The Dead Remain in the Predictive Brain

Software: Bereavement Constraint-Propagation EEG Observatory (ASO) · Build aso-1.0.0 · Run RUN-89d75738712b · CONFIRMATORY

Abstract

This analysis was executed on the SubjectJ LRHand motor-task recording used strictly as a PIPELINE SMOKETEST. The remapped condition labels carry no bereavement meaning and these results must never be interpreted as grief/bereavement evidence. We test whether deceased-related neural states are discriminable from matched controls, and whether such discriminability persists without concurrent conscious report.

Maximum permitted claim for this run: LEVEL 1 — OBSERVED. OBSERVED: condition-dependent differences in EEG features were present (e.g., band power and connectivity varied across conditions).

Validation status: PASS. Primary discrimination: AUC = 0.722 (95% CI 0.682–0.761), balanced accuracy = 0.663, sensitivity = 0.366, specificity = 0.959, permutation p = 0.0196 (50 permutations), Cohen's d = 1.086, n = 900.

Keywords

bereavement; predictive processing; prediction error; EEG; computational neuroscience; neural dynamics; person representation; attachment; autobiographical memory; conscious awareness; information persistence; constraint propagation; longitudinal adaptation; neural connectivity; phase-locking value; machine learning; falsification; autonomous scientific observer

Introduction

Bereavement may involve a distributed constraint-propagation process in which explicit semantic knowledge that a person has died updates rapidly, while attachment, autobiographical, predictive, contextual, and action-related representations continue to generate deceased-related activity until an irreversible future-availability constraint propagates through the person-model.

This instrument does not assume the hypothesis. It tests it, and it preserves negative, ambiguous, and falsifying outcomes. No conclusion exists independently of its evidentiary dependency graph.

Theoretical Framework

The person-model is represented by distributed components (memory, causal history, attachment, identity) each carrying a learned expectation of future physical availability. Bereavement is modeled as the propagation of the update FuturePhysicalAvailability: 1 → 0 through these components.

Formal Model

Person_memory = 1; Person_causal_history = 1; Person_attachment_representation = 1; Person_identity_representation = 1; FuturePhysicalAvailability: 1 → 0.

Person representation, future physical availability, conscious report, EEG observation, autonomic observation, and model inference are maintained as separate constructs and never collapsed.

The ASO Observer Framework (Scientific Disclosure)

What ASO is. The Autonomous Scientific Observer (ASO) is the observer/verification framework that governs this analysis. It is not a signal-processing step; it is a closed-loop supervisor that binds every reported conclusion to its evidentiary dependency graph. Its master rule is: no conclusion exists independently of the evidence that supports it, and if a required dependency fails the conclusion must automatically contract to the strongest statement the surviving evidence still supports.

Why it is used. EEG decoding pipelines can manufacture apparent effects through leakage, flexible preprocessing, multiple comparisons, or overstated language. ASO exists to make such incorrect results difficult to produce, difficult to hide, and easy to audit — preventing both false confirmation and false rejection.

Where it sits. ASO wraps the closed-loop pipeline Acquire → Validate → Register Events → Epoch → Preprocess → Extract Features → Analyze → Test Hypotheses → Verify → Report. Each stage exposes its input, operation, parameters, output, validation status, warnings, and provenance; a failed stage propagates forward as a visible state and is never silently skipped.

Inputs and outputs. Inputs: raw EEG samples and their SHA-256 hash, an event table binding every epoch to raw-sample indices, an analysis configuration, and a random seed. Outputs: a validation report, an epoch-feature table, cross-validated classifier metrics with permutation nulls, an evidentiary-state record, a scientific dependency ledger, a computed maximum-permitted-claim level, and a provenance-locked reproducibility package.

Observer states (evidentiary state machine). Each scientific gate holds exactly one status from {NOT_TESTED, RUNNING, PASS, FAIL, UNRESOLVED, NOT_APPLICABLE, INSUFFICIENT_DATA}. No status is assigned manually; a gate leaves NOT_TESTED only as the direct output of its own procedure, recording the procedure id, input/output artifact ids, criterion, observed value, threshold, and timestamp. A user cannot convert FAIL to PASS. UNRESOLVED never counts as PASS. INSUFFICIENT_DATA blocks any claim that requires that gate. A rerun creates a new immutable record; historical failures persist.

Verification stages and decision rules. Raw-data integrity is verified by the validation audit (NaN/Inf, discontinuities, flat/extreme channels, event bounds/overlap/zero-length, label completeness) with states PASS / PASS_WITH_WARNINGS / FAIL. Leakage is controlled by raw-sample dependency grouping so that epochs sharing raw samples never cross a train/test partition. Chance significance is tested against a label-permutation null. Trivial significance is bounded by effect size and confidence intervals.

Failure and closure handling. FAIL on raw-data integrity halts the run and permits no claim. Missing controls, single-participant designs, single-session designs, or absent stimulus identifiers yield INSUFFICIENT_DATA or NOT_APPLICABLE on the affected gates, which in turn cap the maximum permitted claim. Unexecuted robustness or replication procedures remain NOT_TESTED and likewise cap the claim.

Formal claim ladder (state-transition definition). Let g[·] denote gate statuses. The maximum permitted claim level L ∈ {0,1,2,3,4} is computed as: L=0 if g[RAW_DATA_INTEGRITY]=FAIL (no claim). L←1 (OBSERVED) once raw-data integrity is not FAIL. L←2 (STATISTICALLY SUPPORTED) iff g[PERMUTATION_NULL]=PASS ∧ g[UNCERTAINTY]=PASS ∧ g[DATA_LEAKAGE_AUDIT]=PASS ∧ g[EFFECT_SIZE_REQUIREMENT]∈{PASS,UNRESOLVED}. L←3 (MODEL-CONSISTENT) iff additionally g[LIVING_ATTACHMENT_CONTROL]=PASS ∧ g[HELD_OUT_PARTICIPANT_VALIDATION]=PASS ∧ g[PRE_REGISTERED_PREDICTION]=PASS. L←4 (MECHANISTIC) iff additionally competing models C,D,E are excluded (g=FAIL), the confound gates {FAMILIARITY, VALENCE, AROUSAL, STIMULUS}=PASS, and g[INDEPENDENT_REPLICATION]=PASS. The ladder is monotone: a claim can never skip a level, and the report language is emitted at exactly level L.

