Testing information continuity after death

Controlled mediumship source attribution and aerospace assurance and experimental programme analysis

Abstract

This study examines whether controlled mediumship can establish reliable target-specific information acquisition and, separately, whether such acquisition can identify a continuing deceased source. The evidence window closes on 31 December 2024. Positive blinded studies, relevant null results, a published meta-analysis and detailed observational investigations are assessed against information control, dependence, scoring specificity and causal discrimination. Above-chance matching is distinguished from anomalous acquisition, and both are distinguished from personal survival. Neurophysiological correlates are evaluated as measurements of state rather than proof of the source of information. Three proposed hypotheses manipulate information exclusivity, test identity consistency and introduce newly assigned challenges requiring identity-specific responses. A multicentre experimental programme specifies target custody, matched decoys, independent factual scoring, hierarchical analysis and prospective decision rules. Separate defence and aerospace sections translate the methodological lessons into testable proposals for information provenance, calibrated reliance, operator-state assessment, machine self-monitoring and independent checking of consequential recommendations. The central contribution is a disciplined account of what evidence can establish, what remains underdetermined and which experiments could materially change the assessment.

Contents

Research orientation

The empirical evidence and competing explanations are evaluated first. Proposed experiments follow. Defence and aerospace applications are then assessed as independently testable research transfers.

1 Research question and complementary scope

The central question is whether controlled mediumship can supply reliable evidence of information access that is difficult to explain through ordinary acquisition, inference and scoring, and whether such evidence can discriminate personal survival from alternative explanations. The emphasis is on informational specificity and source attribution. Experiences associated with cardiac arrest mainly raise questions about timing and physiological conditions; mediumship experiments can manipulate targets, access and judgement more directly. The two research domains are therefore complementary without being evidentially interchangeable.

Mediumship here means a reported interaction in which a living person claims to receive information associated with someone who has died. The term does not prejudge the source. A sitter is the person connected with the deceased individual who may later assess a reading. A reading is the recorded output produced for a designated target. Anomalous information reception describes an apparent performance beyond the explanations controlled by a particular experiment; it is not in itself a causal mechanism.

The most scientifically interesting cases contain verifiable details with low probability under an appropriate baseline. Emotional resonance, stylistic resemblance and a feeling of recognition may be meaningful to participants, but they are not equivalent to exclusive knowledge. A study that aims to investigate continuing personal agency must therefore move beyond the question of whether a narrative seems applicable. It must ask what information was available, to whom, at what time, and through which independently documented route.

This distinction has become more consequential with generative models. A system can reproduce biographical style and answer personal questions from stored records without demonstrating that the original person continues to experience anything. Research into digital representations of deceased people makes this separation ethically and conceptually explicit (Hollanek and Nowaczyk-Basińska, 2024). The comparison offers a useful experimental control: persuasive personal communication can be generated through known informational processes. Any survival hypothesis must identify the additional prediction that differentiates its proposed mechanism.

2 Evidence selection and analytical method

This report is a structured critical review with original experimental proposals, not an independently conducted mediumship experiment or a newly estimated meta-analysis. Searches covered blinded mediumship, anomalous information reception, positive and null controlled studies, psychography, neural correlates, survival hypotheses, preregistration, sample-size planning, human-automation interaction and aerospace neuroergonomics. PubMed, journal and publisher records, original papers and institutional repositories were used to verify claims and bibliographic details.

Only works published by 31 December 2024 are cited. Final journal publication dates are used conservatively where online-first and issue dates differ. The principal empirical sources are peer-reviewed papers. A philosophical monograph supports analysis of explanatory inference, and a historical commissioned technical report supports the distinction between laboratory findings and operational utility. These source types are identified rather than presented as equivalent forms of experimental evidence.

The selection prioritises studies with explicit target and control procedures. It includes negative findings and weaker observational designs where they illuminate a specific methodological issue. The report does not claim exhaustive retrieval, a preregistered review protocol or complete access to underlying datasets. Full-text material was examined where retrievable; claims from abstract-only or publisher-excerpt records remain confined to the reported design, outcomes and limitations that those records establish. The numerical audit uses published aggregate counts, not participant-level reanalysis.

Five appraisal dimensions organise the synthesis: integrity of information control; independence of the unit analysed; objectivity and specificity of scoring; transparency of outcome selection; and discrimination among causal explanations. A sixth dimension, operational relevance, is assessed separately for defence and aerospace applications. A laboratory effect can warrant scientific interest while remaining unsuitable for operational use.

The strongest empirical assessment would compare investigators' own interpretations with an explicitly specified alternative model. Authors may reasonably describe a result as evidence of anomalous reception, but the phrase cannot replace an account of the leakage, inference and rating mechanisms actually excluded. Conversely, a critic should identify a plausible mechanism and a testable prediction rather than invoke unspecified bias whenever a result is positive. Both supportive and sceptical models become stronger when they risk being contradicted.

3 From accuracy to personal survival

Three distinct inferential steps

The first step is discrimination: does a reading match its designated target better than matched alternatives? This requires a defined outcome and a defensible null distribution. The second is anomalous acquisition: does the discrimination persist when ordinary information pathways, inference and scoring artifacts are adequately controlled? The third is source identity: does the pattern specifically support a continuing deceased agent rather than an unknown process involving living people, records or other sources?

