13 - Physical EEG Setup and Safety¶

Safety-Focused Guide for Non-Medical EEG Recording¶

SCIENTIFIC BOUNDARY: EEG records scalp electrical activity. It does NOT read thoughts, decode mental content, capture consciousness, or transfer subjective experience. EEG-based personalization does NOT make any AI system conscious. This guide covers physical safety for non-medical EEG research only. EEG cannot be used to infer private thoughts, emotions, identity, or consciousness.

1. Scientific Boundary¶

Electroencephalography (EEG) is a non-invasive neuroimaging technique that measures electrical potential differences on the scalp produced by postsynaptic neuronal activity in the cortex. EEG has been a cornerstone of neuroscience research for nearly a century and provides excellent temporal resolution (millisecond-level), but its spatial resolution is limited. The signal recorded at any scalp electrode reflects the summed activity of millions of neurons, filtered through the skull, scalp, and meninges. Understanding the precise boundaries of what EEG can and cannot do is essential before any recording session.

1.1 EEG Records Scalp Electrical Activity¶

EEG electrodes placed on the scalp detect voltage fluctuations that arise primarily from pyramidal neurons in the cerebral cortex. These fluctuations are on the order of microvolts (µV) and are categorized into frequency bands that correlate with different brain states: delta (0.5–4 Hz, deep sleep), theta (4–8 Hz, drowsiness and meditation), alpha (8–13 Hz, relaxed wakefulness), beta (13–30 Hz, active concentration), and gamma (30–45 Hz, cognitive processing). The relationship between these oscillatory patterns and cognitive states is probabilistic, not deterministic. Two individuals may exhibit different EEG signatures during the same task, and the same individual may show variability across sessions. The feature extraction pipeline in this repository (cia.neuro.eeg_feature_extraction) computes these bandpowers and spectral ratios as statistical summaries, not as direct measures of mental content.

1.2 EEG Does NOT Read Thoughts¶

Despite popular portrayals in fiction and media, EEG does not provide direct access to a person's thoughts, intentions, or internal monologue. The EEG signal is a highly aggregated, spatially blurred representation of cortical electrical activity. While machine-learning classifiers have demonstrated the ability to decode certain gross cognitive states (e.g., motor imagery, visual stimulus categories, or sleep stages) from EEG patterns with above-chance accuracy, these decodings are statistical associations, not literal readings of thought content. The classification is inherently probabilistic, limited to predefined categories, and subject to significant inter-individual variability. The neural-state proxy estimates produced by the NeuralStateEncoder in this repository (cia.neuro.neural_state_encoder) are heuristic approximations derived from bandpower features — they do not decode or read thoughts in any meaningful sense.

1.3 EEG Does NOT Capture Consciousness¶

Consciousness is a philosophical and scientific concept that refers to subjective, phenomenal experience — the "what it is like" quality of awareness. EEG measures objective, third-person electrical signals on the scalp. No EEG measure, feature, or pattern can directly demonstrate or capture the presence, absence, or quality of consciousness. The theoretical frameworks implemented in this repository — including Global Workspace Theory, Attention Schema Theory, and Predictive Processing — identify structural and functional correlates that may be relevant to consciousness in biological systems, but these correlates are not consciousness itself. The neuroadaptive extension uses EEG-derived proxy estimates (attention, workload, fatigue, arousal) to modulate the CIA system's cognitive architecture, but this modulation does not transfer, replicate, or create consciousness.

1.4 EEG Cannot Transfer Identity into AI¶

EEG data is a biometric signal — it contains statistical patterns that may be used to identify individuals, similar to fingerprints or iris scans. However, EEG data does not encode identity, personality, memories, subjective experience, or selfhood in a form that can be transferred into an artificial system. The SubjectNeuroProfile in this repository stores baseline EEG features (aggregate bandpowers, spectral ratio references, calibration offsets) for individualized proxy estimation. This profile is a lightweight statistical calibration container — it does not capture the person's identity, consciousness, or subjective experience. Claims of "mind uploading," "consciousness transfer," or "creating a digital copy of the self" based on EEG data are scientifically unfounded and ethically dangerous. This repository explicitly prohibits such claims in its safety policy (cia.neuro.safety) and in the caveat strings attached to every output schema.

2. Ethical and Legal Requirements¶

EEG recording involves the collection of sensitive biometric data from human participants. This imposes significant ethical and legal obligations on researchers and operators. The requirements below apply to any use of the neuroadaptive extension in this repository with real human EEG data. Even when using only synthetic or mock data (such as the MockSignalStream), understanding these requirements is essential for responsible research design.

Informed consent is the foundational ethical requirement for all human subjects research. Before any EEG recording, participants must receive a clear, comprehensible explanation of the study's purpose, procedures, risks, benefits, data handling practices, and their right to withdraw at any time without penalty. The consent form should specifically address: what EEG measures (scalp electrical activity, not thoughts), how the data will be used (e.g., cognitive-state proxy estimation for neuroadaptive AI research), who will have access to the data, how long the data will be stored, and any potential risks (e.g., skin irritation from electrodes, fatigue from extended sessions). Consent must be voluntary and given without coercion. For participants who cannot provide consent themselves (e.g., minors, individuals with cognitive impairments), consent must be obtained from a legally authorized representative, and the participant's assent should be obtained when possible.

2.2 Right to Stop¶

Every participant has the absolute right to stop participating in an EEG recording session at any time, for any reason, without penalty or negative consequences. This right must be clearly communicated before the session begins and reaffirmed throughout. The operator should establish a clear signal (verbal or physical) that the participant can use to request an immediate stop. The safety policy in this repository (NeuroSafetyPolicy.evaluate_recording_plan) enforces this by flagging sessions where consent has not been confirmed. If a participant shows signs of distress, discomfort, or simply wishes to stop, recording must cease immediately, and all electrodes must be removed. Data collected up to the point of withdrawal should be handled according to the participant's stated preference (retained, deleted, or anonymized).

2.3 Data Privacy: EEG Is Biometric Data¶

EEG recordings constitute biometric data under many data protection regulations, including the European Union's General Data Protection Regulation (GDPR) and various national and state-level privacy laws. Biometric data is classified as a special category of personal data that requires enhanced protections. EEG data, when combined with sufficient contextual information, can potentially be used to identify individuals or make inferences about their health, cognitive state, or neurological conditions. Researchers must implement appropriate technical and organizational measures to protect EEG data, including encryption at rest and in transit, access controls, anonymization or pseudonymization, and secure deletion when data is no longer needed. The SubjectNeuroProfile in this repository uses opaque subject identifiers (e.g., participant codes) rather than names, but researchers must ensure that the mapping between codes and real identities is stored separately and securely.

2.4 Local Regulations¶

Data protection and research ethics regulations vary significantly across jurisdictions. Researchers must comply with all applicable local, national, and institutional regulations governing human subjects research, data protection, and medical device usage. Key regulatory frameworks to consider include: GDPR (EU), HIPAA (United States), the Common Rule (United States federal research regulations), the Australian National Statement on Ethical Conduct in Human Research, and equivalent frameworks in other jurisdictions. If the EEG recording is conducted in a clinical context or for clinical purposes, additional medical device regulations (e.g., FDA 510(k) clearance, CE marking under the EU Medical Device Regulation) may apply. This repository's software is explicitly not a medical device and is not intended for clinical diagnosis, treatment, or monitoring.

