16 - Neuroadaptive Experiments¶

Potential EEG-Informed Experiments for the CIA Neuroadaptive Extension¶

SCIENTIFIC BOUNDARY: This document describes potential research experiments using EEG-informed proxy estimates to modulate the CIA cognitive architecture. All experiments use proxy-based measurements and do not demonstrate consciousness, subjective experience, or sentience in any system. EEG does not read thoughts directly. Results from these experiments are theory-derived architectural indicator measurements subject to significant limitations. No claims of mind-reading, consciousness transfer, or AI self-awareness are supported.

1. Attention Modulation Experiment¶

1.1 Objective¶

Investigate whether the CIA system's attention controller responds measurably to EEG-derived attention proxy estimates from a human participant performing a visual attention task. The experiment tests whether the neuroadaptive conditioning loop produces systematic changes in the attention controller's weight distribution when the participant's measured attention level changes.

1.2 Design¶

The participant performs a sustained visual attention task (e.g., the Sustained Attention to Response Task, SART) while wearing an EEG headset. EEG data is segmented into windows, preprocessed, and fed through the feature extraction and neural-state encoding pipeline to produce attention proxy estimates. The NeuroadaptiveConditioner converts these estimates into control signals that modulate the AttentionController's salience and goal-relevance weights.

Conditions:

Condition	Description	Expected EEG Pattern
High attention	Participant actively attending to task stimuli	Elevated alpha/beta ratio, high signal quality
Low attention	Participant in resting state or distracted	Reduced alpha/beta ratio, variable signal quality

Dependent variables: - Change in AttentionController._weights.salience and AttentionController._weights.goal_relevance - Magnitude of attention_bias in the NeuroadaptiveControlSignal - CIA indicator scores for the ATTENTION_SCHEMA and GLOBAL_BROADCAST categories

1.3 Procedure¶

Record a 2-minute resting-state baseline with eyes open to establish the participant's baseline EEG features.
Build a SubjectNeuroProfile from the resting-state data using SubjectProfileManager.update_profile().
Begin the visual attention task. For each 2-second EEG window: a. Preprocess the window using EEGPreprocessor.clean(). b. Extract features using EEGFeatureExtractor.extract(). c. Encode the neural state using NeuralStateEncoder.encode(). d. Convert to a control signal using NeuroadaptiveConditioner.convert_state_to_control_signal() with the subject profile. e. Apply the control signal to the AttentionController using apply_to_attention_controller(). f. Log the attention weights, control signal values, and proxy estimates.
Compare attention weights between high-attention and low-attention windows.
Restore original weights using restore_attention_controller_weights().

1.4 Expected Outcomes¶

When the participant is actively attending (high attention proxy), the conditioner should produce a positive attention_bias, which should increase the salience and goal_relevance weights on the attention controller. When the participant is distracted or resting (low attention proxy), the weights should return toward baseline or shift in the opposite direction. This would demonstrate that the neuroadaptive loop successfully propagates EEG-informed attention information into the CIA system's computational parameters.

1.5 Caveats¶

Proxy-based: The attention proxy is derived from EEG bandpower heuristics (alpha/beta ratio), not from direct measurement of attention. The alpha/beta ratio is a well-known but noisy correlate of attentional state with significant individual variability.
No subjective experience: Changes in attention controller weights are computational parameter adjustments, not evidence that the system experiences attention or that the participant's attention has been "transferred."
Confounding factors: EEG attention proxies can be affected by eye movements, muscle tension, environmental noise, caffeine, time of day, and many other factors that are unrelated to actual attentional engagement.
Artifact sensitivity: The attention proxy is sensitive to EEG artifacts (blinks, muscle activity, line noise) that may produce spurious proxy changes unrelated to the participant's actual attentional state.
Causal direction: Even if attention weights change consistently with EEG proxies, this does not establish that the neuroadaptive conditioning improves the CIA system's performance or produces meaningful behavioral outcomes.

2. Workload Modulation Experiment¶

2.1 Objective¶

Investigate whether the CIA system's global workspace capacity can be modulated by EEG-derived workload proxy estimates from a participant performing tasks at varying difficulty levels. This experiment tests the neuroadaptive conditioning of the workspace bottleneck based on the participant's measured cognitive load.