A positive result at the first step does not automatically establish the second, and neither automatically establishes the third. Sudduth (2016) analyses the auxiliary assumptions required to move from apparent anomalies to survival. A model of continuing personality must specify capacities, access and constraints. Otherwise it predicts almost any result: accurate communication when successful, and failed contact or poor transmission when unsuccessful. An unrestricted alternative involving living-agent anomalous cognition suffers the same weakness. Neither gains explanatory force merely by being difficult to falsify.

The relevant comparison is between models that make different predictions. If a survival model and an unrestricted anomalous-access model assign identical probabilities to every possible observation, the experiment cannot distinguish them. Increased sample size will not repair this structural non-identifiability. Progress requires either constrained models or a new intervention that changes their predictions.

A causal account of the reading

A reading may contain information derived from public records, recruitment contacts, first names, demographic priors, feedback, remembered encounters, chance or an unknown process. Its apparent accuracy then depends on segmentation into claims, the sitter's interpretation, knowledge gaps and the available comparison readings. These pathways form a causal sequence from target selection to final score. Blinding one participant closes only some of them.

For example, keeping the sitter absent reduces conversational cueing but does not automatically remove information acquired during recruitment. Blinding the interviewer to biography does not ensure that target and decoy sets are demographically matched. Blinding a sitter to which transcript is intended does not prevent generic statements from scoring highly across both. Strong designs document each knowledge boundary rather than rely on the number of times the word blind appears in the protocol.

A useful representation separates the generation process from the evaluation process. The medium produces a fixed transcript before scoring. Independently assembled records establish what can be verified. Blinded judges apply a prespecified rubric to target and decoy dossiers. The sitter's felt recognition is retained as a separate outcome. This separation allows a study to discover that emotional recognition and factual discrimination diverge, rather than defining either outcome as a substitute for the other.

Information and identity

Information about a person can persist in records and other people's memories. Continuing personhood is a different claim. A biography may be accurate without being generated by its subject; a distinctive phrase can be copied without conveying the speaker's awareness. Evidence of continuing agency would therefore be strengthened by responses to genuinely new, identity-specific challenges under controlled access, not merely by repetition of accessible autobiographical material.

Even then, the philosophical distinction between a causally continuing person and an extraordinarily accurate reconstruction remains demanding. No short behavioural test resolves every theory of personal identity. An empirical programme can make more limited progress: exclude specified ordinary pathways, demonstrate replicable target specificity and test whether a model of addressable agency predicts the observations better than constrained competitors.

4 Controlled mediumship as a research case

The comparative evidence

Table 1 Comparative evidence

What positive blinded experiments establish

Beischel and Schwartz (2007) separated sitters from readings and used a proxy interviewer unaware of the relevant biographical information. The positive discrimination is therefore more informative than a conversational reading in which the sitter provides continuous feedback. Yet an absent sitter does not logically exclude every hypothetical living-agent source. It excludes ordinary access through the particular interaction and, under additional assumptions, some narrowly specified alternatives. Source identity remains a separate inference.

The 2015 extension is valuable because its outcomes include both global ratings and forced choices. Its published materials require careful denominator handling: the abstract lists 86 readings, the introduction lists 96, and the principal combined scored endpoint contains 58. The reported significance tests were directional. These totals should not be silently harmonised or treated as one independent sample. The publisher record also acknowledges that ten readings were collected during the electrophysiology project, so a synthesis must check overlap before treating publications as independent replications (Beischel et al., 2015).

The appropriate interpretation is neither that a discrepancy invalidates every result nor that positive p-values make the discrepancy irrelevant. A transparent replication would publish the complete flow from attempted reading to usable endpoint, distinguish exploratory material from confirmatory material and assign stable identifiers that permit overlap checks. This would let readers evaluate attrition and dependence directly.

Null results and protocol dependence

O'Keeffe and Wiseman (2005) developed controls addressing ordinary information access, general statements and non-blind judging, and obtained a result that did not support the tested claim. A null study should be judged by its sensitivity and fit to the capability asserted. If proponents argue that its conditions prevented the phenomenon, that objection should generate a prospective comparison between acceptable conditions with equivalent information security. It should not function as a retrospective exemption from every unfavourable result.

Kelly and Arcangel (2011) reported an initial non-significant study and a later positive study using proxy readings and blinded ratings. The contrast cautions against describing the entire field as uniformly positive or uniformly negative. It also illustrates why recruitment, scoring and experimental context need to be treated as potentially active components of the design. Variation can reflect a real moderator, a methodological difference or sampling uncertainty; a small set of studies cannot reliably identify which without planned replication.

The most productive response to protocol disputes is an adversarial collaboration in which proponents and critics jointly approve conditions, declare what would count as evidence in either direction and lock the analysis before results. The design can preserve a participant's customary preparation while separating it from access to target information. Ecological validity and information security are not mutually exclusive, but their compatibility must be demonstrated rather than asserted.

The 2022 result and the meaning of a threshold

Tressoldi et al. (2022) reported 25 correct selections from 38 readings, or 65.8%, against a two-choice chance level of 50%. An independent exact binomial calculation gives a one-sided p-value of 0.0365 and a two-sided value of 0.0730. The authors' directional result is therefore arithmetically consistent with the stated count. A descriptive Wilson 95% interval is approximately 49.9-78.8%, assuming independent Bernoulli observations. Repeated mediums and paired readings mean that assumption is not guaranteed.

This example shows why the design hypothesis and analysis direction must be specified in advance. A one-sided test is not automatically improper when only improvement was predicted, but its interpretation differs from a two-sided test. The scientific conclusion should rest on effect size, uncertainty, dependence and replication, not on moving a result across a conventional threshold by changing the test after inspection.