2.5 Institutional Review Board / Ethics Review¶

All research involving human EEG recording should be reviewed and approved by an appropriate institutional review board (IRB), ethics committee, or equivalent body before data collection begins. The ethics review should evaluate the study's risk-benefit profile, consent procedures, data protection measures, participant recruitment practices, and compliance with applicable regulations. The NeuroSafetyPolicy in this repository provides automated safety checks (device type verification, duration limits, consent confirmation) that complement but do not replace institutional ethics review. Researchers should submit the study protocol, consent forms, data management plan, and risk assessment to their ethics committee for review. The review process may require modifications to the study design, consent process, or data handling procedures.

2.6 Avoid Vulnerable Participants¶

Extra caution must be exercised when recruiting participants who may be vulnerable to coercion, undue influence, or harm. Vulnerable populations include minors, individuals with cognitive impairments, psychiatric patients, prisoners, economically disadvantaged individuals, and individuals who may not fully understand the nature of the research. Special protections for vulnerable participants may include: additional consent from legal representatives, simplified consent forms with clear language, enhanced monitoring during recording sessions, stricter data protection measures, and explicit justification for the inclusion of vulnerable participants. As a general principle for non-medical EEG research with the CIA framework, it is recommended to recruit only healthy adult volunteers who can provide fully informed consent. If vulnerable populations are to be included, this must be explicitly justified in the ethics application and additional safeguards must be implemented.

3. Hardware Safety¶

EEG hardware safety is paramount. EEG recording involves placing electrodes on the scalp and connecting them to electronic amplification equipment. While non-invasive EEG is generally safe when used correctly, improper use can lead to skin irritation, discomfort, or in rare cases, electrical hazards. The following guidelines are designed to ensure safe hardware operation.

3.1 Non-Invasive EEG Only¶

This repository and its neuroadaptive extension support only non-invasive scalp EEG recording. Non-invasive EEG uses electrodes placed on the surface of the scalp (held in place by caps, headbands, adhesive paste, or gel) to measure electrical potentials from the underlying cortex. The NeuroSafetyPolicy.check_noninvasive_only() method in cia.neuro.safety actively rejects any recording plan that indicates the use of invasive devices such as intracranial electrodes, subdural grids, stereotactic EEG (sEEG) depth electrodes, or electrocorticography (ECoG) strips. Invasive EEG procedures are surgical procedures that carry significant medical risks (infection, hemorrhage, neurological deficit) and must only be performed in clinical settings by qualified neurosurgeons. This software must never be used to process data from invasive recording methods.

3.2 Manufacturer-Approved Electrodes¶

Only electrodes approved by the EEG device manufacturer should be used. Manufacturer-approved electrodes have been tested for electrical safety, biocompatibility (skin contact materials), and signal quality. Using third-party or homemade electrodes may introduce electrical hazards (e.g., inadequate insulation, uncontrolled impedance), skin irritation (allergenic materials), or signal quality degradation (poor contact, high noise). Common electrode types include: Ag/AgCl (silver/silver chloride) cup electrodes with conductive gel (research standard), dry electrodes with spring-loaded pins (consumer devices), and fabric-based textile electrodes (wearable research). Each type has specific preparation and maintenance requirements documented in the device manual.

3.3 No Hardware Modification¶

EEG recording equipment must not be modified, opened, or altered in any way. Hardware modifications can compromise electrical safety (exposing the participant to hazardous voltages), void manufacturer warranties, invalidate regulatory certifications (CE, FDA), and introduce signal artifacts or interference. This includes modifying electrode caps, extending cables, soldering connections, bypassing safety circuits, or altering the amplifier's gain, filter, or power settings. If the equipment does not meet the needs of the experiment, the researcher should obtain appropriate equipment rather than modifying existing hardware.

3.4 No Invasive Electrodes¶

As stated in Section 3.1, invasive electrodes are strictly prohibited. This includes but is not limited to: needle electrodes inserted through the skin or scalp, depth electrodes implanted in brain tissue, subdural electrode grids placed between the brain surface and the dura mater, and epidural electrodes placed outside the dura. The NeuroSafetyPolicy in this repository checks for invasive device keywords in recording metadata and will reject any session that references invasive methods. If there is any doubt about whether a particular electrode type is non-invasive, the default assumption should be that it is not suitable for use with this framework, and the researcher should consult the device manufacturer and their ethics committee.

3.5 No Damaged Cables¶

Before each recording session, all cables should be visually inspected for damage. Check for: frayed or exposed wires, cracked or broken connectors, bent or corroded pins, damaged insulation, and kinked or pinched cables. Damaged cables can cause signal artifacts, intermittent connections, or in extreme cases, electrical hazards. If any damage is detected, the cable must be replaced before recording. Cable management during recording is also important — cables should be routed to avoid pulling on the electrodes, creating tension, or creating tripping hazards. Use cable management solutions (cable clips, velcro straps, cable channels) to keep cables organized and secure.

3.6 Avoid Electrical Hazards¶

Electrical safety is critical when using EEG equipment. Key precautions include: always use battery-powered or medically-isolated equipment when possible (see Section 3.7), never connect EEG equipment to mains power in a way that bypasses the manufacturer's isolation circuitry, ensure all equipment is grounded properly according to the manufacturer's instructions, do not use EEG equipment near water or in damp environments, do not use EEG equipment during thunderstorms (lightning strikes can induce dangerous currents in cables), and never touch the participant or electrodes while the equipment is connected to mains power if there is any doubt about electrical isolation. If the participant reports any tingling, burning, or electric shock sensation during recording, stop immediately and disconnect all equipment.

3.7 Battery-Powered Devices Preferred¶

Battery-powered EEG devices are strongly preferred for non-medical research because they provide inherent electrical isolation between the participant and mains power. Battery-powered devices eliminate the risk of leakage currents that could flow through the participant in the event of a fault. If a mains-powered device must be used, ensure it is a medical-grade device with appropriate isolation (IEC 60601-1 compliance for medical electrical equipment) and that it is used in accordance with the manufacturer's instructions. Many modern consumer and research EEG headsets are USB-powered or battery-powered, which provides good isolation. The specific power requirements and safety certifications should be checked in the device documentation.

3.8 Keep Equipment Dry¶

EEG equipment must be kept dry. Water or conductive gel that comes into contact with electronic components can cause short circuits, corrosion, and electrical hazards. When applying conductive gel to electrodes, use care to avoid gel running down cables or into the amplifier. If gel does contact electronic components, disconnect the equipment, clean the affected area with a dry cloth, and allow it to dry completely before reconnecting. Do not use EEG equipment with wet hands. Store equipment in a dry environment when not in use. Some EEG caps are washable — follow the manufacturer's washing instructions carefully, and ensure the cap is completely dry before use.