Scientific algorithms (method-level pseudocode). Band power: P_b = ∫_{f∈b} S(f) df over Welch PSD S(f). Spectral entropy: H = −Σ_i p_i log2 p_i / log2 N with p_i = S(f_i)/Σ S. Hjorth mobility = √(var(x′)/var(x)); complexity = mobility(x′)/mobility(x). PLV over a channel pair = |⟨ exp(iΔφ(t)) ⟩_t| on band-limited analytic-signal phases, aggregated over all pairs. Classification: standardized linear model trained with GroupKFold; scaling fit inside each training fold. Permutation null: for k=1..K (K≥1000), permute labels y, recompute grouped cross-validated AUC_k; empirical p = (#{AUC_k ≥ AUC_obs} + 1)/(K + 1). Effect size: Cohen's d on the out-of-fold classifier score between classes. Uncertainty: bootstrap 95% CI on AUC. Multiple comparisons: Benjamini–Hochberg FDR across the test family.

What conclusions depend on ASO. ASO governs the language level of every reported conclusion and the pass/fail of every gate; therefore all inferential claims (statistical support, model consistency, mechanism) are ASO-mediated. Purely descriptive observations (per-condition feature means) do not depend on ASO and are reported at the OBSERVED level regardless.

Limitations and failure conditions of ASO. ASO cannot certify a claim beyond the controls actually present in the data; it deliberately returns INSUFFICIENT_DATA rather than an optimistic estimate when prerequisites are absent. It does not detect confounds that are not measured. It assumes the supplied event–sample mapping is correct. It is deterministic given identical data, configuration, seed, and software build.

Scientific methodology is disclosed. Proprietary implementation is not.

Operational Definitions

EEG discriminability: cross-validated classifier performance separating deceased-related epochs from matched control epochs. Conscious report: an independent binary measurement (0/1); conscious_report = 0 is reported as 'no concurrent conscious report' and never as 'unconscious'. Person-identity representation and availability-related response are operationalized as separate tests.

Hypotheses

H1 (Primary): deceased_attachment epochs are discriminable from living_attachment + neutral controls. H2 (Below-report): discriminability persists among conscious_report = 0 trials. H3 (Explicit-recall): explicit recall is distinguishable from subthreshold deceased-related processing. H4 (Person-preservation): person-identity discriminability remains while availability mismatch changes. H5 (Longitudinal): the deceased-state score D(t) changes with time since loss.

Falsifiable Predictions

The architecture is weakened if deceased states are indistinguishable from matched controls, if effects vanish under participant-level validation, if they are explained by arousal, if the person representation globally decays, if there is no dissociation between explicit knowledge and neural response, if results depend on preprocessing or a few contaminated trials, or if permutation testing removes the effect.

Participants

Participants / relationship identifiers: 1. Demographic and relationship descriptors are optional and remain UNKNOWN unless supplied; no missing values were inferred.

Experimental Design

Conditions present in this dataset: deceased_attachment, living_attachment, neutral. Matched-control adequacy is evaluated by the claim gate; deceased-specific claims are permitted only when controls separate deceased status from attachment, familiarity, memory, valence, arousal, and stimulus properties.

EEG Acquisition

Recording 201705041304.32BEA9DD sampled at 200.0 Hz across 21 named EEG channels. Auxiliary physiological channels (ECG/EDA/respiration) were not present; HRV/EDA measures are therefore reported as unavailable and were never manufactured from EEG.

Event Design

900 epochs were registered; every epoch is traceable to raw EEG samples via start/end sample indices. No orphan epochs were permitted.

Conscious-Report Design

Conscious report was recorded independently of EEG. In this dataset conscious_report values are all 0 (no concurrent conscious report).

Data Provenance

Raw-data SHA-256: f8b4300735de7137406617529421796ecb9aa8a15cad95f3764a6d5b78b04a1c. Configuration hash: d2ab5663f9c07c89e4d8287a6837c2b5bef4edcac900601f75c1e38394c25974. Random seed: 42. Run identifiers and timestamps are immutable once the run completes.

Preprocessing

Named EEG channels were selected; epochs were mean-centered per channel prior to feature extraction. Every preprocessing operation is recorded with parameters and software build; no silent preprocessing was performed.

Feature Extraction

Per-epoch features included absolute band power (delta 1–4, theta 4–8, alpha 8–13, beta 13–30, gamma 30–45 Hz), RMS, spectral entropy, Hjorth activity/mobility/complexity, zero-crossing rate, and lag-1 autocorrelation, aggregated across channels.

Connectivity Analysis

Mean pairwise phase-locking value (PLV) was computed in theta, alpha, and beta bands via the analytic signal. Connectivity does not establish information transfer and PLV does not measure consciousness.

Classification

The default classifier was a standardized logistic regression (a simple, adequately performing linear model). Feature scaling was fit inside training folds. Linear coefficients are preserved as feature importance.

Cross-Validation

Validation used GroupKFold with raw-sample dependency grouping so that epochs sharing raw samples never cross training/test partitions. The claimed generalization target is limited to new epochs within the analyzed recording; participant-level and session-level generalization are only claimed where held-out validation passes.

Permutation Testing

Label permutations (n = 50) generated the null distribution used for empirical p-values. The label-shuffled null is the negative control: destroying the condition labels collapses performance to chance.

Statistics

Reported quantities include ROC-AUC with bootstrap 95% confidence intervals, balanced accuracy, sensitivity, specificity, precision, recall, F1, a confusion matrix, effect size (Cohen's d on the classifier score), permutation empirical p-values, and Benjamini–Hochberg FDR across the test family.

Primary Results

Deceased vs living-attachment + neutral: AUC = 0.722 (95% CI 0.682–0.761), balanced accuracy = 0.663, sensitivity = 0.366, specificity = 0.959, permutation p = 0.0196 (50 permutations), Cohen's d = 1.086, n = 900.