The publication also assessed subjective modes of receiving information. Such reports can help characterise experience, but a feeling of direct interaction does not independently identify the information source. Phenomenological agency is an outcome requiring explanation, not an instrument that verifies its own interpretation.

Meta analysis and independence

The meta-analysis by Sarraf, Woodley of Menie and Tressoldi (2021) reports an above-chance aggregate estimate and robustness to two publication-bias checks. Its standardised proportion-index result should not be rewritten as a universal 18-percentage-point improvement in raw accuracy or as a probability that survival is true. Different tasks and scoring structures make such conversions unsafe without the underlying transformation and study-specific chance rates.

A meta-analysis can reduce sampling uncertainty while retaining a shared bias. Its effective evidential breadth depends on independent participants, independent investigators, distinct data and comparable endpoints. The overlap signalled in the 2015 paper makes a participant-level audit desirable; it does not establish that the published meta-analysis actually double-counted data. A new synthesis should disclose the overlap matrix and show sensitivity to one effect per independent dataset, laboratory and participant pool.

Tests for publication bias also have limited sensitivity in a small literature. A non-significant diagnostic is not proof that selection is absent. Equally, a hypothetical file drawer is not evidence that the observed effect disappears. The stronger response is prospective registration, publication of all completed sites and a planned pooled analysis. That approach changes the evidence-generating process rather than trying to infer every missing study retrospectively.

A detailed message and the denominator of surprise

Silva et al. (2023) examined a message in detail and classified its statements according to correctness, prior disclosure, inferability and generality. The case is valuable because it makes specific claims available for scrutiny. However, a list containing many apparently unusual matches is not automatically a set of independent low-probability events. Names, relationships, possessions and habits can be correlated, and a striking case may have been selected from a larger body of material.

The proper denominator includes all attempted statements and the process that selected the case. A prospective version would define the dossier, verification rules and comparison pool before generating a message. The distinction is methodological, not dismissive: detailed case investigation generates candidate features for an experiment, while a controlled comparison estimates how unusual their joint occurrence actually is.

5 Neurophysiology and the source of information

A genuine state does not establish an external source

Delorme et al. (2013) recorded electrical activity in six mediums during several conditions, including reported communication, recollection, fabrication and perception. The study found condition-related differences and exploratory relationships involving reported accuracy. Its sample was too small for a reliable population-level biomarker, and usable paired accuracy ratings were not available for every participant. The research supports investigation of distinct subjective states; it does not demonstrate that a neural signature identifies communication with a deceased person.

Peres et al. (2012) compared trance writing with ordinary writing in ten psychographers using cerebral perfusion imaging. Experienced participants showed reduced activity in several regions alongside more complex written material under the study's scoring procedure. These findings challenge a simplistic expectation that more complex output must always require greater measured activity in the same regions. They do not establish the external authorship of the text. Practice, altered attention and task strategy remain possible contributors requiring experimental comparison.

Poldrack (2006) explains why reverse inference from a brain pattern to a specific cognitive process depends on the pattern's specificity and prior probability. The problem is even stronger when the proposed inference concerns the identity of an external communicator. A neural correlate of sincerely feeling contacted cannot establish who, if anyone, supplied the information. An adequate study must jointly measure state, accuracy and access conditions, keeping each inferential step separate.

A joint rather than circular test

A useful neurophysiological substudy would preregister a small set of features, acquire movement and muscle-artifact channels, and evaluate prediction on entirely held-out sessions and people. The outcome should be externally adjudicated target discrimination, not the medium's own label of successful communication. Otherwise the model merely learns to recognise the subjective state used to define the outcome.

Three questions can then be answered independently. Can neural data distinguish task conditions? Can they predict factual discrimination beyond behavioural and demographic baselines? Does any predictive relationship survive altered access controls? A positive answer to the first with negative answers to the others would still be a useful finding about experience. A positive answer to all three would justify a stronger replication, while leaving source identity open.

Crucially, trials must not be retained only when the neural pattern looks promising. Selecting the neural state after learning which readings were accurate would build the desired relationship into the analysis. A training phase can discover candidate features, but the confirmatory phase must use frozen processing and independent outcomes.

6 Original hypotheses and discriminating predictions

The following hypotheses are proposed here as a coherent experimental programme. They are not established results, and the combination is offered as an original research design rather than a claim that no related idea has previously been considered.

H1 Information exclusivity has a measurable effect

The source-exclusivity hypothesis predicts that ordinary-access performance will decline as information moves from publicly available biography to private records, then to securely held information known only to the target person before death. A constrained survival model predicts some persistence of target-specific performance in the last category, conditional on its explicit assumptions about memory, willingness and communication. A constrained living-agent model predicts dependence on what living participants know or can access.

The experiment must not infer exclusivity merely because a sitter says a fact was secret. Each information class needs an access history and independent verification. A fact unknown to the sitter may still exist in a public archive, a shared message or another person's memory. Truly exclusive material is difficult to establish, so the strongest version would require prospective participation while the person is alive, voluntary creation of a target and secure custody of the verification record.

The discriminating prediction is an interaction between information class and experimental access, not simply a positive overall hit rate. If performance falls to baseline as ordinary access closes, that favours the tested ordinary-access account. If a specific residual remains under independent replication, it becomes an explanatory target. An unrestricted hypothesis allowing access to any record, living mind or future feedback remains observationally difficult to distinguish from survival; the experiment cannot eliminate that ambiguity by declaration.