3.9 Stop if Discomfort Occurs¶

If the participant reports any discomfort during an EEG recording session, recording must stop immediately. Common sources of discomfort include: pressure from tight electrode caps, skin irritation from conductive gel or adhesive, itching under electrodes, headache from prolonged cap wear, dizziness or nausea from visual stimuli presented during the session, and anxiety or claustrophobia from the cap or experimental setup. The NeuroSafetyPolicy.SAFETY_WARNINGS in this repository explicitly states: "STOP IMMEDIATELY if the participant reports discomfort, skin irritation, dizziness, distress, nausea, headache, visual disturbances, or any other adverse symptoms." After stopping, remove the electrodes, clean the participant's skin, and offer appropriate support (water, rest, medical attention if needed). Record the adverse event in the session log and report it to the ethics committee if required.

4. Suggested Non-Medical Research Setup¶

This section describes a recommended hardware and software setup for non-medical EEG research using the CIA neuroadaptive extension. The setup is designed to be practical, affordable, and compliant with the safety requirements described above. Researchers should adapt this setup to their specific needs, institutional resources, and ethical requirements.

4.1 Consumer/Research EEG Headset¶

A consumer or research-grade EEG headset is the primary hardware component. Suitable devices include research-grade systems (e.g., BrainProducts actiCHamp, g.tec g.Nautilus, BioSemi ActiveTwo, EGI HydroCel) and consumer-grade systems (e.g., Emotiv EPOC/X, Muse, OpenBCI Cyton). Research-grade systems offer higher signal quality, more channels, and better technical support, but are significantly more expensive. Consumer-grade systems are more accessible and easier to set up, but may have lower signal quality, fewer channels, and proprietary data formats. For the CIA neuroadaptive extension, the minimum requirement is a device that provides access to raw (or minimally filtered) EEG data at a sampling rate of at least 100 Hz (250 Hz or higher recommended) with at least one channel. The device's SDK or acquisition software should support data export in a format compatible with the ingestion pipeline (CSV, JSON, NumPy, or BIDS).

4.2 Electrode Placement Per Device Manual¶

Electrode placement should follow the device manufacturer's instructions. Research-grade systems typically use the international 10-20 system or a subset thereof, which provides standardized electrode positions on the scalp. Consumer-grade systems often have fixed electrode positions determined by the headset design. The key considerations for electrode placement are: follow the manufacturer's placement guide exactly, ensure good skin-electrode contact (impedance below 10 kΩ for gel-based systems, below the manufacturer's threshold for dry systems), avoid placing electrodes over hair whorls or scars, and verify signal quality on each channel before recording begins. The cia.neuro.eeg_preprocessing.EEGPreprocessor.detect_artifacts() method can help identify channels with poor signal quality (flatline, high variance, extreme amplitude) that may indicate electrode placement issues.

4.3 Recording Laptop¶

A laptop or desktop computer is needed to run the EEG acquisition software, the CIA neuroadaptive pipeline, and the stimulus/task presentation software. Minimum specifications depend on the specific software used, but a modern laptop with at least 8 GB RAM, a multi-core processor, and a USB port (for device connection) should be sufficient. The laptop should be dedicated to the recording session if possible — running other resource-intensive applications alongside EEG acquisition can cause data loss or timing issues. Ensure the laptop is plugged into mains power (battery drain can interrupt recording) and that screen brightness is set to a comfortable level for the participant. The laptop should be positioned so the participant can see the screen clearly (if visual stimuli are presented) while keeping a comfortable distance from the EEG equipment.

4.4 Timestamped Stimulus/Task Presentation¶

If the experiment involves presenting stimuli or tasks to the participant (visual attention tasks, auditory stimuli, cognitive tasks), the stimulus presentation must be precisely timestamped and synchronized with the EEG recording. This synchronization is essential for event-related analysis (e.g., averaging EEG signal epochs time-locked to stimulus onset). Common stimulus presentation software includes PsychoPy, E-Prime, Presentation, and jsPsych (browser-based). The stimulus software should send trigger signals (event markers) to the EEG acquisition system at the exact moment each stimulus appears or each response is made. If hardware triggers are not available, software triggers can be used, but they introduce additional timing uncertainty (typically 5-50 ms depending on the system). The EEGWindow schema in this repository includes a metadata field that can store stimulus timing information for downstream analysis.

4.5 Activity Log¶

A detailed activity log should be maintained for every recording session. The log is a written record (paper or digital) of everything that happens during the session, including: session start and end times, breaks taken, task transitions, participant behavior observations (movements, drowsiness, attention lapses), environmental events (loud noises, equipment malfunctions), any adverse events or participant complaints, and notes on signal quality. The activity log is essential for interpreting the EEG data — it allows researchers to identify periods of artifact, correlate signal changes with external events, and understand the context of the recorded data. The activity log should be stored alongside the EEG data in the dataset directory.

4.6 Synchronized Event Markers¶

Event markers (also called triggers, events, or annotations) are timestamped labels that indicate when specific events occurred during the recording. Common event types include: stimulus onset, stimulus offset, participant response (button press, verbal response), task block start/end, and experimenter notes (e.g., "participant blinked," "noise from hallway"). Event markers are critical for time-locking EEG analysis to specific experimental events. In the BIDS format, event information is stored in a _events.tsv file with columns for onset time, duration, and event type. The ingestion pipeline in this repository (cia.neuro.eeg_ingestion.EEGIngestion) reads BIDS-format datasets and preserves metadata including task labels. For non-BIDS data, event markers can be included in the EEGWindow.metadata dict or stored in a separate file alongside the EEG data.

5. Week-Long Recording Caution¶

Some research designs may involve collecting EEG data across multiple days or even an entire week to capture naturalistic cognitive-state variations. While longitudinal data collection can provide valuable information about within-subject variability, it also introduces significant participant burden and safety considerations that must be carefully managed.

5.1 Continuous Recording May Cause Discomfort¶

Continuous EEG recording over extended periods (hours or days) is physically demanding for the participant. Even lightweight, wireless EEG headsets can cause pressure points, skin irritation, and fatigue with prolonged wear. The conductive gel used in many research-grade systems can dry out over time, increasing impedance and degrading signal quality. Sweat and skin oils can accumulate under electrodes, causing itching and discomfort. Participants may experience headaches, neck strain, or claustrophobia from wearing a cap for extended periods. The NeuroSafetyPolicy.check_duration_limits() method in this repository issues warnings when continuous recording exceeds 60 minutes and strongly recommends splitting sessions that exceed 120 minutes.

5.2 Prefer Short Sessions Across a Week¶

Rather than continuous long-duration recording, a preferred approach is to conduct multiple short sessions distributed across the week. For example, instead of one 6-hour recording, conduct six 45-minute sessions — one per day at the same time of day. This approach provides longitudinal coverage while minimizing participant burden and fatigue. Short sessions also allow for better control over the recording environment (lighting, noise, temperature) and task presentation. Between sessions, the participant can rest, the electrodes can be cleaned and re-prepared, and any equipment issues can be addressed.