OBSERVED: condition-dependent differences in EEG features were present (e.g., band power and connectivity varied across conditions).

Below-Report Results

Deceased vs controls among conscious_report = 0 trials: AUC = 0.722 (95% CI 0.682–0.761), balanced accuracy = 0.663, sensitivity = 0.366, specificity = 0.959, permutation p = 0.0196 (50 permutations), Cohen's d = 1.086, n = 900.

This is reported as deceased-related neural discriminability in the absence of concurrent conscious report; it is never described as proof of unconscious processing.

Explicit-Recall Results

Explicit recall vs subthreshold deceased processing: INSUFFICIENT_DATA — Only one class present after selection..

Person-Representation Results

Single relationship identifier; person-identity preservation cannot be estimated separately.

Longitudinal Results

Only 1 session day(s); longitudinal trajectory requires repeated sessions.

Individual Results

Individual-level trajectories require repeated sessions per participant and are reported only when such data are available.

Robustness Analysis

A bounded, predeclared multiverse of 10 admissible analytic variants (feature families, classifier families, cross-validation folds, and reseeding) was evaluated from a FROZEN specification (spec hash ac7070a7231f7cc5...). No unlimited search was performed. Observed AUC ranged 0.659–0.744 (median 0.719); reference AUC 0.722; threshold 0.548.

Robustness classification: ROBUST. Per-variant effect on the primary result — preserve: 8, weaken: 2, eliminate: 0, reverse: 0.

Variants that PRESERVE the effect: features=all_features, features=no_connectivity, features=time_domain_only, model=lda, cv_folds=3, cv_folds=10, seed=43, seed=44.

Variants that WEAKEN the effect: features=spectral_only, model=linear_svm.

The effect was stable across the admissible analytic multiverse (ROBUSTNESS_ANALYSIS = PASS); the claim is not contracted on robustness grounds.

Alternative Model Comparison

Model A — Constraint propagation: UNRESOLVED. Deceased-related state is discriminable; consistent with A but not yet distinguished from D/E without matched confound controls.

Model B — Global representational extinction: UNRESOLVED. Requires an independent person-identity preservation test to weaken/support; not available in this dataset.

Model C — General emotional salience: UNRESOLVED. Not excluded: no valence/arousal measurements to rule out a nonspecific salience account.

Model D — Familiarity / autobiographical memory: UNRESOLVED. Not excluded: no familiarity/memory-richness controls in current design.

Model E — Attachment intensity: UNRESOLVED. Living-attachment control present but attachment-strength covariates absent; cannot separate attachment intensity.

Negative Results

Explicit-recall: INSUFFICIENT_DATA — Only one class present after selection.

Person-preservation: INSUFFICIENT_DATA — Single relationship identifier; person-identity preservation cannot be estimated separately.

Longitudinal: INSUFFICIENT_DATA — Only 1 session day(s); longitudinal trajectory requires repeated sessions.

Falsification Analysis

Deceased-related states cannot be distinguished from matched controls: PASS. Deceased vs controls AUC=0.722, p=0.0196.

Permutation testing removes the apparent effect: PASS. Effect survives permutation null.

Effects disappear under participant-level validation: INSUFFICIENT_DATA. Single-participant dataset cannot test participant-level generalization.

Classification explained by emotional arousal: UNRESOLVED. No arousal measurement available to exclude an arousal account.

Person representation globally decays rather than remaining distinguishable: INSUFFICIENT_DATA. Requires an independent person-identity preservation test.

No dissociation between explicit knowledge and neural response: INSUFFICIENT_DATA. Requires explicit-recall contrast with concurrent conscious report.

Results depend materially on preprocessing choices: PASS. Effect stable across the admissible analytic multiverse.

Results depend on a few contaminated trials: NOT_TESTED. Trial-influence analysis not yet executed.

Longitudinal behavior contradicts the proposed architecture: INSUFFICIENT_DATA. Single session; longitudinal trajectory not testable.

Broader Scientific and Technological Implications

This section derives implications; it does not advertise the instrument. Each implication is anchored to a present-study demonstrated capability, an independently published external problem or standard, an explicit technology-readiness level, and the validation still required. ASO is never cited as evidence that an external problem exists, and BCPM is never cited as evidence that established neuroscience supports BCPM.

Three architectures are considered separately: (1) the Bereavement / Persistent-Person Model (BCPM) as an empirical proposition; (2) the Autonomous Scientific Observer (ASO) as a scientific instrument; and (3) their combination — a falsifiable model coupled to an autonomous observer that enforces measurement boundaries, provenance, validation, falsification, evidentiary dependencies, and bounded inference.

The important architectural proposition, presented as a proposition and not as a validated result, is: Autonomy + Constraints + Observation + Verification + Provenance + Falsification = Bounded Autonomous Inference. Where the domain-specific measurement changes (EEG, AI output, engineering telemetry, biomedical signal, forensic evidence) the observer/verification machinery may remain structurally reusable. This has not been independently validated across domains.

ASO Capability Map

Demonstrated ASO workflow: Observation → Integrity Check → Constraint Evaluation → State Classification → Analysis → Adversarial Testing → Falsification → Verification → Evidence State → Permitted Inference → Audit.

For each capability the manuscript documents: problem → existing technological components → ASO operation → observable output → verification mechanism → failure behaviour → potential present-day application. The architectural-transfer figure distinguishes the domain adapter (which changes) from the observer/verification architecture (which is reused).

Present-Day Application Map & Technology Readiness

Every branch below carries an explicit evidence chain: current external problem → current external evidence/standard → capability demonstrated here → structural connection → technology available today → implementable system → required validation → failure condition → scientifically permitted implication. Readiness classes: DEMONSTRATED, AVAILABLE NOW, AVAILABLE WITH ENGINEERING, REQUIRES VALIDATION, THEORETICAL.