2.2 Design¶

The participant performs an n-back working memory task at multiple difficulty levels (0-back, 1-back, 2-back, 3-back). Each difficulty level imposes a different cognitive workload, which is expected to produce different EEG signatures (higher n-back → higher theta power, lower alpha power → higher workload proxy). The NeuroadaptiveConditioner converts the workload proxy into a workspace_priority_bias that modulates the GlobalWorkspace capacity.

Conditions:

Condition	Task	Expected EEG Pattern
Low workload	0-back (simple detection)	Lower theta/beta ratio, higher alpha power
Medium workload	1-back or 2-back	Moderate theta/beta ratio
High workload	3-back or higher	Higher theta/beta ratio, reduced alpha

Dependent variables: - Change in GlobalWorkspace._capacity (effective broadcast capacity) - Magnitude and direction of workspace_priority_bias - Number of items broadcast per cognitive cycle - CIA indicator scores for GLOBAL_BROADCAST category

2.3 Procedure¶

Record resting baseline for subject profile creation.
For each difficulty level (randomized order): a. Present the n-back task for 2 minutes. b. Process EEG windows in real-time (or offline with windowed analysis). c. Compute workload proxy estimates for each window. d. Apply neuroadaptive conditioning to the global workspace. e. Run CIA cognitive cycles with the conditioned workspace and log workspace capacity and broadcast counts.
Compare workspace capacity and broadcast behavior across difficulty levels.
Restore original workspace capacity after each block.

2.4 Expected Outcomes¶

Higher workload (higher n-back difficulty) should produce a more negative workspace_priority_bias, causing the workspace capacity to decrease. This implements the theoretical principle that cognitive load narrows the workspace bottleneck — fewer items can be simultaneously held in global access when cognitive resources are strained. Lower workload should allow the workspace to expand to its full capacity.

2.5 Caveats¶

Proxy-based: The workload proxy is derived from the inverse of the alpha/beta ratio heuristic. The relationship between EEG spectral features and cognitive workload is well-studied but imprecise. Different individuals may show different EEG signatures for the same subjective workload level.
No consciousness claim: Narrowing the workspace capacity based on workload is a computational resource management strategy, not a demonstration that the system experiences cognitive load or that the participant's workload has been transferred.
Task generalization: The workload proxy may not generalize across task types. A workload proxy calibrated on n-back tasks may not accurately reflect workload during reading, conversation, or visual search.
Hysteresis and lag: EEG workload proxies may lag behind the actual cognitive state transition (e.g., switching from 0-back to 3-back) by several seconds, limiting the temporal precision of the neuroadaptive response.
Capacity changes vs. performance: Even if workspace capacity changes are reliably observed, this does not demonstrate that the neuroadaptive conditioning improves or degrades the CIA system's task performance. Separate behavioral experiments would be needed to assess functional impact.

3. Prediction Violation with EEG Proxy¶

3.1 Objective¶

Investigate whether the CIA system's predictive world model learning rate can be modulated by EEG-derived prediction error proxy estimates from a participant exposed to expected and unexpected events. This experiment bridges Predictive Processing theory with the neuroadaptive extension.

3.2 Design¶

The participant performs an oddball task (frequent standard stimuli interspersed with rare deviant stimuli) while EEG is recorded. Deviant (unexpected) stimuli typically elicit larger prediction error responses in the EEG — increased theta and gamma power around 200-400 ms post-stimulus. The NeuroadaptiveConditioner converts the prediction error proxy into a prediction_sensitivity_adjustment that modulates the PredictiveWorldModel's learning rate.