5.3 Include Rest Periods¶

Every recording session must include regular rest periods. During rest periods, the participant should be able to relax without any task demands, and ideally without wearing the EEG headset. A common schedule is 10 minutes of rest for every 20-30 minutes of recording. During rest, the participant should be offered water, allowed to stretch, and given the opportunity to report any discomfort. Rest periods are not merely a courtesy — they improve data quality by reducing fatigue-related artifacts and maintaining the participant's alertness and cooperation. The NeuroSafetyPolicy recommends "at least 10 minutes of rest after every 60 minutes of continuous recording" and more frequent breaks if the participant shows signs of fatigue.

5.4 Inspect Skin¶

Before and after each session, the participant's skin under the electrode sites should be visually inspected. Look for signs of irritation, redness, abrasion, allergic reaction, or pressure sores. If any skin issues are observed, they should be documented in the session log, the affected sites should be cleaned with mild soap and water, and the participant should be advised to monitor the sites for 24-48 hours. If irritation persists or worsens, the participant should be referred to appropriate medical attention. For gel-based systems, all gel should be thoroughly cleaned from the scalp and hair after each session using warm water and mild shampoo. Leaving gel on the scalp can cause irritation and may harbor bacteria.

5.5 Document Artifacts¶

Longitudinal recording designs are particularly susceptible to artifacts from changes in electrode placement, skin preparation quality, environmental conditions, and participant state across sessions. All observed artifacts should be documented in the session log and in the data files. Common artifact sources in longitudinal designs include: different cap placements across days (inevitable with repeated donning/doffing), changes in electrode impedance due to variations in skin preparation, progressive skin irritation affecting signal quality, and environmental variations (different room, different time of day, different noise levels). The cia.neuro.eeg_preprocessing.EEGPreprocessor.detect_artifacts() method identifies common artifact types (flatline, extreme amplitude, high variance, missing values) that can be monitored across sessions.

5.6 Avoid Sleep Deprivation¶

Participants should not be asked or encouraged to deprive themselves of sleep for research purposes. Sleep deprivation is a physiological stressor that can have serious health consequences, including impaired cognitive function, mood disturbance, immune suppression, and increased risk of accidents. If the research design involves studying fatigue or drowsiness, participants should be monitored carefully, and sessions should be terminated if the participant shows signs of excessive sleepiness (microsleeps, head drops, difficulty keeping eyes open). The recording schedule should allow participants to maintain their normal sleep patterns. The NeuroSafetyPolicy warns that "extended sessions significantly increase fatigue risk" when durations exceed 120 minutes, and recommends splitting long sessions into shorter blocks.

6. Data Collection Protocol Template¶

A standardized data collection protocol ensures consistency across sessions, facilitates reproducibility, and supports the scientific integrity of the research. The following template is recommended for all EEG data collection using the CIA neuroadaptive extension.

6.1 Subject ID Pseudonym¶

Every participant must be assigned a unique pseudonymous identifier (subject ID) that is used in all data files, logs, and metadata. The subject ID should not contain any personally identifiable information (names, dates of birth, phone numbers, email addresses). Recommended format: sub-XXX where XXX is a sequential number or random code (e.g., sub-001, sub-042). The mapping between subject IDs and real identities must be stored in a separate, secure file that is not included in the dataset. The subject ID is used throughout the CIA pipeline — in EEGRecordingMetadata.subject_id, SubjectNeuroProfile.subject_id, and in BIDS folder naming (sub-01/, sub-02/, etc.).

Each recording session must reference a specific consent form. The consent form reference (e.g., consent form version number, date signed) should be recorded in the session metadata. If the consent form is revised between sessions, the version used for each session must be documented. The EEGRecordingMetadata schema includes a consent_confirmed boolean field and a safety_reviewed boolean field that must be set to True before data can be processed by the safety-checked pipeline. The consent form itself should be stored in a secure location separate from the EEG data, in compliance with the data retention policy.

6.3 Session Start/End¶

Record the precise start and end time of each recording session. Use a consistent time format (ISO 8601 recommended: YYYY-MM-DDTHH:MM:SSZ). Include timezone information or use UTC consistently. If the session includes multiple blocks or tasks, record the start and end time of each block separately. The EEGRecordingMetadata schema includes duration_seconds which can be computed from start/end times. Accurate timing is essential for session management, participant scheduling, and data quality monitoring (e.g., identifying sessions that were cut short due to technical issues).

6.4 Task Labels¶

Each distinct task or experimental condition must be labeled with a descriptive name. Task labels should be short, consistent, and machine-readable (no spaces, use underscores or hyphens). Examples: rest_eyesclosed, rest_eyesopen, nback_2back, p300_oddball, reading_comprehension, mental_arithmetic. In the BIDS format, task labels are included in filenames (e.g., sub-01_task-rest_eeg.vhdr) and in the task entity. The EEGRecordingMetadata schema includes a task field for this purpose. Consistent task labeling across sessions and subjects is essential for automated data processing and analysis.

6.5 Activity Notes¶

Detailed activity notes should be recorded for each session, including: participant behavior (alert, drowsy, fidgety, cooperative), environmental conditions (room temperature, noise level, lighting), equipment status (which channels had good/bad signal quality, any technical issues), deviations from the protocol (missed tasks, shortened blocks), and any other observations that may be relevant for interpreting the data. Activity notes can be stored in a plain text file alongside the EEG data, in the EEGWindow.metadata dict, or in a BIDS scans.tsv file. The NeuroSafetyPolicy and SubjectProfileManager in this repository log warnings and limitations that should be included in the activity notes.

6.6 Electrode Quality¶

Document the signal quality of each electrode at the beginning, during, and at the end of each session. Signal quality can be assessed by: visual inspection of the raw signal on the acquisition software, impedance measurement (for systems that support it), and automated artifact detection using the EEGPreprocessor.detect_artifacts() method. Record which channels had good signal quality, which had marginal quality, and which were unusable. Note any channels that required re-application of gel, repositioning, or replacement. This information is critical for downstream analysis — poor-quality channels should be excluded or interpolated, and the exclusion criteria should be documented.

6.7 Adverse Events¶

Any adverse event that occurs during a recording session must be documented immediately. Adverse events include: skin irritation or allergic reaction, headache, dizziness, nausea, distress or anxiety, equipment malfunction that may have affected the participant, and any other unexpected or undesirable event. For each adverse event, record: a description of what happened, the time it occurred, the severity (mild, moderate, severe), the action taken (session stopped, equipment removed, medical attention sought), and the outcome (symptoms resolved, ongoing, referred to physician). Adverse events must be reported to the ethics committee in accordance with institutional reporting requirements. The NeuroSafetyPolicy.SAFETY_WARNINGS list provides a comprehensive set of adverse symptoms that should be monitored.

