changed relations? Separate significance decisions in those conditions cannot answer the interaction question. Dose, lag, temporal reversal, and source-type comparisons would receive separately powered rounds instead of dividing the initial families among every combination.

Development would finish when the worlds were learnable, the withheld answers resisted local inference, ordinary transfer under changed appearances was measured, and recording was reliable. These criteria give the pilots an endpoint independent of finding an unusual distant response.

Dose, acquisition, and exposure

Source participants would demonstrate learning on new instances of the rule and in a delayed retention test. A comparison group would encounter the same sequences without being required to acquire the rule, matching duration, displays, motor activity, and reward as closely as the task permits.

Replay and incidental exposure can themselves teach, so both groups would take the same later mastery test. Active problem solving also changes attention, expectation, errors, and physiology. Comparing the groups estimates the consequences of two training experiences; it cannot isolate consciousness by subtracting one from the other. A machine-source arm could instantiate or learn the policy repeatedly on independent hardware. Humans would remain blind to the machine-selected policy until receiver records were sealed, allowing a later comparison of source programs with and without designated human acquisition.

Each source condition would have its own randomized dose assignments. Mastery and completion tell us what the assigned program achieved. They would be reported alongside its total effect, without removing assigned groups that learned less or failed to finish.

Concealing and auditing the assignment

Receiver materials, schedules, and readout times would be fixed before source training and independently of the chosen policy or the source’s progress. If receiver testing began when the source reached mastery, differences in learning speed could alter test time and create an ordinary route from source assignment to receiver outcome. The schedule would therefore proceed as registered, including when source acquisition fell short.

Source policy would be drawn only after receiver packages were fixed. Receiver staff, participants, and data-quality inspectors would remain blind to policy and dose. Separate implementations and the absence of a live shared service would limit software routes between the two sides. Recruitment screening would check for overlapping participants, and task materials would stay unpublished until the relevant readouts were sealed.

The audit would include people and metadata. A coordinator who knows both conditions, a troubleshooting conversation, or a revealing filename can transmit the answer. An independent team would try to recover source assignments from information available before each diagnostic response: packages, logs, timestamps, schedules, and staff behavior. The receiver responses themselves are the proposed readout and would be analyzed separately. Successful decoding from the preceding inputs would identify a route to remove or measure before making an unexplained-transfer claim.

Calibration would use separate families given weak, deliberate ordinary information. Blinded analysis should recover those signals. Sham assignments with neither source training nor deliberate delivery provide a chance-level check. Source learning without ordinary delivery is the experimental target, so its outcome cannot be stipulated to be chance as another calibration requirement.

Receiver records would be sealed before the assignment key was released for scoring. A cryptographic commitment can establish when a record became fixed; it cannot establish that the record is truthful or that information never entered earlier. Independent implementation, custody, and observation would address those questions separately.

Sealing records before disclosure blocks later ordinary communication from changing those recorded responses. It does not exclude a proposed anomalous influence attributed to someone who learns the assignment afterward, including an analyst at unsealing. The post-readout source-assignment control below probes accounts in which later learning can align an earlier record with the assigned policy; any such account would need to state its prediction for that control.

Accounting for accumulated experience

A history-dependent hypothesis changes the meaning of a control. Tomorrow’s supposedly untreated task may already have been trained today. Receiver exposure can contribute new experience, and reusing materials in replication may change the quantity being tested.

An exposure ledger would record who encountered each family, under which policy, with what feedback, and when. Pilots and confirmations would use fresh families. Opposite-policy training would enter as a deliberate history rather than an assumed washout. Simultaneous controls and randomized schedules would separate assigned training from calendar trends. Families would also be separated under the registered structural metric; residual similarities and cross-family exposure would remain explicit. Renaming a known graph would not make it new.

The ledger records what happened. It is not a license to adjust away source mastery or receiver learning that occurred after assignment. Those events may mediate the effect. The primary estimate concerns the total assigned-program effect, which could be amplified by histories generated at the receiver. It would not establish direct transfer from the original source to every individual receiver.

Hudgens and Halloran, and later Aronow and Samii, develop causal methods for experiments in which one unit’s treatment affects another’s outcome.²⁹³⁰ Although the proposed route here is unusual, the statistical problem of interference has this established precedent. Exposure to other units’ assignments belongs in the design and interpretation instead of being hidden within an independent-trials model.

The principal randomization test would hold raw receiver responses fixed, replay the permissible source assignments, and recalculate source agreement for each reassignment. It would preserve the actual schedule, family blocks, and dose allocations. The global intervention-null says that no assigned source history changes any receiver outcome; its test cannot shuffle individual trials as if each carried its own assignment. Fitting a particular similarity kernel would introduce further spillover assumptions, to be specified and examined in sensitivity analyses.