Conditions:

Condition	Stimulus Type	Expected EEG Pattern
Expected	Frequent standard (80%)	Lower gamma power, prediction error proxy near baseline
Unexpected	Rare deviant (20%)	Elevated gamma power, higher prediction error proxy

Dependent variables: - Change in PredictiveWorldModel._learning_rate - Magnitude of prediction_sensitivity_adjustment - Rate of prediction error reduction in the world model after expected vs. unexpected events - CIA indicator scores for PREDICTIVE_MODELING category

3.3 Procedure¶

Record resting baseline for subject profile creation.
Present the oddball paradigm with standard and deviant stimuli.
For each EEG window: a. Preprocess, extract features, encode neural state. b. Compute prediction error proxy from gamma bandpower. c. Apply neuroadaptive conditioning to the predictive world model's learning rate. d. Present a text description of the stimulus event to the CIA system. e. Log the learning rate, prediction error, and CIA system's response.
Compare learning rate changes and world model adaptation speed between standard and deviant trials.

3.4 Expected Outcomes¶

During deviant trials (unexpected events), the elevated gamma power should produce a higher prediction error proxy, which should increase the prediction_sensitivity_adjustment, raising the world model's learning rate. This should cause the CIA system's predictive model to update its hypotheses more rapidly in response to unexpected events — a computational analogue of the Predictive Processing principle that prediction error drives belief updating.

3.5 Caveats¶

Proxy-based: The prediction error proxy is derived from gamma bandpower, which is a rough correlate of prediction error responses in the EEG literature. Gamma power is also affected by attention, arousal, muscle activity, and environmental electromagnetic interference. The proxy is a noisy and indirect measure of prediction error.
No phenomenal surprise: The CIA system does not "feel" surprise when its learning rate increases. The learning rate adjustment is a purely computational mechanism — an optimization parameter change, not an emotional or phenomenal response.
Temporal alignment: The oddball paradigm requires precise temporal alignment between EEG windows and stimulus events. Without event markers, it is difficult to determine which EEG windows correspond to standard vs. deviant trials. The BIDS _events.tsv file should be used for precise synchronization.
Gamma band challenges: Gamma band (30-45 Hz) is the noisiest EEG frequency band and is particularly susceptible to muscle artifact (EMG contamination from jaw clenching, neck tension). Careful artifact detection and removal is essential before using gamma power as a prediction error proxy.
Model convergence: Increasing the learning rate makes the predictive model more responsive but also more volatile — it may oscillate or fail to converge if the learning rate is too high. The learning rate adjustment should be bounded (as implemented in the conditioner, clamped to [0.01, 0.99]).

4. Fatigue-Aware Workspace Throttling¶

4.1 Objective¶

Investigate whether the CIA system can implement a "fatigue-aware" mode that progressively reduces its computational scope (workspace capacity, attention focus) as the participant's measured fatigue increases during a sustained task. This experiment explores a potential application of neuroadaptive AI for human-computer collaboration safety.

4.2 Design¶

The participant performs a sustained cognitive task (e.g., psychomotor vigilance task, PVT) for 30-60 minutes while EEG is recorded. Fatigue is expected to increase over time, producing characteristic EEG changes: increased theta power, decreased alpha and beta power, and increased theta/beta ratio. The NeuroadaptiveConditioner converts the fatigue proxy into an uncertainty_adjustment and a negative attention_bias, effectively reducing the CIA system's computational engagement as the participant becomes more fatigued.

Conditions:

Phase	Duration	Expected State
Alert (baseline)	0-10 minutes	Low fatigue proxy, normal system behavior
Transition	10-30 minutes	Gradually increasing fatigue proxy
Fatigued	30-60 minutes	High fatigue proxy, reduced system engagement

Dependent variables: - Time course of fatigue proxy (theta/beta ratio) - Progressive changes in workspace capacity and attention weights - Rate of uncertainty_adjustment increase over time - Relationship between subjective fatigue ratings (collected periodically) and EEG fatigue proxy

4.3 Procedure¶

Record resting baseline for subject profile creation.
Begin the sustained vigilance task.
Every 2 minutes, collect a subjective fatigue rating (e.g., Karolinska Sleepiness Scale, 1-9).
For each EEG window: a. Preprocess, extract features, encode neural state. b. Apply neuroadaptive conditioning to attention, workspace, and welfare monitor. c. Log all proxy estimates, control signals, and conditioned parameters.
Analyze the time course of fatigue proxy vs. subjective ratings.
Analyze the progressive changes in CIA system parameters.
Restore all original parameters after the session.