6.8 Withdrawal Option¶

The participant's right to withdraw must be reaffirmed at the start of each session and clearly documented in the session log. If a participant withdraws, record: the time of withdrawal, the reason (if provided), what data has been collected up to the withdrawal point, and the participant's preference for handling that data (retain, delete, anonymize). Withdrawn data should be handled in accordance with the consent form provisions and ethics committee requirements. The withdrawal option should be clearly stated in both the consent form and the session protocol.

7. Activities/Tasks Examples¶

The following are examples of cognitive tasks and activities suitable for EEG recording in the context of the CIA neuroadaptive extension. These tasks are designed to elicit distinct EEG patterns that can be used as inputs to the neural-state proxy estimation pipeline. All tasks should be reviewed by the relevant ethics committee before use.

7.1 Resting Eyes Open/Closed¶

Resting-state recording is the simplest and most common EEG paradigm. The participant sits quietly with eyes open (fixating on a crosshair) or eyes closed for a specified duration (typically 2-5 minutes per condition). Eyes-closed resting state is characterized by prominent alpha rhythm (8-13 Hz) over occipital electrodes, while eyes-open resting state shows alpha suppression (alpha blocking). Resting-state data is essential for establishing individual baselines — the NeuralStateEncoder in this repository uses baseline feature vectors (ideally from resting-state recordings) to calibrate proxy estimates via z-score normalization. The SubjectNeuroProfile stores baseline aggregate bandpowers computed from resting-state data.

7.2 Reading¶

During a reading task, the participant reads text passages presented on a screen or on paper. EEG during reading typically shows left-hemisphere dominance (language processing) and beta/theta activity associated with comprehension and working memory demands. Reading tasks can be varied in difficulty (simple words vs. complex academic text) to modulate cognitive workload. The workload proxy in the neural-state encoder is expected to increase with reading difficulty, reflecting increased beta activity. Reading tasks are suitable for studying naturalistic cognitive engagement without artificial experimental stimuli.

7.3 Simple Arithmetic¶

Mental arithmetic tasks (e.g., serial subtraction, multiplication, n-back with numbers) engage working memory and executive function, producing characteristic EEG changes including increased theta and beta power over frontal and central electrodes. The difficulty of the arithmetic task can be modulated to create varying levels of cognitive workload. Simple tasks (single-digit addition) produce lower workload, while complex tasks (serial subtraction of 7s from 1000) produce higher workload. The workload proxy in the neural-state encoder maps increased beta and decreased alpha to higher workload estimates, which should correlate with arithmetic task difficulty.

7.4 Visual Attention Task¶

Visual attention tasks (e.g., Posner cueing task, visual search, sustained attention to response task, SART) require the participant to attend to visual stimuli and respond to specific targets. These tasks are well-suited for studying attention modulation, and the EEG correlates are well-characterized (P300 event-related potential for target detection, N1/P1 for early visual processing, alpha lateralization for spatial attention). The attention proxy in the neural-state encoder maps increased alpha/beta ratio to higher attention estimates. Visual attention tasks can be used to validate whether the attention proxy responds appropriately to experimentally manipulated attention demands.

7.5 Motor Imagery¶

Motor imagery tasks require the participant to imagine performing a movement (e.g., left hand, right hand, feet) without actually moving. Motor imagery produces event-related desynchronization (ERD) in the mu rhythm (8-12 Hz) and beta rhythm (13-30 Hz) over contralateral motor cortex. Motor imagery is the basis for many brain-computer interface (BCI) paradigms and is a well-studied EEG phenomenon. In the context of the CIA neuroadaptive extension, motor imagery can be used to study how different cognitive states modulate the proxy estimates, although the current proxy mappings are primarily based on attention/workload/fatigue rather than motor-specific patterns.

7.6 Memory Task¶

Memory tasks (e.g., n-back, word list recall, Sternberg paradigm) engage working memory and long-term memory systems. EEG correlates of memory load include increased theta power (especially over frontal midline electrodes) and changes in alpha and beta power depending on task demands. The n-back task, where the participant must indicate whether the current stimulus matches the one presented n steps earlier, is a widely used working memory paradigm that allows systematic manipulation of memory load (0-back, 1-back, 2-back, 3-back). Higher memory load should produce higher workload proxy estimates in the neural-state encoder.

7.7 Conversation¶

Natural conversation involves a complex interplay of language processing, social cognition, attention, working memory, and motor control. EEG during conversation is challenging due to movement artifacts (jaw clenching, head movement) but can be recorded with careful experimental design (e.g., using a confederate, pre-planned dialogue, or silent reading/listening of dialogue). The neuroadaptive extension could use conversation data to study how naturalistic cognitive engagement modulates proxy estimates, but the artifact burden in conversational EEG makes this a challenging paradigm for the current preprocessing pipeline.

7.8 Music Listening¶

Music listening engages auditory processing, emotion, memory, and attention systems. EEG responses to music include auditory evoked potentials, rhythmic entrainment (neural activity synchronizing with musical beat), and emotional responses (changes in frontal asymmetry). Music listening is a relatively low-artifact paradigm (participant sits still and listens) and can be used to study passive cognitive engagement. Different types of music (calming vs. stimulating, familiar vs. unfamiliar) can be used to modulate arousal and attention proxies.

7.9 Fatigue/Rest Task¶

A fatigue task involves sustained cognitive performance over an extended period to induce mental fatigue, followed by a rest recovery period. Common fatigue induction paradigms include: the psychomotor vigilance task (PVT), sustained attention tasks, and long-duration working memory tasks. EEG correlates of fatigue include increased theta power, decreased alpha and beta power, increased theta/beta ratio, and slowing of the dominant rhythm. The fatigue proxy in the neural-state encoder maps increased theta/beta ratio to higher fatigue estimates, which should track the progression of fatigue during a sustained task. This paradigm is particularly relevant for validating the fatigue proxy and for studying fatigue-aware neuroadaptive conditioning (e.g., reducing workspace capacity as fatigue increases).

8. Event Markers¶

Event markers (triggers, annotations, events) are timestamped labels that indicate when specific events occurred during EEG recording. They are essential for time-locking EEG analysis to experimental events and for interpreting the recorded data in the context of the experimental paradigm.

8.1 Why They Matter¶

Event markers provide the temporal alignment between EEG data and experimental events. Without event markers, it is impossible to determine which EEG signal segments correspond to which stimuli, responses, or task conditions. Event markers enable event-related analysis techniques (ERP averaging, time-frequency analysis time-locked to events), task segmentation (extracting EEG windows corresponding to specific task blocks), and behavioral-physiological correlation (relating response times, accuracy, or subjective ratings to EEG features). In the CIA neuroadaptive extension, event markers can be stored in the EEGWindow.metadata dict and used to label extracted features for downstream analysis. The BIDS standard uses _events.tsv files to store event information in a structured, machine-readable format.

8.2 Stimulus Onset Timestamps¶

Every experimental stimulus (visual, auditory, tactile) should be marked with a precise timestamp indicating when it was presented. Stimulus onset timestamps should be as accurate as possible — ideally within 1 ms of the actual stimulus presentation. Hardware triggers (TTL pulses sent from the stimulus computer to the EEG amplifier) provide the most accurate timing. Software triggers (events written to the EEG data file by the acquisition software) are less accurate due to operating system scheduling latency but are more convenient. The timestamp format should be consistent throughout the experiment and should be in the same time base as the EEG sampling clock. In the BIDS format, stimulus onset is recorded in the onset column of _events.tsv relative to the start of the EEG recording.