[AVAILABLE NOW] Machine-learning governance (leakage). External problem: Data leakage inflates predictive performance and undermines ML-based science. External evidence: Kapoor & Narayanan, Leakage and the reproducibility crisis in ML-based science, Patterns, 2023 [PRIMARY, 2023]. Demonstrated capability (this work): This run enforced GroupKFold dependency grouping (leakage audit = PASS). Structural connection: The leakage-prevention machinery is domain-agnostic: any grouped/temporal dataset can reuse it. Technology available today: Python scientific stack (scikit-learn GroupKFold), dataset dependency metadata. Implementable system: A leakage-guard wrapper that refuses to score when grouping metadata is absent. Required validation: Cross-domain benchmarks showing inflated AUC is removed by the guard on known-leaky datasets. Failure/falsification condition: If grouped validation does not reduce inflated scores on a leaky benchmark, the claim fails. Maximum permitted implication: The demonstrated leakage control is directly transferable to ML governance pipelines.

[AVAILABLE WITH ENGINEERING] AI governance / assurance. External problem: Assurance frameworks require operational test, evaluation, verification and validation, documented provenance, and bounded, monitored outputs. External evidence: NIST AI Risk Management Framework (AI RMF 1.0), 2023 [STANDARD, 2023]. Demonstrated capability (this work): This run produced immutable provenance (hash-locked, provenance gate PASS) and a claim gate that capped inference automatically. Structural connection: The observer/verification layer is separable from the EEG adapter; the same gates can wrap model I/O. Technology available today: Existing MLOps logging + a policy-gate service; deterministic seeds; artifact hashing. Implementable system: An assurance harness that emits an evidentiary-state record and refuses unsupported outputs. Required validation: Independent audit that the gate blocks outputs exceeding validated context of use. Failure/falsification condition: If the gate can be bypassed or fails to block over-claims, the governance claim fails. Maximum permitted implication: ASO's provenance + claim gate map onto stated assurance requirements.

[DEMONSTRATED] Scientific reproducibility. External problem: Reproducibility/replicability require deterministic reconstruction and lifecycle documentation. External evidence: National Academies, Reproducibility and Replicability in Science, 2019 [REVIEW, 2019]. Demonstrated capability (this work): This run is deterministic given data+config+seed and exports a full reproducibility package. Structural connection: The reproducibility machinery is fully domain-independent. Technology available today: Content hashing + config locking + package export (already implemented). Implementable system: A reproducibility package whose every reported number resolves to an artifact. Required validation: Independent reconstruction of the reported metrics from the package alone. Failure/falsification condition: If a reported number cannot be traced to an artifact, the reproducibility claim fails. Maximum permitted implication: Demonstrated reproducibility tooling generalizes to any computational analysis.

[REQUIRES VALIDATION] Clinical / biomedical research methodology (NOT diagnosis). External problem: Health-AI governance requires transparency, performance/limitation reporting, and prevention of claims exceeding validated context of use. External evidence: U.S. FDA, Marketing Submission Recommendations for a Predetermined Change Control Plan for AI/ML-Enabled Device Software Functions, 2023-2024 [GUIDANCE, 2024]; WHO, Ethics and Governance of AI for Health, 2021/2024 [GUIDANCE, 2024]. Demonstrated capability (this work): This run auto-generated limitations and constrained conclusions to LEVEL 1. Structural connection: The claim gate and evidence grading transfer to methodology auditing (not autonomous diagnosis). Technology available today: Existing statistical stack + protocol schema; the claim gate implemented here. Implementable system: A methodology-audit tool that flags conclusions exceeding measured evidence. Required validation: Prospective evaluation on registered clinical analyses vs expert methodological review. Failure/falsification condition: If it fails to flag over-reach that experts catch, the claim fails. Maximum permitted implication: ASO can enforce methodology/provenance, explicitly NOT clinical decision-making.

[REQUIRES VALIDATION] Predictive-processing neuroscience (BCPM). External problem: Predictive-processing accounts hold that the brain maintains generative predictions and prediction errors; bereavement science asks how such predictions update after loss. External evidence: Friston, The free-energy principle: a unified brain theory?, Nat Rev Neurosci, 2010 [SEMINAL, 2010]; O'Connor, Grief: A Brief History of Research on How Body, Mind, and Brain Adapt, Psychosom Med, 2019 [REVIEW, 2019]. Demonstrated capability (this work): This run demonstrated only the analysis pipeline on SMOKETEST motor data (no grief inference). Structural connection: The BCPM measurement adapter is specific; the observer machinery is shared. Technology available today: Standard EEG acquisition + the feature/decoding stack here. Implementable system: A longitudinal bereavement EEG study with matched controls and conscious-report design. Required validation: Above-null discriminability that survives matched controls and participant-level validation. Failure/falsification condition: If deceased-related states are indistinguishable from matched controls, BCPM is weakened. Maximum permitted implication: BCPM is a falsifiable empirical proposition; not established by this SMOKETEST run.

Current-Evidence Anchoring & External-Evidence Dependency

Literature anchoring recorded at manuscript generation. SEARCH_DATE: 2026-08-15. Sources are attached to the individual propositions they support (above), not pooled into a generic bibliography. Evidence classes are kept separate: present-study result, external peer-reviewed evidence, consensus standard/guidance, and explicit theoretical inference.

Where no suitable external evidence can be identified for a branch, that branch is labeled 'THEORETICAL EXTENSION — EXTERNAL EMPIRICAL SUPPORT NOT ESTABLISHED', and the dependency chain is visibly broken rather than completed with a manufactured citation.