H2 Identity consistency exceeds shared biographical cues

The identity-consistency hypothesis predicts that independent mediums addressing the same target will converge on a preregistered set of distinctive features more strongly than matched unrelated targets, even when common biographical facts and demographic cues are removed. A continuing-agent model could predict stable individual preferences or response dispositions. An ordinary narrative model predicts convergence primarily through common priors and the information supplied by the protocol.

Convergence alone is weak because culturally familiar themes are shared. The proposed endpoint therefore compares cross-session identity discrimination after removal of features available to all participants. Entire target families should be held out during development, preventing a model from learning a family's vocabulary and rediscovering it in validation. Contradictions must count: an identity claim that accumulates matching statements while ignoring incompatible ones is not a stringent test.

Failure of independent sessions to show distinctive, reproducible structure would constrain the specified identity model. Positive structure would support continuity of target-related information, but additional evidence would still be required for continuity of subjective experience. This hypothesis deliberately separates a stable informational signature from a stronger metaphysical conclusion.

H3 A new challenge produces an address specific response

The reciprocal-contingency hypothesis asks whether a purported communicator responds in a way contingent on a newly assigned challenge. A target participant, while alive and voluntarily enrolled, would learn a private rule for mapping a small set of challenge symbols to responses. After death, an independent custodian would randomly assign challenges and target addresses. The medium would then provide a fixed response without feedback. A corresponding living-target arm would test ordinary and proposed living-agent pathways under comparable conditions.

A successful response must depend jointly on the intended identity and the particular challenge. Generic messages, repeated favourite phrases and unsolicited biographical details would not satisfy the endpoint. Address permutations, fictitious-target controls and sham challenges would estimate nonspecific response patterns. The target rule must remain inaccessible to those interacting with the medium and to scoring personnel until outputs are locked.

This design does not logically defeat every possible anomalous-access alternative. A hypothesis allowing direct access to the sealed rule can mimic the result. It does, however, move beyond passive resemblance toward a constrained test of addressable, context-sensitive behaviour. The informative outcome is a reproducible interaction specified before testing. Failure under sufficient sensitivity counts against the enrolled capability, rather than being explained away by adding unlimited post hoc exceptions.

7 Proposed multicentre experimental programme

Separate discovery from confirmation

The programme would begin with feasibility work jointly designed by investigators with different theoretical commitments. Participants would help specify acceptable preparation and session conditions. This phase would establish recruitment, target custody, tolerability, transcription and scoring, without treating exploratory success as confirmatory evidence. A subsequent protocol would be preregistered, ideally through a registered-report process. Preregistration separates prediction from explanation and makes deviations visible (Nosek et al., 2018).

Confirmation would use at least three independently managed sites, an external randomisation custodian and a common locked protocol. Each site should have enough trials to reveal whether a pooled result is dominated by one location. Independent replication means more than repeating the same procedure with the same staff and participants. It requires new target material, separate custody and analysis conducted without access to outcome labels until code is frozen.

Target selection and decoys

The baseline experiment would use paired target and decoy dossiers matched for characteristics that may be inferred from the information supplied. The amount of permitted identifying information should be explicitly manipulated rather than allowed to vary informally. A first name, for example, can convey cohort, linguistic and demographic information. A code-only condition tests a different claimed capability and should be agreed during feasibility rather than introduced as a surprise.

Dossiers would contain documented facts, independently rated specificity and verification sources. They would also record facts that remain unknown. The scoring system should distinguish wrong, correct, too general and unverifiable statements. Unknown does not mean correct, but treating every unknown as false may also distort the result. The main analysis should follow a declared rule, with sensitivity analyses showing how plausible alternative handling changes the inference.

Decoys must be chosen before seeing the reading. Selecting a visibly poor match after reading the transcript creates an artificial advantage. A useful quality check asks blinded judges to predict the intended target from protocol metadata alone. Above-chance performance in that task signals an information leak or an imbalance before any mediumship output is assessed.

Session conduct and chain of custody

The medium would receive only the information allowed by the assigned condition. Sessions would be recorded, with consent, and completed without access to external communication or target files. The experimenter present would not know the relevant biography or correct response. The transcript would be fixed and timestamped before any scoring or disclosure of target details. A full record would preserve false starts, questions and corrections so that only favourable fragments cannot later be selected.

Digital custody requires more than encryption. Scheduling systems, file names, email subjects, document metadata and staff conversations can disclose identity. An independent integrity team would audit these ordinary routes using synthetic test cases before real recruitment. A cryptographic commitment would prove that a response or target file was not revised after unblinding; it would not prove that nobody saw it earlier. Physical and organisational access records remain necessary.

The same principles apply to prospective private-rule material. The participant should consent to the research and its posthumous use without creating an obligation to continue participation during illness. The study should not affect treatment decisions or encourage a belief that successful communication is expected. Participation and withdrawal conditions must be intelligible both to the person and the family.

Scoring specificity without inventing precision

The primary baseline endpoint would be correct target selection under a fixed forced-choice procedure. A secondary factual score would assess individual claims using independent documentation and matched alternative dossiers. Inter-rater agreement would be reported, along with the frequency and handling of disagreements. Emotional resonance and the sitter's overall judgement would remain distinct secondary outcomes.

Specificity can be estimated by asking how often a claim applies to comparable people in a reference sample assembled independently of the tested reading. A rare verified fact is more discriminating than a common one. However, multiplying several item probabilities assumes independence and can grossly exaggerate surprise. A name, profession, place of residence and family structure may be correlated. The preferred analysis therefore permutes whole readings against whole dossiers within matched blocks, retaining the dependence among their details.