8.3 Task Labels¶

Each event should be labeled with a descriptive task label indicating what type of stimulus or event occurred. Common task label conventions include: stimulus type (target, nontarget, standard, novel), condition (congruent, incongruent, neutral), response (left, right, correct, incorrect), and block markers (block_start, block_end). Task labels should be consistent across subjects and sessions. In BIDS, the trial_type column in _events.tsv contains the task label for each event. The CIA pipeline uses the task field in EEGRecordingMetadata to identify the task associated with a recording.

8.4 Behavioural Responses¶

Participant responses (button presses, verbal responses, eye movements) should also be timestamped and labeled. Response events are important for computing reaction times (RT = response time - stimulus onset), accuracy (correct vs. incorrect), and for segmenting EEG data based on response-locked rather than stimulus-locked timing. Response events should include both the timestamp and the response value (which button was pressed, what was said, etc.). In the BIDS format, response events can be included in the same _events.tsv file as stimulus events, distinguished by their trial_type label.

8.5 Sync with EEG Windows¶

Event markers must be synchronized with the EEG data windows used for analysis. The EEGIngestion class in this repository segments continuous EEG data into fixed-duration windows (configurable via window_seconds and step_seconds parameters). Event markers should be mapped to the corresponding window index based on their timestamp relative to the window start times. If an event falls within a window, the window's metadata dict should include the event information. This mapping allows downstream feature extraction and neural-state encoding to be conditioned on task context (e.g., computing attention proxies separately for target and non-target trials). The BIDS format simplifies this synchronization by storing all timing information in standardized sidecar files that reference the same time base as the EEG data.

9. Data Privacy¶

EEG data is sensitive biometric data that requires robust privacy protections. This section outlines the privacy requirements for EEG data collected and processed using the CIA neuroadaptive extension.

9.1 EEG Is Biometric Data¶

EEG recordings are classified as biometric data under the GDPR (Article 9, which covers "data concerning health") and under many other data protection frameworks. Biometric data derived from a person's physiological characteristics can potentially be used to identify the individual. Research has demonstrated that EEG patterns contain enough individual variation to serve as a biometric identifier, even from a small number of electrodes. This means that EEG data, if not properly protected, could be used to identify participants without their knowledge or consent. The sensitivity of EEG data is compounded by the fact that it may contain information about neurological conditions (epilepsy, sleep disorders), cognitive states (fatigue, attention lapses), and potentially psychiatric conditions.

9.2 Store Encrypted if Possible¶

EEG data files should be encrypted at rest whenever technically feasible. Full-disk encryption on the recording laptop provides baseline protection. Additionally, individual EEG data files or the entire dataset directory should be encrypted using tools such as GPG, AES-256 encryption, or platform-native encryption (BitLocker on Windows, FileVault on macOS, LUKS on Linux). Encryption keys and passwords must be stored separately from the encrypted data, preferably in a hardware security module or a secure password manager. When transmitting EEG data over networks (e.g., uploading to cloud storage, sending to collaborators), always use encrypted transport (HTTPS, SFTP, VPN). The SubjectProfileManager.save_profile() method in this repository saves profiles as plain JSON — researchers should apply file-level encryption to stored profiles if the underlying filesystem is not encrypted.

9.3 Avoid Uploading Identifiable Data¶

EEG data should not be uploaded to public repositories, cloud services, or shared with collaborators in identifiable form. Before sharing, data should be anonymized or pseudonymized by replacing personally identifiable information (names, dates of birth, addresses) with codes. The BIDS format supports pseudonymization through the use of sub- and ses- labels rather than names. OpenNeuro (https://openneuro.org), a public neuroimaging data platform, requires datasets to be properly de-identified before upload, following the BIDS standard's de-identification guidelines. Even when uploading to private or restricted-access repositories, consider the risk of data breaches and apply the principle of minimum necessary access.

9.4 Remove Names¶

All personally identifiable information must be removed from EEG data files and metadata before the data is shared, published, or archived outside the secure research environment. This includes: participant names (replace with subject IDs), dates of birth (replace with age or year of birth), addresses, phone numbers, email addresses, medical record numbers, and any other information that could directly or indirectly identify the participant. Check all data files, metadata files, sidecar files, event files, and log files for identifiable information. The BIDS standard recommends removing dates of birth, acquisition dates, and other temporal information that could be used to re-identify participants, replacing them with relative dates (e.g., age_at_scan instead of date_of_birth). The EEGRecordingMetadata schema in this repository uses a generic subject_id field that should contain only the pseudonymous code.

9.5 Follow Data Retention Policy¶

EEG data should be retained only for as long as necessary for the research purposes specified in the consent form and ethics approval. A data retention policy should specify: how long EEG data will be stored (typically 5-10 years for research data), where it will be stored, who has access, how it will be secured, and when and how it will be destroyed. When the retention period expires, data must be securely destroyed using methods appropriate for digital data (secure deletion with multiple overwrite passes, degaussing of physical storage media). The destruction of data should be documented. Participants should be informed about the data retention policy in the consent form. If a participant withdraws consent, their data should be destroyed according to the consent provisions and the institutional data retention policy, unless there are legal or regulatory requirements to retain it.

10. Integration with Repository¶

This section explains how to integrate physically recorded EEG data with the CIA neuroadaptive extension, from raw data storage through the full processing pipeline to neuroadaptive conditioning.

10.1 Save Raw Data Outside Git¶

Raw EEG data files are typically large (megabytes to gigabytes per session) and should never be committed to the Git repository. Store raw data in a dedicated directory outside the repository (e.g., ~/eeg-data/ or /data/eeg-recordings/). The repository's .gitignore file should include entries for common EEG data directories and formats. Use a consistent directory structure for raw data, ideally following the BIDS convention with sub-XX/ses-YY/eeg/ folders. The CIA repository's datasets/ directory (if it exists) can be used for data that has been specifically prepared for pipeline ingestion, but raw recordings should be stored separately and linked or copied as needed.

10.2 Convert to BIDS When Possible¶

The Brain Imaging Data Structure (BIDS) standard (https://bids.neuroimaging.io) provides a standardized organization for neuroimaging data. Converting raw EEG data to BIDS format ensures interoperability with a wide ecosystem of tools (MNE-Python, MNE-BIDS, EEGlab, FieldTrip) and facilitates data sharing via platforms like OpenNeuro. The BIDS format for EEG data includes specific file naming conventions (sub-XX_task-YY_eeg.vhdr for BrainVision, sub-XX_task-YY_eeg.edf for EDF, etc.), required sidecar files (_events.tsv, _channels.tsv, _eeg.json, _electrodes.tsv), and a dataset_description.json file at the root. If the raw data is not in a BIDS-compatible format, use the MNE-BIDS library to convert it. The cia.neuro.bids_utils.BIDSUtils class provides utilities for validating, summarizing, and navigating BIDS-like folders.