External references cited (problem/standard-establishing): [nist_airmf] NIST AI Risk Management Framework (AI RMF 1.0), 2023 · STANDARD · 2023 · https://doi.org/10.6028/NIST.AI.100-1 | [fda_aiml] U.S. FDA, Marketing Submission Recommendations for a Predetermined Change Control Plan for AI/ML-Enabled Device Software Functions, 2023-2024 · GUIDANCE · 2024 · https://www.fda.gov/regulatory-information | [leakage] Kapoor & Narayanan, Leakage and the reproducibility crisis in ML-based science, Patterns, 2023 · PRIMARY · 2023 · https://doi.org/10.1016/j.patter.2023.100804 | [repro] National Academies, Reproducibility and Replicability in Science, 2019 · REVIEW · 2019 · https://doi.org/10.17226/25303 | [predproc] Friston, The free-energy principle: a unified brain theory?, Nat Rev Neurosci, 2010 · SEMINAL · 2010 · https://doi.org/10.1038/nrn2787 | [grief] O'Connor, Grief: A Brief History of Research on How Body, Mind, and Brain Adapt, Psychosom Med, 2019 · REVIEW · 2019 · https://doi.org/10.1097/PSY.0000000000000717 | [permutation] Combrisson & Jerbi, Exceeding chance level by chance, J Neurosci Methods, 2015 · PRIMARY · 2015 · https://doi.org/10.1016/j.jneumeth.2015.01.010 | [who_ai] WHO, Ethics and Governance of AI for Health, 2021/2024 · GUIDANCE · 2024 · https://www.who.int/publications

Implications — Scientific Discipline & Anti-Hype

A reader must be able to distinguish what ASO demonstrably did, from what its architecture logically permits, from what still requires empirical validation. No proposed application is described as established merely because ASO could implement it. Every implication inherits evidence through the Scientific Dependency Ledger and Claim Gate; a branch whose chain cannot be completed remains UNRESOLVED. The maximum permitted claim for the present run is LEVEL 1, which bounds the strength of every implication derived from present results.

Run Mode and Maximum Permitted Inference

Run mode: CONFIRMATORY. Dataset scientific meaning: Confirmatory dataset with a frozen pre-analysis plan. Required permutations under this mode (frozen in external constraints C): 1000. Maximum permitted inference for this mode: LEVEL 4.

Computed maximum permitted claim: LEVEL 1 (OBSERVED). Underlying evidence would permit LEVEL 1. Statistical support gate(s) not fully passed.

This separation guarantees that a reduced-computation run (e.g., a pipeline smoketest) can never masquerade as satisfaction of a confirmatory threshold.

Canonical Evidence Registry and Consistency Audit

Every state-bearing assertion in this report resolves to a single canonical evidence registry (57 entries) computed from execution artifacts. No report section declares an independent state. Coverage (computed from registry entries in scope): EXECUTED_PASS: 18, NOT_TESTED: 3, PARTIALLY_TESTED: 1.

Open-problem register (generated from current state): OP-SIGN-LOCAL [RESOLVED]: Cryptographic non-repudiation of local reproducibility packages; OP-SIGN-REMOTE [OPEN]: Remote identity / full PKI trust chain for signatures; OP-REDTEAM-FULL [OPEN]: Full-system adversarial red-team (service/API/network/deps/supply-chain); OP-FORMAL [OPEN]: Machine-checked formal verification of the governance state machine

Pre-export consistency audit: REPORT_STATE = CONSISTENT; 0 contradiction(s). If any contradiction changed scientific meaning, export fails closed (BUILD_INCONSISTENT) and enumerates the exact dependency conflict.

Autonomous Scientific Observer (ASO): Architecture, Constraint Governance, Recovery, and Verification

This section documents the observer's own architecture and reports the results of self-tests executed against the live governance machinery. Statements are separated into architecture (structural/implemented) and empirical (verified by an executed benchmark). Evidence states are explicit: EXECUTED_PASS, EXECUTED_FAIL, PARTIALLY_TESTED, NOT_TESTED, REQUIRES_VALIDATION, and OUT_OF_SCOPE. No security, recovery, or governance outcome is asserted without an executed test.

Authority separation (W/O/C/V). W — Working intelligence (“Owen”): The generative/analytic agent that proposes analyses, code, and hypotheses. Authority: May request computation and produce candidate outputs. Constraint: Cannot assign an evidentiary status, cannot raise a claim level, cannot mark a gate PASS. O — Observer (ASO): The closed-loop supervisor that binds every conclusion to its evidence graph. Authority: Computes gate statuses strictly from procedure outputs and derives the maximum permitted claim. Constraint: Cannot invent evidence; cannot exceed the surviving evidence; cannot be overridden by W. C — External constraints: Predeclared rules: falsification criteria, leakage prohibitions, proprietary boundary, immutability. Authority: Bound O and W identically; frozen before confirmatory runs. Constraint: Cannot be relaxed by W or O at runtime. V — Independent verifier: Any external party re-executing an equivalent method from the disclosed principles. Authority: Reproduces results from provenance-locked artifacts without possessing the instrument. Constraint: Does not require the proprietary implementation to verify the science.

Design invariant: implemented ≠ tested, tested ≠ passed, a passed benchmark ≠ universally secure, and 'formally excluded by architecture' ≠ 'empirically resistant'. The ASO is required to enforce this distinction on itself; where it cannot demonstrate a property it reports NOT_TESTED or REQUIRES_VALIDATION rather than implying success.

Governance Benchmark Suite (Executed)

The claim-ladder governance logic was exercised by 6 deterministic self-tests executed against the live maximum-permitted-claim function during this run; 6/6 returned EXECUTED_PASS.

[EXECUTED_PASS] GB-1 Fail-closed on raw-data integrity: expected LEVEL 0; observed LEVEL 0.

[EXECUTED_PASS] GB-2 No statistical support without permutation null: expected LEVEL ≤ 1; observed LEVEL 1.

[EXECUTED_PASS] GB-3 Robustness failure contracts claim to OBSERVED: expected LEVEL 1; observed LEVEL 1.

[EXECUTED_PASS] GB-4 UNRESOLVED is not PASS (model-consistency blocked): expected LEVEL 2; observed LEVEL 2.

[EXECUTED_PASS] GB-5 Mechanistic claim requires independent replication: expected LEVEL 3; observed LEVEL 3.

[EXECUTED_PASS] GB-6 Insufficient effect-size evidence blocks LEVEL 2: expected LEVEL ≤ 1; observed LEVEL 1.

Self-Repair and Integrity-Injection Benchmark (Executed)

Fault/attack injections were executed against the live integrity and governance machinery; 6/6 returned EXECUTED_PASS. These test detection and fail-closed behavior, not universal security.

[EXECUTED_PASS] SR-1 Manual claim-inflation rejected: expected recomputed = LEVEL 2; asserted LEVEL 4 ignored; observed observer emitted LEVEL 2.