A source policy chosen after readout

A diagnostic control would reverse the order of assignment and observation. Receivers would respond and their records would be sealed before an independent randomizer drew the source policy and initiated its training. These draws would use fresh random bits independent of every earlier source assignment. Taking the remaining signs from a prebalanced schedule could make a future sign predictable from earlier histories, creating a correlation under interference that this control is meant to avoid. The forward-looking postulate predicts no alignment with the independent later draw.

An apparent alignment would first call for checks of the randomizer, custody, analysis, and any conditioning on later events. It would not identify retrocausality. A clean outcome would answer the procedural question posed by this control, while whole-history models would still need their own explicit predictions.

The ordinary source-first experiment is timelike: light and conventional messages have time to travel between laboratories. Any claim of faster-than-light influence would require another experiment, with suitable spacetime separation and control of the relevant events. A surprising result in the present design could not supply that evidence.

Decisions before disclosure

A registered report would place the protocol and analysis before reviewers while the outcome was unknown.³¹ Proponents of historical transfer, skeptical experimentalists, learning researchers, statisticians, and information-leakage specialists would help design it. Cogitate provides a recent example of competing theoretical commitments examined through a shared preregistered experiment.³² Such collaboration gives disagreement a role in constructing the test, before there is a result to dispute.

The protocol would declare its primary contrast, the postulate’s predicted positive direction, the two-sided test, exclusions, missing-data treatment, stopping rule, and analysis. Structural transfer, learning interference, source type, and lag comparisons would have an explicit order and multiplicity policy. Neural decoding or an alternative similarity metric discovered afterward could motivate another study. They could not replace a failed declared endpoint. Selecting an analysis on the same observations used to judge it creates a well-established inference problem.³³³⁴

For the first main round, we propose a two-sided randomization threshold of $p\leq0.005$, accompanied by a signed effect estimate and a two-sided 99.5% confidence interval constructed under the registered analysis. Given the absence of established evidence for this transfer, leakage or procedural error is a more credible initial explanation for a single threshold crossing. A crossing would justify an information audit and replication of the same signed assignment effect across new families under independent implementation, not establish a transmission route. Alignment has the direction predicted by the postulate; anti-alignment contradicts it. Because anti-alignment was not the predicted direction, a negative-direction finding would require a newly registered replication with that direction declared before it could support further mechanism or structural-transfer claims. Leaving out each site in turn would assess dominance and heterogeneity, not demand significance in every smaller subset. A structural-transfer claim would also require replication of its specific interaction.

An estimate too imprecise to distinguish zero from the design target would call for a newly registered study or an extension already allowed by a preplanned sequential rule. It would not justify collecting indefinitely until significance appeared. A two-sided 99.5% confidence interval lying wholly between minus and plus two percentage points would exclude effects of the design-target magnitude in either direction under the analysis’s assumptions. An interval excluding only the positive target would leave a negative effect of that magnitude unresolved. A reproducible smaller effect might still warrant work if its arbitrary content were secure. Repeated failures at adequate precision would end this version of the proposal in its declared task domain and dose range. Changing the relevant similarity afterward would begin a new hypothesis and a new public record.

From a signal to a cause

Suppose the assigned-content effect survives. We would then need experiments that make it change in predictable ways.

A response map would vary source learning, the interval before receiver testing, physical separation, and the relation between the tasks. Opposite training introduced at a controlled time could weaken a preference, reverse it, or leave it following accumulated exposure. A proposed mechanism would have to account for the pattern across these interventions.

Distance and timing could distinguish candidate carriers. A finite-speed influence predicts a delay related to separation, provided the effect is strong and temporally precise enough to resolve it. Coarse observations that fit a delay-free model say little about instantaneous propagation. Shielding, apparatus changes, and transfers between sites would test particular electromagnetic, acoustic, digital, or material accounts. Each control needs a candidate cause that it would distinguish and a measurement showing what it changed.

The source’s retention offers another comparison. Does the receiver track the number of past learning episodes, the source’s current performance, or recent reactivation of the lesson? Interference tasks and tests after retention declines could separate those histories. An accumulation model can allow the source to forget while its past learning continues to matter. A mechanism that requires an active representation predicts dependence on the source’s present state. The same initial lesson could therefore support different experiments on what remains of it.