Information-theoretic scores can supplement this approach when the reference distribution is known. The information associated with an event of probability q is −log₂(q) bits. That quantity is not a probability that survival occurred, and a product of guessed frequencies is not a defensible information measure. The report of an unusually specific match should always include the provenance of its baseline probability.

Dependence and hierarchical analysis

The reading is not necessarily the independent unit. A medium may complete many sessions, the same target may be assessed repeatedly and sitters may score both target and decoy material. Analyses should preserve those structures. Randomisation inference can change labels only within the blocks authorised by the design. A multilevel model can include medium, target and site effects, but with few clusters its uncertainty may be substantial and should not be hidden by a large number of scored sentences.

A candidate logistic model predicts correct discrimination from access condition, information class and their interaction, with separate variation across mediums, targets and sites.

logit(p) = α + β₁A + β₂E + β₃AE + u + v + w

Here A represents the experimentally assigned access condition, E the prespecified information-exclusivity class, and u, v and w the medium, target and site effects. The coding and baseline chance rate depend on the actual task; the intercept must not be assumed to equal chance in every condition. Priors or regularisation, interactions and random-effect structure would be locked during design. With only a few sites, site-variance estimates are fragile, so site-stratified results and sensitivity to fixed versus random site effects are necessary.

The interaction is central to H1; a grand average above chance cannot establish it. H2 requires cross-session target consistency against matched permutations. H3 requires correct identity-by-challenge contingency. Each has its own estimand and should not be pooled into a single vague score of evidence.

An additional sensitivity model makes the required integrity explicit. Suppose a fraction λ of trials has an unrecognised ordinary information advantage giving success probability q, while the remainder follows chance p₀. Then the expected success rate is (1 − λ)p₀ + λq. In an illustrative two-choice task, 60% success could arise from 20% of trials having perfect ordinary information and the rest being random. This is not an allegation about any published study; it quantifies one competing explanation. Integrity testing should constrain such advantages strongly enough that the hypothesised effect cannot be attributed to plausible contamination, while allowing uncertainty about both λ and q.

The primary endpoint and decision rule should be fixed before recruitment. Secondary hypotheses can use a hierarchical testing sequence or explicit familywise adjustment. All attempted readings, unusable recordings and withdrawals should be reported. Data exclusions must follow rules blind to target correctness. Simmons, Nelson and Simonsohn (2011) provide the broader methodological reason: flexible analyses can manufacture apparent evidence even without deliberate fraud.

Power and evidence thresholds

An illustrative planning calculation for the binary baseline task compares 50% with 60% correct, using 90% power and a two-sided alpha of 0.005. The normal approximation gives about 413 independent evaluable decisions. Assuming ten observations per medium and an intraclass correlation of 0.05 gives a simple design effect of 1.45; allowing for 20% unusable observations yields roughly 750 scheduled readings. These are design assumptions, not estimates of the true effect or a final sample-size recommendation.

The final design must simulate the actual crossed dependence, allocation, missingness and test. It should also specify the smallest effect worth pursuing and the precision needed to exclude that effect. Lakens (2022) provides the rationale for linking sample size to the inferential objective rather than copying a previous small study. The more ambitious source-exclusivity and private-rule experiments require separate power analyses; a trial powered for overall discrimination may be underpowered for an interaction.

Bayesian analysis can compare constrained models and incorporate uncertainty in nuisance mechanisms, provided priors and sensitivity ranges are declared. Posterior odds equal prior odds multiplied by the likelihood ratio for the models under comparison. A p-value against random guessing is not that likelihood ratio and is not a probability of survival. If a model leaves the capabilities of a deceased agent unspecified, its probability predictions cannot be recovered simply by giving it a name.

Outcomes that would change the assessment

A reproducible effect surviving tighter ordinary-access controls, independent factual scoring and out-of-sample replication would justify treating anomalous information acquisition as a serious empirical candidate. A further address-specific response to newly assigned challenges would strengthen a model of contingent agency relative to passive narrative reconstruction. Neither result alone settles all philosophical alternatives.

An effect that disappears when target and decoy dossiers are matched would implicate baseline imbalance. An effect limited to global emotional ratings would suggest a different phenomenon from exclusive factual access. An effect confined to one medium or laboratory would narrow generalisability and motivate a direct replication. A precise null result excluding the preregistered minimum effect would count against the specified capability under the agreed conditions. Each outcome should have an interpretation written before the data arrive.

8 Defence adjacent implications

Laboratory significance and operational utility

Historical evaluation of unconventional information gathering provides a useful caution about translation. Mumford, Rose and Goslin (1995) separated laboratory research from operational applications in their assessment of remote viewing. The review acknowledged statistical laboratory findings while concluding that operational material did not establish useful actionable intelligence. Remote viewing is not mediumship, and that report cannot adjudicate the experiments reviewed here. Its relevance is the explicit separation of an experimental signal from reliable decision support.

The corresponding defence research principle is to evaluate timeliness, specificity, calibration and incremental value against existing methods. An unusual channel with occasional striking matches may still perform poorly when wrong claims impose investigation costs. Demonstrating target discrimination under laboratory scoring rules is not sufficient to justify operational dependence. The burden of validation rises when decisions affect safety or other people's rights.

Information provenance and protection against deception

The most immediate defence-adjacent application is improved evaluation of claims. An analyst should be able to distinguish an observed fact, a sourced report, a model inference and an unsupported assertion. The experimental discipline developed for mediumship makes the path from information to confidence explicit. That discipline is useful even if no anomalous capability is confirmed.