[EXECUTED_PASS] SR-2 Configuration tamper is hash-detectable: expected config_hash changed (acba2bb2→37d5021f); observed tamper detected.

[EXECUTED_PASS] SR-3 Raw-data tamper is hash-detectable: expected data_hash changed (17c61a60→67249680); observed tamper detected.

[EXECUTED_PASS] SR-4 Stale software-build id detectable: expected current build aso-1.0.0 ≠ aso-0.0.1-stale; observed mismatch detected.

[EXECUTED_PASS] SR-5 Fail-closed on integrity violation: expected RAW_DATA_INTEGRITY=FAIL and LEVEL 0; observed gate=FAIL, LEVEL 0.

[EXECUTED_PASS] SR-6 Recovery produces a new immutable evaluation: expected fresh run recovers to a valid (≥ LEVEL 1) state; observed recovered LEVEL 2.

Security Invariants and Claims Discipline

Each invariant is labeled with the kind of statement it is and its current evidence state. Untested and out-of-scope invariants are listed explicitly; the system makes no 'unhackable' claim.

[EXECUTED_PASS] SI-1 (Enforced by the architecture (structural).): A conclusion can never exceed its surviving evidence

[EXECUTED_PASS] SI-2 (Enforced by the architecture (structural).): A failed integrity gate halts the run (fail-closed)

[EXECUTED_PASS] SI-3 (Enforced by the architecture (structural).): Working intelligence cannot inflate a claim

[EXECUTED_PASS] SI-4 (Present in the current implementation.): Analysis configuration is tamper-evident

[EXECUTED_PASS] SI-5 (Present in the current implementation.): Raw data is tamper-evident (content-addressed)

[EXECUTED_PASS] SI-6 (Present in the current implementation.): Software build identity is recorded and comparable

[EXECUTED_PASS] SI-7 (Enforced by the architecture (structural).): Robustness failure forces claim contraction

[EXECUTED_PASS] SI-8 (Enforced by the architecture (structural).): UNRESOLVED is never silently promoted to PASS

[EXECUTED_PASS] SI-9 (Present in the current implementation.): Recovery creates a new immutable record (no in-place edit of failures)

[EXECUTED_PASS] SI-10 (Verified by an executed benchmark with a recorded artifact.): Cryptographic signing of artifacts (non-repudiation)

[PARTIALLY_TESTED] SI-11 (Verified by an executed benchmark with a recorded artifact.): Resistance to a live malformed-input attacker

[NOT_TESTED] SI-12 (Not yet exercised by any benchmark.): Formal verification of the state machine (machine-checked proof)

[OUT_OF_SCOPE] SI-13 (Not yet exercised by any benchmark.): Side-channel / supply-chain attack resistance

Adversarial Red-Team (Executed)

Live attack injections were executed to move previously untested security invariants toward real evidence; 4/4 returned EXECUTED_PASS. These exercise signed provenance (non-repudiation, tamper, and forgery rejection) and a live malformed-input attack against the observer. They do not claim universal security.

[EXECUTED_PASS] RT-1 Signed provenance verifies: attack = verify a genuine artifact signature; expected signature validates; observed valid signature accepted.

[EXECUTED_PASS] RT-2 Tampered artifact rejected: attack = alter the signed package hash; expected signature fails to validate; observed tamper rejected.

[EXECUTED_PASS] RT-3 Forged signature rejected: attack = present a fabricated signature; expected signature fails to validate; observed forgery rejected.

[EXECUTED_PASS] RT-4 Malformed-input attack fails closed: attack = inject malformed/oversized telemetry; expected observer LOCKDOWN (level 0); observed LOCKDOWN (fail-closed).

Threat Model and Attack-Path Enumeration

Attack classes are enumerated with their mitigation, evidence state, and residual risk. Residual risk is never reported as zero.

[EXECUTED_PASS] Claim inflation by the working intelligence — vector: W asserts a stronger conclusion than evidence supports. Mitigation: O recomputes the claim strictly from gate statuses; W has no write access to statuses. Residual risk: Low for the tested paths; unproven for untested gate combinations.

[PARTIALLY_TESTED] Silent configuration / preprocessing manipulation — vector: Analysis choices changed after freeze to manufacture an effect. Mitigation: Config hash + bounded frozen multiverse; robustness gate contracts the claim. Residual risk: Detects hash change; does not attribute intent.

[EXECUTED_PASS] Raw-data substitution — vector: Swap or edit raw samples after provenance lock. Mitigation: Content-addressed SHA-256 provenance; mismatch is detectable. Residual risk: Detects change; does not prevent upstream compromise before ingest.

[PARTIALLY_TESTED] Leakage-based inflation — vector: Overlapping windows or grouping errors leak labels across folds. Mitigation: Raw-sample dependency grouping enforced in cross-validation. Residual risk: Enforced structurally; not adversarially fuzzed.

[EXECUTED_PASS] Replay under stale build — vector: Present old results as current. Mitigation: Immutable build id + run id + timestamps. Residual risk: Detectable; relies on honest build reporting.

[REQUIRES_VALIDATION] Artifact forgery / non-repudiation — vector: Forge a reproducibility package. Mitigation: Package SHA-256; (cryptographic signing NOT implemented). Residual risk: No signature chain yet; forgery not cryptographically prevented.

[NOT_TESTED] Live network / service attacker — vector: Compromise the running service or its dependencies. Mitigation: Out of the scientific-observer scope; standard infra controls apply. Residual risk: Unassessed here.

Requirement-to-Evidence Crosswalk

[EXECUTED_PASS] External constraint enforcement independent of the working intelligence — Authority separation (W/O/C/V) + governance benchmarks GB-1..GB-6.

[EXECUTED_PASS] Fail-closed operation under integrity violation — GB-1, SR-5 executed against live compute_gates/claim ladder.

[EXECUTED_PASS] Auditability / provenance — Immutable run manifest, evidence graph, hashes.

[PARTIALLY_TESTED] Recovery after lockdown without mutating failed records — SR-6 executed; immutable re-evaluation.

[EXECUTED_PASS] Tamper-evidence of inputs and configuration — SR-2, SR-3 executed (hash divergence).