Experiments with known carriers

A parallel physics program would create a history whose carrier could be deliberately preserved, moved, or partly erased, then test what a new reader recovers. Its success criteria are independent of the distant-learning outcome: distinguish prepared histories on held-out trials, identify which interventions preserve or erase readable information, and reproduce those results. An adequately precise null that excludes a declared readout accuracy at the tested delays and disturbances also meets those criteria: it establishes limits on what remains readable. A null in the behavioral experiment would leave those results intact.

Simulations could begin with the chaotic billiards and protected magnetic domains discussed earlier. We would introduce disturbances that preserve or disrupt the suspected means of retention, then ask which traces remain readable within a fixed observation period. For the billiard, dephasing and energy relaxation need separate tests; for the magnetic domains, the comparison would include perturbations that break the protecting symmetry. A surviving trace would matter experimentally only if a specified measurement could distinguish the earlier preparations. These tests would connect long-time mathematical persistence to information a later observer could recover.³⁶

Structured quantum environments allow a system’s current state and its environmental history to be manipulated separately. White and colleagues have already demonstrated process-tensor characterization and improved control on a quantum processor with memory.³⁵ A source qubit could write one of two randomized preparations into a characterized environment. After resetting or replacing the source, a fresh second probe would attempt to distinguish those preparations on held-out trials. Varying the delay, perturbing the environment, and replacing the probe would reveal which prepared distinctions remained readable. Process-tensor interventions and kernel reconstruction could help identify the relevant temporal dependence.⁷⁹⁸ The characterized environment carries the information between probes. The experiment asks what it retains after a reset and what the second probe can read.

Living tissue offers a corresponding program. Researchers could write a reproducible bioelectric pattern, disturb selected electrical or intercellular routes, and measure the subsequent developmental response. Existing planarian manipulations and optical voltage measurements provide systems in which to begin.¹²¹⁵ Extending the test to physically separate cultures would introduce another claim. It would require newly assigned distinguishable target patterns, independent batches, and a separate account of transmission; spontaneous resemblance between two regenerates would leave the arbitrary-content question unanswered.

These platforms allow persistence, transfer, and successful readout to be changed separately. They would also show, under controlled conditions, how an overlooked environment can produce an apparently unexplained dependence on history. That is knowledge the distant-learning experiment cannot obtain from a behavioral signature alone.

Which physical organization is shared?

An effect that survives changes of appearance while following relational structure would justify tests across different media. The source might navigate a layout while the receiver encounters an equivalent sequence of sounds. Ordinary within-person transfer would first establish that both tasks express the same learned relation. The distant experiment would then ask whether the proposed influence survives the change of medium.

A graph specifies a correspondence for the investigator. A physical explanation must identify how each system realizes it, what couples those realizations, and why some transformations preserve the effect while others destroy it. Writing a similarity kernel makes those dependencies explicit; explaining their physical origin remains the theoretical task.

Human and machine sources could help locate the relevant organization. Matching signatures would show that acquisition by a designated human source was unnecessary, provided humans had remained blind to the machine’s selected policy until readout was sealed. They would not settle every possible role of consciousness in the experiment. A human-only signature would direct the next comparisons toward biological activity, attention, sleep, and retention. Predictive brain signals would still require intervention tests to establish a causal role. The earlier COAT work on neural control and field effects offers related methods, with its original limits intact.¹

Coincidences need a different instrument

Laboratory assignment cannot encompass every question people ask about shared history. Clusters of dates, names, discoveries, and personal encounters can feel unusually well arranged. That significance to an observer can persist even when no statistical claim has been specified.

Diaconis and Mosteller’s Methods for Studying Coincidences examines the difficulties of moving from an observed match to its apparent rarity. Hidden causes, remembered successes, flexible resemblance, and the number of opportunities all matter.³⁶ The striking connections are only the numerator. We also need the connections that could have occurred.

A prospective observatory would define public event streams, store records as they first appeared, and keep event dates separate from publication dates. Before a new observation period, it would fix a small set of exact matching rules. Any semantic-similarity model would have its version and thresholds frozen, with future data excluded from training.

Baseline models would preserve ordinary sources of clustering: weekday schedules, seasonal releases, shared sources, professional networks, duplicate reports, and changes in reporting volume. Temporal-network research offers ways to represent these dependencies.³⁷ Several candidate baselines could be assessed on earlier periods before testing the future interval. Simulations containing planted departures would measure what the detector could find.

A replicated excess would establish that the selected models missed structure in the event stream. Transmission, intention, and global authorship would remain further claims. The observatory could discover patterns worth explaining, while the source experiment would test what changes when an acquired history is deliberately changed. Each project has a question the other cannot answer.