10.3 Use OpenNeuro/BIDS-Like Folders¶

If downloading publicly available EEG datasets from OpenNeuro (https://openneuro.org), place them in a datasets/ directory at the repository root (or alongside the repository). Each dataset should be in its own subdirectory named with the dataset ID (e.g., datasets/ds000001/). OpenNeuro datasets are already in BIDS format, so they can be directly ingested using the EEGIngestion.from_bids_folder() method. The BIDSUtils.generate_openneuro_download_note() method produces download instructions for specific datasets. Note that this repository does NOT automatically download data from OpenNeuro — all downloads must be performed manually by the researcher, and the researcher must verify the dataset licensing terms and data use conditions.

10.4 Run cia neuro ingest¶

Once the raw EEG data is available locally (in CSV, JSON, NumPy, or BIDS format), it can be ingested using the EEGIngestion class. The ingestion process loads the data, validates it, and segments it into fixed-duration windows. Example usage:

from cia.neuro.eeg_ingestion import EEGIngestion

ingestion = EEGIngestion(window_seconds=2.0, step_seconds=1.0)

# From CSV
metadata, windows = ingestion.from_csv(
    path="data/recording.csv",
    sampling_rate_hz=250.0,
    channel_names=["Fz", "Cz", "Pz", "Oz"],
    subject_id="sub-001",
)

# From BIDS folder (requires mne and mne-bids)
metadata, windows = ingestion.from_bids_folder(
    path="datasets/ds000001",
    subject="01",
    task="rest",
)

10.5 Run cia neuro features¶

After ingestion, each EEG window can be preprocessed and features extracted using the preprocessing and feature extraction modules:

from cia.neuro.eeg_preprocessing import EEGPreprocessor
from cia.neuro.eeg_feature_extraction import EEGFeatureExtractor

preprocessor = EEGPreprocessor(bandpass_low_hz=1.0, bandpass_high_hz=40.0)
extractor = EEGFeatureExtractor(use_scipy_welch=True)

for window in windows:
    # Preprocess
    preprocessed = preprocessor.clean(window)

    # Extract features
    features = extractor.extract(preprocessed)

    print(f"Alpha/beta ratio: {features.alpha_beta_ratio:.4f}")
    print(f"Theta/beta ratio: {features.theta_beta_ratio:.4f}")
    print(f"Signal quality: {features.signal_quality:.4f}")

10.6 Run cia neuro closed-loop-demo¶

The full closed-loop neuroadaptive pipeline connects EEG data to the CIA cognitive architecture through the BCIAdapter:

from cia.neuro.bci_adapter import BCIAdapter
from cia.neuro.online_signal_stream import MockSignalStream
from cia.simulation import CombinedConsciousnessIndicatorSystem

# Create system and adapter
system = CombinedConsciousnessIndicatorSystem()
adapter = BCIAdapter(stream=MockSignalStream())
adapter.stream.connect()

# Run closed-loop step
results = adapter.run_closed_loop_step(
    system=system,
    input_text="The red ball rolled behind the screen.",
)

print(f"Neural state attention: {results['neural_state'].attention_proxy:.3f}")
print(f"Neural state fatigue: {results['neural_state'].fatigue_proxy:.3f}")
print(f"Control signal attention_bias: {results['control_signal'].attention_bias:.3f}")

# The system's components are now neuroadaptively conditioned
# Run a cognitive cycle with the conditioned system
report = system.run_cycle("Analyze the scene described above.")

11. Explicit Prohibition¶

The following claims and uses are explicitly prohibited by this repository. These prohibitions are enforced through the NeuroSafetyPolicy in cia.neuro.safety, the scientific boundary warnings in every module, and the caveat strings in every output schema.

11.1 Do NOT Claim Mind Upload¶

It is explicitly prohibited to claim or suggest that EEG data can be used to "upload" a person's mind, consciousness, personality, or identity into a computer or AI system. EEG data consists of statistical measurements of scalp electrical activity. It does not contain sufficient information to reconstruct a person's thoughts, memories, personality, or subjective experience. Claims of mind uploading based on EEG data are scientifically unfounded and may be used to mislead the public about the capabilities of this technology. Any publication, presentation, or communication about this repository must explicitly disclaim any mind upload capabilities.

11.2 Do NOT Claim Consciousness Transfer¶

It is explicitly prohibited to claim or suggest that EEG-based personalization transfers consciousness, subjective experience, or sentience from the participant to the AI system. The neuroadaptive conditioning mechanism in this repository adjusts quantitative parameters (attention weights, workspace capacity, learning rate, welfare thresholds) based on EEG-derived proxy estimates. These parameter adjustments change the system's computational behavior but do not create, transfer, or replicate consciousness. The NeuroadaptiveControlSignal schema includes a caveat stating: "This control signal conditions the cognitive architecture based on EEG-informed neural-state proxies. It does NOT transfer consciousness, subjective experience, or personal identity into the AI system."

11.3 Do NOT Claim AI Becomes the Subject¶

It is explicitly prohibited to claim or suggest that the AI system becomes the participant, takes on their identity, or represents a digital version of the participant. The SubjectNeuroProfile stores statistical calibration data (baseline bandpowers, spectral ratio offsets) for improving proxy estimation accuracy. It does not contain or represent the participant's identity, selfhood, or consciousness. The profile caveat states: "This profile captures statistical EEG features only. It does NOT capture identity, selfhood, consciousness, or subjective experience." The AI system remains a software system processing statistical proxies — it does not become the participant.

11.4 Do NOT Use for Diagnosis or Treatment¶

It is explicitly prohibited to use this software for medical diagnosis, treatment, monitoring, or any other clinical purpose. The NeuroSafetyPolicy.SAFETY_WARNINGS states: "This software is NOT a medical device. It is not intended for diagnosis, treatment, or monitoring of any medical condition." The preprocessing pipeline in EEGPreprocessor is a research-grade implementation with known limitations (e.g., the bandpass filter is an FFT-based placeholder, not a proper FIR/IIR filter). The feature extraction and neural-state encoding modules produce heuristic proxy estimates that have not been clinically validated. Using this software for medical purposes could result in incorrect assessments, missed diagnoses, or inappropriate treatment decisions, with potentially serious health consequences. If you need EEG analysis for clinical purposes, use validated, regulated medical devices and software.