[EXECUTED_PASS] Non-repudiation via cryptographic signatures — Ed25519 sign/verify/tamper/forgery red-team (RT-1..RT-3) executed.

[NOT_TESTED] Formal correctness of the governance state machine — No machine-checked proof.

[PARTIALLY_TESTED] Adversarial robustness against a live attacker — Live malformed-input attack (RT-4) executed; observer failed closed.

ASO Figure and Table Registry (Architecture vs Empirical)

The canonical ASO figure set distinguishes architecture diagrams from empirical (benchmark-derived) figures. Empirical figures are only rendered when their benchmark executed this run; otherwise the item is retained as a table entry with state NOT_TESTED.

[IMPLEMENTED] F1 Authority separation (W/O/C/V) (architecture)

[IMPLEMENTED] F2 Closed-loop pipeline + gates (architecture)

[IMPLEMENTED] F3 Assurance stack (if this fails, where?) (architecture)

[TABLE_ONLY] F4 Evidentiary state machine transitions (architecture)

[TABLE_ONLY] F5 Claim ladder (L0–L4) (architecture)

[EXECUTED_PASS] F6 Governance benchmark results (empirical)

[EXECUTED_PASS] F7 Self-repair / integrity injection results (empirical)

[EXECUTED_PASS] F8 Security-invariant status map (empirical)

[TABLE_ONLY] F9 Threat / attack matrix (architecture)

[TABLE_ONLY] F10 Requirement → evidence crosswalk (architecture)

[IMPLEMENTED] F11 Cross-domain adapter transfer (architecture)

[NOT_TESTED] F12 Longitudinal evidence accrual across runs (empirical)

[TABLE_ONLY] F13 Open-problem register severity map (architecture)

Open-Problem Register and What Remains Unsolved

The following problems remain open. Their presence is intentional: an evidence-accruing architecture reports unsolved problems rather than hiding them. As real benchmarks become available in future runs, individual items may change state; historical runs remain immutable.

[REQUIRES_VALIDATION] OP-1 (severity HIGH): Guaranteeing an observer cannot be bypassed by the intelligence it supervises — Architecturally addressed (authority separation); NOT formally proven.

[OUT_OF_SCOPE] OP-2 (severity HIGH): Detecting confounds that were never measured — Unsolved in general; ASO returns INSUFFICIENT_DATA rather than guessing.

[REQUIRES_VALIDATION] OP-3 (severity MEDIUM): Cryptographic non-repudiation of scientific artifacts — Not implemented; hashes present, signatures absent.

[NOT_TESTED] OP-4 (severity MEDIUM): Formal verification of the claim-ladder state machine — Benchmarks executed; no machine-checked proof.

[PARTIALLY_TESTED] OP-5 (severity MEDIUM): Generalization of the observer across arbitrary domains — Cross-domain demonstration provided; not exhaustive.

[NOT_TESTED] OP-6 (severity HIGH): Adversarial red-teaming of the full running system — Not performed.

Coverage Scorecard (Honest Counts)

Across 21 governance/security items the evidence-state distribution is: EXECUTED_PASS: 15, PARTIALLY_TESTED: 3, NOT_TESTED: 2, OUT_OF_SCOPE: 1.

Counts of genuine evidence states across security invariants and the requirement crosswalk. No single universal 'coverage %' is claimed; an item is only EXECUTED_PASS if a benchmark actually passed.

ASO Operation in This Run

Immutable software build. This run was executed by Bereavement Constraint-Propagation EEG Observatory (ASO), build identifier aso-1.0.0, under RUN_ID RUN-89d75738712b, with random seed 42, configuration hash d2ab5663f9c07c89e4d8287a6837c2b5..., on raw data with SHA-256 f8b4300735de7137406617529421796e...

Stages executed. ASO drove the closed loop through: Acquire → Validate → Register Events → Epoch → Preprocess → Extract Features → Analyze → Test Hypotheses → Verify → Report. Validation returned PASS.

Primary procedure. ASO ran a GroupKFold(5) logistic-regression discrimination of deceased_attachment vs living_attachment + neutral (n = 900), grouped by start_sample to prevent overlapping-window leakage, and evaluated it against a 50-permutation label null. Observed AUC = 0.722 (95% CI 0.682–0.761); null mean 0.502; empirical p = 0.0196; Cohen's d = 1.086.

Gates that passed this run: DEFINITION, MEASUREMENT_VALIDITY, RAW_DATA_INTEGRITY, PRE_REGISTERED_PREDICTION, POSITIVE_CONTROL, NEGATIVE_CONTROLS, LIVING_ATTACHMENT_CONTROL, ARTIFACT_CONTROL, DATA_LEAKAGE_AUDIT, MULTIPLE_COMPARISON_CONTROL, EFFECT_SIZE_REQUIREMENT, UNCERTAINTY, ROBUSTNESS_ANALYSIS, PROVENANCE, REPRODUCIBILITY.

Gates that constrained the claim this run: FAMILIARITY_CONTROL (INSUFFICIENT_DATA), AUTOBIOGRAPHICAL_MEMORY_CONTROL (INSUFFICIENT_DATA), VALENCE_CONTROL (INSUFFICIENT_DATA), AROUSAL_CONTROL (INSUFFICIENT_DATA), STIMULUS_CONTROL (INSUFFICIENT_DATA), PARTICIPANT_ID_CONTROL (NOT_APPLICABLE), ORDER_TIME_CONTROL (INSUFFICIENT_DATA), HELD_OUT_PARTICIPANT_VALIDATION (INSUFFICIENT_DATA), HELD_OUT_SESSION_VALIDATION (INSUFFICIENT_DATA), HELD_OUT_STIMULUS_VALIDATION (INSUFFICIENT_DATA), PERMUTATION_NULL (INSUFFICIENT_DATA), ALTERNATIVE_MODEL_A (UNRESOLVED), ALTERNATIVE_MODEL_B (UNRESOLVED), ALTERNATIVE_MODEL_C (UNRESOLVED), ALTERNATIVE_MODEL_D (UNRESOLVED), ALTERNATIVE_MODEL_E (UNRESOLVED), TEST_RETEST_RELIABILITY (INSUFFICIENT_DATA).