The 1995 American Institutes for Research evaluation of remote viewing provides a related historical distinction. Jessica Utts and Ray Hyman offered sharply different assessments of the laboratory evidence, while the overall report found that operational material had not demonstrated the utility required for intelligence work.³⁸ A statistical anomaly, recoverable arbitrary information, and practical utility require different evidence; identifying a mechanism adds another task. Our proposed policy test aims at recoverable information. A small laboratory advantage, even if reliable, would not establish intelligence-level usefulness.

What becomes of the whole-history wager?

The previous essay considered a possibility about authorship: if it exceeds ordinary functional control, it might concern the global conditions of a history. A subject would be related to a complete lawful history, with no repeated need to add local departures from its dynamics.² That possibility remains distinct from the experimental claim developed here. To connect them, a theory has to do more than describe both in the language of memory.

Wharton and Argaman review quantum models in which boundary conditions and future measurement settings enter an account of an entire spacetime arrangement.³⁹ These models invite questions about whether a sequential causal story exhausts the available descriptions. Their empirical implications depend on the particular model: some global accounts produce statistics indistinguishable from ordinary causal ones. An unexplained correlation alone cannot choose among them.

For a whole-history wager to bear on this experiment, it would need probabilities for the complete arrangement: the randomizer’s output, source learning, receiver choices, and records ultimately analyzed. Those probabilities would have to differ from competitors in a stated way. Preserving the equations obeyed by individual events does not ensure that the ensemble of histories, distribution of settings, or observed no-signalling relations stays the same.

Within the standard classical causal-model framework, Wood and Spekkens examine how explanations of Bell correlations can require fine-tuning to reproduce the observed absence of signalling.⁴⁰ The difficulty arises in the statistics a causal account permits. A global description, like a proposed local interaction, must explain why the observed regularities survive its assumptions.

Global selection by itself predicts neither temporal symmetry nor alignment with a future training assignment. If a theory supplied a probability distribution for the future-assignment control, the data could test it. Until then, the control audits a procedure and poses a demand for a theory; it cannot detect an author outside time.

Ordinary randomization also rests on an assumption: assignment is independent of the relevant prior causes in the way the statistical model requires. This concerns the experiment’s probability structure, without requiring proof of the scientist’s metaphysical free will. A theory that correlates every randomizer with eventual outcomes still owes an account of which correlations should occur. Without one, every possible result can be accommodated afterward.

Experience raises another question. Chalmers’s distinction between functional explanation and subjective experience applies to reading a history as well as storing it.⁴¹ Reproducible access to learned content would leave unexplained why that access feels like anything. And changing a person’s response through a newly discovered channel would establish influence, not by itself the person’s authorship of the response.

COAT distinguishes present control from inherited consequences. Decisions can reshape later dispositions through memory, learning, and institutions. Kane’s account of self-forming actions asks whether some of those decisions also possess a stronger authorship.⁴² An additional route by which dispositions spread would add to the causal inheritance we need to explain. The question of who owns the originating act would remain.

What the program would change

We propose a staged program whose main round would begin only after the feasibility review described above. Its schedule would follow demonstrated recruitment and training capacity, available facilities, and the simulation-based allocations. It would begin with learnable families, verified source acquisition, concealed assignments, and simulated statistical sensitivity. The main round would test arbitrary content at substantial source doses. Independent replication would use new families, code, and investigators. Mechanism studies would pursue signatures that repeated, while the known-memory experiments developed ways to locate a history and test its readability.

This is proposed work. Human and tissue experiments would require appropriate institutional review, and participant materials would describe learning tasks without promising extraordinary abilities. Protocols, code, results, and stimuli after their final use would be released for independent examination while protecting participants.

The central possible result is specific: uninstructed responses follow a particular lesson taught elsewhere, and across fresh families they follow its opposite when the assignment changes. Structural transformations would show which relations preserve the dependence. Interference and delayed readout would constrain how it persists. Independent reproduction would give a theory a phenomenon to explain.

Failure would be specific too. Repeated tests with adequate precision could defeat the claim that these histories, at these doses and under these conditions, carry detectable content into these receivers. That version of the proposal could then be relinquished. A smaller effect or a different task relation would require a separately stated question, not a change in what the completed experiment had tested.

Return to the invented junction. Two exits offer equal local evidence. Somewhere else, people have learned which one pays. We cannot yet say that their past reaches the receiver through an unknown route. We can assign the lesson, change its history, alter its presentation, and test whether a first choice follows. If a reproducible dependence survives those changes and the search for ordinary transmission, the problem will have a precise object: information about an arbitrary lesson, available where it was never taught.

GTCode.com Research

References