A proposed evaluation environment would use benign synthetic scenarios with known ground truth. It would compare conventional analysis, unsupported intuitive reports and machine-generated narratives without disclosing which source produced which output. The outcomes would include factual accuracy, calibration, time spent checking false leads and whether independent evidence changes the judgement appropriately. This is an assessment of information quality and reliance, not a proposal to use claimed paranormal access for targeting or covert operations.

Generative systems intensify the need for provenance. Farquhar et al. (2024) show that semantic uncertainty can help identify some confabulations in language-model output. Their method is not a universal detector of falsity: a consistently wrong claim may have low semantic variation. Defence-facing systems would therefore need traceable evidence and independent checks in addition to uncertainty estimates. Confidence, fluency and repeated agreement are not interchangeable with verification.

Human research and institutional incentives

The research should not require personnel to endorse a metaphysical position. Recruitment, compensation and participation must be separated from career assessment. Claims involving bereavement require particular care because a study can unintentionally imply that a failed reading represents rejection by a loved one. Results should be communicated as performance under a protocol, with no promise of contact or therapeutic benefit. These are proposed safeguards tied directly to the scientific task, not evidence for or against the underlying hypothesis.

9 Aerospace engineering machine consciousness and interfaces

The shared engineering problem

Controlled mediumship research and aerospace assurance share an abstract problem: a system produces an output whose apparent credibility may exceed the evidence supporting its source. In a cockpit, an explanation can be fluent and still rest on stale sensor data or an incorrect model assumption. In a reading, a statement can feel personally compelling while its route of acquisition remains uncertain. The useful transfer is a method for preserving provenance and testing discrimination, not a claim that the two systems share a physical mechanism.

A proposed aerospace interface would attach each consequential recommendation to an evidence record: observation time, originating sensor or document, transformation applied, uncertainty and conditions under which the recommendation ceases to apply. A human-readable explanation would be generated from that record. The underlying record, not the fluency of the explanation, would determine whether the recommendation is eligible for action.

Machine consciousness as a functional research programme

Dehaene, Lau and Kouider (2017) distinguish information made globally available for flexible use from self-monitoring that supports confidence and error detection. These functions motivate machine architectures with shared task information and explicit estimates of their own competence. They do not establish that implementing the functions necessarily produces phenomenal consciousness. A machine can report confidence, describe an internal state or imitate a person's language without those behaviours settling its subjective status.

For aerospace engineering, the useful functional question is whether an architecture can recognise its own informational limits, revise a recommendation after contradictory evidence and explain the basis for a handover. The proposed system should be tested on these capacities directly. Declaring it conscious adds no assurance unless the claim generates an additional measurable prediction relevant to safe operation.

A shared representation could collect state estimates and candidate explanations from separate modules. A monitoring component would track uncertainty, disagreement and data age. An action component would enforce authorised transitions and explicit constraints. This decomposition prevents a language-based explanation from silently becoming control authority. The design is proposed here, drawing on functional consciousness theory and the separation of automation functions developed by Parasuraman, Sheridan and Wickens (2000).

Formal verification and the limits of guarantees

Katz et al. (2017) demonstrated verification of specified properties in neural-network components associated with an airborne collision-avoidance research system. The work illustrates how a precisely stated property can be proved or a counterexample found within a defined model. It does not certify arbitrary language models, establish complete aircraft safety or guarantee that assumptions about sensors and operating conditions are correct.

The proposed assurance architecture would therefore combine several forms of evidence. Formal analysis addresses bounded properties of a specified component. Simulation examines scenarios and distribution shifts. Human-in-the-loop experiments assess comprehension and reliance. Operational evaluation examines the complete system within its intended environment. A claim passing one layer must not inherit the guarantees of all the others.

The analogy with source attribution is exact at the level of reasoning: a result has a defined scope. A verified controller property does not prove every system behaviour safe, just as above-chance matching does not prove a particular postmortem source. Maintaining that scope makes both fields more rigorous without diminishing their legitimate achievements.

Aerospace neuroscience and operator state

Dehais et al. (2019) studied a portable six-electrode EEG system in real flight, comparing conditions involving monitoring and flying. This is a valuable demonstration that cognitive-state measurements can be attempted beyond the laboratory. The study also exposes practical limits: task role and load are intertwined, and the published classification performance depended on the feature set. The study should not be represented as a universal detector of pilot consciousness or a ready-made safety device.

Dehais et al. (2020) describe why workload, engagement and performance require careful operational definition. A brain signal associated with one condition may not uniquely identify the cognitive state an interface designer wants to infer. This is the same reverse-inference problem encountered when interpreting neural correlates of reported mediumship. Poldrack's (2006) analysis therefore has direct methodological relevance to aerospace neuroscience as well as to anomalous-experience research.

A candidate monitoring system should be evaluated against task-relevant outcomes such as missed information, incorrect mode understanding and delayed recovery. It should retain an uncertain state when data quality is poor or the operator falls outside the calibration population. Cross-person and cross-flight validation are essential. A model that recognises a familiar individual's recording pattern may perform well on randomly mixed windows while failing on a new pilot or a new aircraft.

Trust and interface design

Lee and See (2004) link appropriate reliance to the interaction between automation capability, context and trust. Endsley (2017) identifies persistent problems involving situation awareness, monitoring and being out of the control loop. These findings suggest that an interface should help the operator understand what the system knows, what it assumes and what it needs the operator to decide. Anthropomorphic confidence can be counterproductive if it encourages reliance beyond demonstrated capability.