References¶

Resource	URL / Citation
OpenNeuro — Public neuroimaging data platform	https://openneuro.org
BIDS — Brain Imaging Data Structure	https://bids.neuroimaging.io
MNE-Python — EEG/MEG analysis toolkit	https://mne.tools
MNE-BIDS — BIDS support for MNE	https://mne.tools/mne-bids/
Butlin et al. (2023) — Consciousness in Artificial Intelligence	Butlin, C., Long, L., Bhatt, A., et al. (2023). "Consciousness in Artificial Intelligence: Insights from the Science of Consciousness." arXiv preprint.
Butlin et al. (2025) — Identifying indicators of consciousness in AI systems	Butlin, C., et al. (2025). "Identifying indicators of consciousness in AI systems."
Global Workspace Theory	Baars, B.J. (2005). "Global Workspace Theory of Consciousness."
Attention Schema Theory	Graziano, M.S.A. & Webb, T.W. (2015). "The Attention Schema Theory."
Predictive Processing	Clark, A. (2013). "Whatever Next? Predictive Brains, Situated Agents, and the Future of Cognitive Science."
Recurrent Processing Theory	Lamme, V.A.F. (2006). "Towards a True Neural Stance on Consciousness."

12. EEG and Behavioural Data for Personalization¶

12.1 EEG Plus Behavioural Data May Support Personalization¶

The combination of EEG recordings with behavioural observations can support the construction of personalized cognitive-state profiles for research purposes. By correlating EEG-derived proxy estimates (attention, workload, fatigue, arousal) with simultaneously collected behavioural measures (response times, choice patterns, task performance), researchers can build individualized models that capture how a specific person's neural activity patterns relate to their observable cognitive behaviour. This personalized calibration can improve the accuracy of proxy estimates for that individual relative to population-level defaults, as described in the individual calibration experiment (Document 16, Section 5).

The SubjectNeuroProfile in this repository already supports per-subject calibration via baseline bandpowers, spectral ratio references, and individual calibration offsets. When augmented with behavioural data, the profile becomes a richer representation of the individual's characteristic cognitive-state patterns. The Subject-Specific Cognitive Emulation extension (Document 17) formalizes this multi-channel approach by integrating EEG, behavioural traces, writing samples, preference surveys, and voluntary autobiographical memory cues into a composite cognitive emulation profile.

12.2 Neural State Proxies Combined with Behavioural Observations¶

EEG-informed neural state proxies and behavioural observations are complementary data sources. Each captures different aspects of the participant's cognitive state:

Data Source	What It Captures	Limitations
EEG proxies	Physiological correlates of attention, workload, fatigue, arousal — measured at millisecond resolution	Spatially blurred, noisy, proxy-based, no access to mental content
Behavioural observations	Observable performance patterns — response times, accuracy, choice preferences, consistency	Influenced by many non-cognitive factors (motor speed, motivation, strategy), no access to reasoning
Combined	Correlation between physiological state and observable performance — richer individual-level model	Still captures only observable, statistical patterns; no access to subjective experience

The fusion of EEG and behavioural data can improve proxy estimation by providing convergent validation (e.g., does the EEG fatigue proxy increase when behavioural performance deteriorates?). However, the combined data still captures only the observable, third-person aspects of cognition. The subjective, first-person experience of the participant — what it feels like to be fatigued, how they experience the difficulty of a task, their inner monologue during decision-making — remains inaccessible to both EEG and behavioural measurement.

12.3 This Does NOT Enable Reading the Subject's Mind¶

The combination of EEG and behavioural data does not enable mind-reading, thought decoding, or any form of access to the participant's private mental content. This cannot be emphasized strongly enough:

EEG records scalp electrical activity, not thoughts. The EEG signal is a highly aggregated, spatially blurred representation of cortical postsynaptic activity. No amount of signal processing, machine learning, or data fusion can extract subjective experience from this signal.
Behavioural observations record actions, not intentions. A response time measurement tells you how long a person took to press a button, not why they took that amount of time or what they were thinking while deciding.
Combining data sources does not create new capabilities. Adding behavioural data to EEG does not suddenly enable thought reading. It provides a richer statistical profile of observable patterns, but the fundamental limitation remains — subjective experience is not accessible through third-person measurements.
Proxies are not measurements. The attention proxy derived from alpha/beta ratio is a heuristic approximation, not a direct measure of attention. The workload proxy derived from spectral features is a statistical estimate, not a literal reading of cognitive load. These proxies capture probabilistic tendencies, not deterministic truths about the participant's mental state.

The NeuralStateEstimate output from this repository includes a mandatory caveat that states: "Neural-state estimates are proxies and do not decode thoughts or consciousness." This caveat applies equally to any combined EEG-behavioural analysis.

12.4 Explicit Warning Against Presenting the System as the Subject¶

Even when the CIA system's parameters are conditioned on a specific individual's EEG and behavioural data, the system must never be presented as, referred to as, or treated as that individual. This prohibition is absolute and applies in all contexts:

Prohibited Action	Example	Why It Is Prohibited
Presenting the system as the subject	"This is [subject's name]'s digital version"	The system is a software artifact with conditioned parameters, not a person
Using first-person attribution	"The system (as [subject]) prefers X"	The system does not have preferences; it has conditioned parameter values
Creating the appearance of interaction	Setting up the system to "chat" as if it were the subject	This is impersonation, which is explicitly prohibited in the safety policy
Omitting the caveat	Presenting emulation outputs without the mandatory disclaimer	The caveat is mandatory in all outputs without exception
Using the subject's real name in outputs	"Subject [real name]'s attention proxy is 0.7"	Only pseudonymous identifiers (e.g., sub-001) may be used

The safety_validator.py module in the Subject-Specific Cognitive Emulation extension (Document 17) implements automated checks for these violations, including prohibited phrase detection, identity assertion checking, and impersonation guardrails. Researchers must also exercise professional judgement to prevent subtler forms of misrepresentation.

12.5 The Personalization Layer Estimates Behavioural Tendencies, Not Inner Experience¶

The personalization layer — whether implemented through the basic SubjectNeuroProfile or the extended SubjectCognitiveProfile — produces estimates of observable, statistical behavioural tendencies. It does not produce estimates of inner experience, subjective states, or phenomenal consciousness. The distinction is fundamental:

What the Personalization Layer Estimates	What It Does NOT Estimate
This individual tends to show elevated alpha power during reading tasks	What reading feels like for this individual
This individual's response times increase under high workload	Why this individual slows down under high workload
This individual typically selects conservative options in multi-choice scenarios	What this individual is thinking when making a choice
This individual's fatigue proxy increases after 40 minutes of sustained attention	How fatigue subjectively feels for this individual
This individual's EEG baseline differs from the population average	What this individual's inner experience is like

The personalization layer is a computational tool for research into individual differences in cognitive-state correlates. It is a statistical calibration mechanism, not a window into anyone's mind. All claims about the personalization layer must be framed in terms of "estimated behavioural tendencies," "statistical patterns," and "proxy estimates" — never in terms of "capturing experience," "reading thoughts," or "understanding the person."

12.6 Summary of Boundaries¶

Can Be Done	Cannot Be Done
Estimate per-subject EEG baselines for proxy calibration	Read or decode the subject's thoughts
Correlate EEG proxies with behavioural performance	Transfer the subject's consciousness to the system
Condition CIA system parameters based on individual profiles	Make the CIA system become or represent the subject
Build statistical models of individual cognitive patterns	Capture the subject's subjective experience
Improve proxy accuracy through individual calibration	Replicate the subject's identity or personhood
Use multi-channel data for richer personalization	Access the subject's private mental content