4.4 Expected Outcomes¶

The fatigue proxy should increase monotonically over the session (reflecting increasing theta/beta ratio as the participant becomes more fatigued). This should produce progressively more negative attention_bias and progressively more positive uncertainty_adjustment, reducing the CIA system's attentional focus and increasing its uncertainty. The workspace capacity may also decrease. Subjective fatigue ratings should correlate positively with the EEG fatigue proxy, providing convergent validation.

4.5 Caveats¶

Proxy-based: The fatigue proxy (theta/beta ratio) is a well-known correlate of vigilance and sleepiness, but it is not a direct measure of fatigue. Individual variability is substantial, and the theta/beta ratio can be affected by factors other than fatigue (e.g., medication, time of day, individual neurophysiology).
No transfer of experience: The CIA system does not "feel" tired when its uncertainty increases. The parameter changes are purely computational responses to statistical proxy values. The system's "fatigue-aware" behavior is an engineered response, not an emergent property of subjective experience.
Circularity risk: If the CIA system's reduced engagement causes the participant to become less engaged with the task, this could create a positive feedback loop (less engagement → more fatigue → less engagement). This interaction effect should be considered in the experimental design.
Individual differences: Some individuals show minimal theta/beta ratio changes even when subjectively fatigued, while others show large changes. The subject profile and baseline calibration help, but individual differences remain a significant source of noise.
Task specificity: Fatigue manifests differently across tasks. The theta/beta ratio changes observed during a vigilance task may not transfer to a reading task or a conversation task.

5. Individual Calibration Across Sessions¶

5.1 Objective¶

Evaluate the effectiveness of per-subject calibration using the SubjectProfileManager and NeuralStateEncoder.fit_baseline() methods. This experiment tests whether individualized calibration improves the reliability and validity of proxy estimates compared to population-level defaults, and how calibration evolves across multiple sessions.

5.2 Design¶

Multiple participants each complete 3-5 recording sessions on different days. Each session includes a resting-state baseline block followed by a task block (e.g., n-back at a fixed difficulty). The SubjectProfileManager builds and updates each participant's profile across sessions, and the NeuralStateEncoder uses the profile for individualized proxy estimation.

Sessions:

Session	Content	Purpose
Session 1	Resting baseline + task	Initial profile creation
Session 2	Resting baseline + task	Profile update, drift measurement
Session 3-5	Resting baseline + task	Profile refinement, stability assessment

Dependent variables: - Proxy estimate variance across sessions (within-subject stability) - Proxy estimate correlations with behavioral measures (reaction time, accuracy) - Drift score evolution across sessions - Difference between population-level and individualized proxy estimates

5.3 Procedure¶

For each session: a. Record resting-state baseline (eyes open, 2 minutes). b. Record task block (n-back, 10 minutes). c. Preprocess all windows. d. Extract features from all windows. e. Encode neural states using both population-level heuristics and individualized calibration. f. Update the subject profile with the session's features and states. g. Compute drift score between current features and profile baseline. h. Save the updated profile.
Across sessions: a. Compare proxy estimates with and without individual calibration. b. Assess within-subject stability of proxy estimates. c. Assess profile drift and determine whether recalibration is needed. d. Correlate proxy estimates with behavioral measures.

5.4 Expected Outcomes¶

Individualized calibration should reduce the variance of proxy estimates across sessions (because the baseline accounts for each individual's characteristic EEG features). The drift score should remain low across sessions if the participant's EEG features are stable, but may increase if there are systematic changes (different electrode placement, time-of-day effects, caffeine intake). The correlation between proxy estimates and behavioral measures should be stronger with individualized calibration than without.

5.5 Caveats¶

Proxy-based: Individual calibration improves the accuracy of proxy estimates relative to each individual's baseline, but the underlying heuristics (alpha/beta ratio → attention, theta/beta ratio → fatigue) remain approximate. Better calibration of an approximate proxy does not make it an exact measurement.
No identity capture: The SubjectNeuroProfile captures statistical EEG features for proxy calibration. It does not capture the participant's identity, consciousness, or subjective experience. The profile caveat explicitly states this.
Session-to-session variability: Even with calibration, EEG features can vary substantially across sessions due to differences in electrode placement, skin preparation, environmental conditions, participant state, and many other factors. Some of this variability is irreducible.
Calibration data requirements: The NeuralStateEncoder.fit_baseline() method requires multiple feature vectors for meaningful standard deviation estimates. With very few data points, the baseline statistics may be unreliable, and z-score normalization may amplify noise rather than reduce it.
Overfitting risk: If calibration is performed on a very small dataset (e.g., a single session), the baseline may overfit to the specific conditions of that session and not generalize well to other sessions, tasks, or days.