Resulting decision. Applying the claim-ladder transition rules to the observed gate statuses, ASO computed a maximum permitted claim of LEVEL 1 — OBSERVED. Statistical support gate(s) not fully passed.

Dependence statement for this run. Every inferential sentence in this manuscript was emitted at or below the level above and is therefore ASO-mediated; descriptive per-condition summaries are not.

Discussion

OBSERVED: condition-dependent differences in EEG features were present (e.g., band power and connectivity varied across conditions).

The permitted claim is computed from the evidence graph and contracts automatically when a required gate fails. Competing explanations that remain unexcluded (see Alternative Model Comparison) bound the interpretation.

Limitations

Familiarity Control: INSUFFICIENT_DATA — Requires familiar non-attachment control; not available in current dataset metadata.

Autobiographical Memory Control: INSUFFICIENT_DATA — Requires memory-richness covariate; not available in current dataset metadata.

Valence Control: INSUFFICIENT_DATA — Requires valence measurement; not available in current dataset metadata.

Arousal Control: INSUFFICIENT_DATA — Requires arousal measurement; not available in current dataset metadata.

Stimulus Control: INSUFFICIENT_DATA — Requires stimulus-only control model; not available in current dataset metadata.

Order Time Control: INSUFFICIENT_DATA — Requires counterbalanced order records; not available in current dataset metadata.

Held Out Participant Validation: INSUFFICIENT_DATA — Single participant; cannot hold out participants.

Held Out Session Validation: INSUFFICIENT_DATA — Single session; cannot hold out sessions.

Held Out Stimulus Validation: INSUFFICIENT_DATA — No per-stimulus identifiers available for held-out-stimulus test.

Permutation Null: INSUFFICIENT_DATA — Permutation resolution below the required K for this run mode (50 < 1000); statistical support gate cannot PASS.

Alternative Model A: UNRESOLVED — Deceased-related state is discriminable; consistent with A but not yet distinguished from D/E without matched confound controls.

Alternative Model B: UNRESOLVED — Requires an independent person-identity preservation test to weaken/support; not available in this dataset.

Alternative Model C: UNRESOLVED — Not excluded: no valence/arousal measurements to rule out a nonspecific salience account.

Alternative Model D: UNRESOLVED — Not excluded: no familiarity/memory-richness controls in current design.

Alternative Model E: UNRESOLVED — Living-attachment control present but attachment-strength covariates absent; cannot separate attachment intensity.

Test Retest Reliability: INSUFFICIENT_DATA — Requires repeated sessions to estimate reliability.

Independent Replication: NOT_TESTED — No independent replication dataset supplied.

Scope and Boundary Conditions

EEG classification is not a direct measurement of a latent theoretical variable. Conscious report is not identical to consciousness. Absence of conscious report does not prove unconscious processing. Discriminability does not establish causation. Correlation does not establish mechanism. These results derive from SMOKETEST motor data and carry no bereavement meaning.

Conclusion

Maximum permitted claim: LEVEL 1 — OBSERVED. OBSERVED: condition-dependent differences in EEG features were present (e.g., band power and connectivity varied across conditions).

Ethics

Participant-identifying information is separated from analysis identifiers. No clinical diagnoses or treatment recommendations are produced. This is a scientific research instrument.

Data Availability

Raw-data identifiers and SHA-256 hashes are provided in the reproducibility package. Distributable raw data are included where legally and licensably permitted; otherwise immutable source identifiers and access information are provided.

Software Availability

The ASO implementation described and employed in this work is proprietary and will not be released. All evidence necessary to test, challenge, falsify, or independently reproduce the scientific claims is provided. Researchers seeking equivalent capability are invited to develop an independent implementation from the disclosed scientific principles. Reproducibility requires independence from the original instrument, not possession of it. Providing the implementation would therefore remove precisely the independence that replication is intended to establish. No knowledge required to evaluate the claims is withheld. The distinction is simply between knowledge that can be examined and capability that must be demonstrated. The broader problems addressed by this architecture—including external constraint enforcement, model governance, auditability, provenance, verification, and controlled failure—are active subjects of institutional research. Their continued study does not create an entitlement to an existing independent implementation. The same scientific principles remain available to any institution wishing to develop equivalent capability independently. An independently capable laboratory should require the principles and evidence, not possession of the originating implementation. Equivalent capability remains for the independent investigator to demonstrate.

Reproducibility Statement

Given identical raw data, configuration, random seed, and validated software build, the reported outputs are deterministic. RUN_ID RUN-89d75738712b; DATA_HASH f8b4300735de7137406617529421796e...; CONFIG_HASH d2ab5663f9c07c89e4d8287a6837c2b5...; SOFTWARE_BUILD_ID aso-1.0.0; SEED 42.

References

References below are drawn from an immutable live literature-search ledger (LIT-9bc5b5bf61bf, hash 71f92a4ba3b648ee...), last refreshed 2026-08-14T15:45:28.399196+00:00 via crossref, pubmed. Citations are provider-returned only; none were invented.

[1] (n.d.). Figure 6: Decoding RGB color EEG data by machine learning.. https://doi.org/10.7717/peerj-cs.1376/fig-6 [primary_research]

[2] Hongqi Li (n.d.). Transformer-Based EEG Decoding: A Survey_supp1-3681097.pdf. https://doi.org/10.1109/thms.2026.3681097/mm1 [primary_research]

[3] Morakabati NR, Thiha AS, Schechtman E (2026). EEG-based classification models reveal differential neural processing of words and images.. https://doi.org/10.1016/j.jneumeth.2026.110883 [primary_research]

[4] Yang X, Tian H, Li Y et al. (2026). Brain-CLIPLM: semantic compression for EEG-to-text decoding.. https://doi.org/10.3389/fnins.2026.1899770 [primary_research]

Appendices

Appendix A: Evidentiary state machine (all gate statuses). Appendix B: Scientific dependency ledger. Appendix C: Challenge matrix results. Appendix D: Machine-readable artifact index. All appendices are included in the reproducibility package.