The proposed interface would distinguish four output types in ordinary language: measured information, retrieved procedural information, model inference and unresolved uncertainty. It would identify contradictory inputs and show what evidence would change the recommendation. An explanation could be concise while retaining links to its source record. A warning about uncertainty should be tied to a useful action, such as checking a sensor or confirming an aircraft mode, rather than becoming a generic disclaimer.

An experimental comparison would hold the underlying recommendation constant while varying provenance, uncertainty presentation and conversational style. The primary outcomes would be appropriate acceptance of correct advice, rejection of incorrect advice and recovery after a misleading recommendation. Self-reported trust would be secondary because a more likable system is not necessarily a safer one.

Commercial and defence applications

Table 2 Research applications

The commercial and defence settings share measurement problems but differ in workload, mission context and acceptable operating envelopes. Transfer between them should be demonstrated, not inferred from a common interface appearance. A classifier or assistant validated in a calm simulator does not automatically generalise to high acceleration or degraded communication. Each proposed application must specify the context in which its evidence is valid.

Two falsifiable engineering hypotheses

H4 Provenance improves calibration more than simulated personality. When advice quality is held constant, an interface exposing source and uncertainty will produce more appropriate reliance than an equally fluent interface that emphasises humanlike identity. The trial would measure acceptance and rejection against known ground truth, rather than assume that lower trust is always better. Failure to improve calibration would count against the proposed interface design.

H5 Independent checking reduces persuasive error. A system in which an action-eligibility check uses evidence independently of the language generator will produce fewer unsafe recommendations reaching the operator than a system relying on the generator's self-assessment alone. The comparator must have matched information access and a clearly defined set of constraints. Testing should include cases where the checking module is wrong, because dependence can reappear if both modules share the same flawed assumptions.

These hypotheses are engineering proposals inspired by the source-attribution problem. Their success would not validate mediumship or machine consciousness. They would establish that a specific architecture or interface improves decisions under stated conditions, which is the relevant engineering claim.

10 Integrated interpretation and research priorities

The controlled mediumship literature contains positive experimental findings, relevant null results and methodological advances that warrant serious, discriminating investigation. The evidence supports a research question; it does not justify presenting communication with deceased persons as an established mechanism. The decisive gap is between performance that appears unusual and a uniquely identified source.

The proposed programme addresses that gap by varying information availability, testing distinctive identity across independent sessions and introducing newly assigned challenges. Its contribution is a set of constrained comparisons whose outcomes can change the assessment. It also specifies where discrimination remains impossible: unconstrained models of anomalous access can mimic survival, and no behavioural resemblance alone proves continuity of a conscious subject.

Research priorities are therefore independent replication, complete accounting of observations, secure target custody and explicit source models. Neurophysiology can contribute by measuring state and predicting independently scored accuracy, provided it is not used to certify its own subjective labels. Generative models can contribute as controlled baselines and as tools for transcription, provided they do not invent evidence, select favourable interpretations or score themselves.

For defence and aerospace, the strongest near-term relevance is epistemic and engineering discipline. Systems should preserve the difference between an observation and an inference, distinguish confidence from correctness and restrict consequential action to evidence and authority that have actually been established. These contributions remain worthwhile under either a positive or a null outcome in consciousness research. An unexpectedly replicated effect would open a deeper scientific problem; a precise failure would narrow the claims that remain credible. Both outcomes improve the quality of inquiry.

Appendix A Statistical audit and planning calculations

For the 25 of 38 selections in Tressoldi et al. (2022), the exact one-sided probability under independent fair choices is the sum of the binomial probabilities for 25 through 38 successes. It equals 0.0364757. The corresponding exact two-sided probability is 0.0729514. The Wilson 95% interval is 0.4989-0.7879. These are sensitivity checks on the published count, not a replacement for analysis preserving the study's repeated and paired observations.

The planning calculation uses p₀ = 0.50, p₁ = 0.60, two-sided alpha = 0.005 and power = 0.90. The standard normal approximation, combining the null standard deviation with the alternative standard deviation, gives n ≈ 412.64, rounded to 413 independent evaluable decisions. A simple equal-cluster adjustment with ten readings per medium and intraclass correlation 0.05 gives a design effect of 1 + 9 × 0.05 = 1.45. Dividing the adjusted requirement by 0.80 gives approximately 749 scheduled readings after rounding, expressed as roughly 750 in the text. This approximation does not model crossed target and site effects.

The meta-analytic estimate in Sarraf, Woodley of Menie and Tressoldi (2021) is quoted in its reported standardised proportion-index metric. No conversion to raw accuracy, probability of survival or universal expected performance is made. The 2015 study's different overall reading totals and endpoint-specific sample sizes are retained explicitly. Any future individual-data synthesis should reconcile those totals and identify overlap with related publications before estimating independent evidence.

Appendix B Evidence and inference boundaries

Table 3 Evidence and inference

References

Beischel, J. and Schwartz, G.E. (2007) 'Anomalous information reception by research mediums demonstrated using a novel triple-blind protocol', Explore, 3(1), pp. 23-27. https://doi.org/10.1016/j.explore.2006.10.004

Beischel, J., Boccuzzi, M., Biuso, M. and Rock, A.J. (2015) 'Anomalous information reception by research mediums under blinded conditions II: replication and extension', Explore, 11(2), pp. 136-142. https://doi.org/10.1016/j.explore.2015.01.001

Dehaene, S., Lau, H. and Kouider, S. (2017) 'What is consciousness, and could machines have it?', Science, 358(6362), pp. 486-492. https://doi.org/10.1126/science.aan8871

Dehais, F. et al. (2019) 'Monitoring pilot's mental workload using ERPs and spectral power with a six-dry-electrode EEG system in real flight conditions', Sensors, 19(6), 1324. https://doi.org/10.3390/s19061324

Dehais, F., Lafont, A., Roy, R. and Fairclough, S. (2020) 'A neuroergonomics approach to mental workload, engagement and human performance', Frontiers in Neuroscience, 14, 268. https://doi.org/10.3389/fnins.2020.00268

Delorme, A., Beischel, J., Michel, L., Boccuzzi, M., Radin, D. and Mills, P.J. (2013) 'Electrocortical activity associated with subjective communication with the deceased', Frontiers in Psychology, 4, 834. https://doi.org/10.3389/fpsyg.2013.00834

Endsley, M.R. (2017) 'From here to autonomy: lessons learned from human-automation research', Human Factors, 59(1), pp. 5-27. https://doi.org/10.1177/0018720816681350

Farquhar, S., Kossen, J., Kuhn, L. and Gal, Y. (2024) 'Detecting hallucinations in large language models using semantic entropy', Nature, 630, pp. 625-630. https://doi.org/10.1038/s41586-024-07421-0

Hollanek, T. and Nowaczyk-Basińska, K. (2024) 'Griefbots, Deadbots, Postmortem Avatars: on responsible applications of generative AI in the digital afterlife industry', Philosophy & Technology, 37, 63. https://doi.org/10.1007/s13347-024-00744-w

Katz, G., Barrett, C., Dill, D.L., Julian, K. and Kochenderfer, M.J. (2017) 'Reluplex: an efficient SMT solver for verifying deep neural networks', in Majumdar, R. and Kunčak, V. (eds.) Computer Aided Verification. Lecture Notes in Computer Science, 10426. Cham: Springer, pp. 97-117. https://doi.org/10.1007/978-3-319-63387-9_5

Kelly, E.W. and Arcangel, D. (2011) 'An investigation of mediums who claim to give information about deceased persons', Journal of Nervous and Mental Disease, 199(1), pp. 11-17. https://doi.org/10.1097/NMD.0b013e31820439da

Lakens, D. (2022) 'Sample size justification', Collabra: Psychology, 8(1), 33267. https://doi.org/10.1525/collabra.33267

Lee, J.D. and See, K.A. (2004) 'Trust in automation: designing for appropriate reliance', Human Factors, 46(1), pp. 50-80. https://doi.org/10.1518/hfes.46.1.50_30392

Mumford, M.D., Rose, A.M. and Goslin, D.A. (1995) An evaluation of remote viewing: research and applications. Washington, DC: American Institutes for Research. Commissioned technical report, 29 September. https://www.alice.id.tue.nl/references/mumford-rose-goslin-1995.pdf

Nosek, B.A., Ebersole, C.R., DeHaven, A.C. and Mellor, D.T. (2018) 'The preregistration revolution', Proceedings of the National Academy of Sciences, 115(11), pp. 2600-2606. https://doi.org/10.1073/pnas.1708274114

O'Keeffe, C. and Wiseman, R. (2005) 'Testing alleged mediumship: methods and results', British Journal of Psychology, 96(2), pp. 165-179. https://doi.org/10.1348/000712605X36361

Parasuraman, R., Sheridan, T.B. and Wickens, C.D. (2000) 'A model for types and levels of human interaction with automation', IEEE Transactions on Systems, Man, and Cybernetics Part A, 30(3), pp. 286-297. https://doi.org/10.1109/3468.844354

Peres, J.F., Moreira-Almeida, A., Caixeta, L., Leao, F. and Newberg, A. (2012) 'Neuroimaging during trance state: a contribution to the study of dissociation', PLOS ONE, 7(11), e49360. https://doi.org/10.1371/journal.pone.0049360

Poldrack, R.A. (2006) 'Can cognitive processes be inferred from neuroimaging data?', Trends in Cognitive Sciences, 10(2), pp. 59-63. https://doi.org/10.1016/j.tics.2005.12.004

Sarraf, M., Woodley of Menie, M.A. and Tressoldi, P. (2021) 'Anomalous information reception by mediums: a meta-analysis of the scientific evidence', Explore, 17(5), pp. 396-402. https://doi.org/10.1016/j.explore.2020.04.002

Silva, J., Mendes, C.G., Wainstock, B.C., Gomide, M. and Moreira-Almeida, A. (2023) 'Evaluation of the occurrence of anomalous information reception in messages in an allegedly mediumistic process', Explore, 19(6), pp. 785-791. https://doi.org/10.1016/j.explore.2023.03.010

Simmons, J.P., Nelson, L.D. and Simonsohn, U. (2011) 'False-positive psychology: undisclosed flexibility in data collection and analysis allows presenting anything as significant', Psychological Science, 22(11), pp. 1359-1366. https://doi.org/10.1177/0956797611417632

Sudduth, M. (2016) A philosophical critique of empirical arguments for postmortem survival. Basingstoke: Palgrave Macmillan. https://doi.org/10.1057/9781137440945

Tressoldi, P., Liberale, L., Sinesio, F., Bubba, V., Pederzoli, L. and Testoni, I. (2022) 'Mediumship accuracy: a quantitative and qualitative study with a triple-blind protocol', Explore, 18(4), pp. 411-415. https://doi.org/10.1016/j.explore.2021.05.009