6. Controls and Limitations¶

6.1 Essential Controls¶

Every neuroadaptive experiment described in this document should include the following controls:

No-EEG control: Run the CIA system without any neuroadaptive conditioning (default parameters) and compare indicator scores, attention weights, workspace capacity, and learning rate against the conditioned conditions. This establishes whether the neuroadaptive conditioning produces any meaningful change above baseline variability.

Sham conditioning control: Apply random or constant control signals (e.g., attention_bias = 0.0 for all windows) while still processing the EEG data through the full pipeline. This controls for any effects of the processing pipeline itself (e.g., computational overhead, timing changes) that are unrelated to the actual proxy values.

Cross-subject comparison: Run the experiment with multiple participants to assess individual variability in proxy estimates and neuroadaptive responses. Results that are consistent across participants are more robust than results from a single participant.

Order and learning effects: Counterbalance or randomize the order of experimental conditions across participants and sessions to control for order effects, learning effects, and fatigue.

Artifact control: Use the EEGPreprocessor.detect_artifacts() method to identify and flag windows with significant artifacts. Exclude or separately analyze artifacted windows to ensure that results are not driven by non-neural signals.

6.2 Statistical Limitations¶

Sample size: EEG experiments typically require 10-30 participants for adequate statistical power, depending on the effect size. Single-subject experiments (case studies) can be informative for proof-of-concept but cannot be generalized.

Multiple comparisons: If multiple proxy estimates, multiple conditions, or multiple time windows are compared, appropriate corrections for multiple comparisons (e.g., Bonferroni, FDR) should be applied to control the false positive rate.

Effect size: Even if neuroadaptive conditioning produces statistically significant changes in CIA system parameters, the effect sizes may be small. The practical significance of small parameter changes should be carefully evaluated.

Reproducibility: Results should be reproducible across sessions, participants, and EEG hardware configurations. If results are specific to a single participant or hardware setup, their generalizability is limited.

6.3 Technical Limitations¶

Sampling rate requirements: The current feature extraction pipeline requires a minimum sampling rate of approximately 100 Hz to resolve the gamma band (30-45 Hz). Lower sampling rates will miss gamma-band information entirely.

Channel count requirements: The proxy heuristics are based on aggregate (mean across channels) bandpowers. With very few channels (1-2), the aggregate values are noisy and may not represent the underlying neural activity accurately. At least 4-8 channels are recommended for reasonable proxy quality.

Preprocessing limitations: The FFT-based bandpass filter in EEGPreprocessor is a placeholder without proper roll-off characteristics. For publication-quality research, MNE-Python's filtering or scipy.signal filters should be used instead.

No real-time processing: The current pipeline processes data offline (from files). Real-time neuroadaptive conditioning would require streaming data integration, which is documented as a future extension.

6.4 Ethical Limitations¶

No medical claims: None of the experiments described in this document should be framed as medical assessments, diagnostic tools, or therapeutic interventions. The CIA system is not a medical device and the proxy estimates are not clinical measures.

Participant welfare: The fatigue-aware throttling experiment (Section 4) is particularly sensitive — it involves intentionally studying participants while they become fatigued. Researchers must ensure that participants are never pushed beyond safe fatigue levels and that sessions can be terminated at any time at the participant's request.

Data sensitivity: EEG data collected for these experiments is biometric and must be stored, processed, and shared in accordance with applicable data protection regulations (GDPR, HIPAA, etc.). The data privacy guidelines in the EEG safety document (Document 13) apply.

Informed consent: All experiments require informed consent that explicitly describes the EEG recording procedures, the neuroadaptive conditioning mechanism (adjusting AI system parameters based on EEG), and the data handling practices.

References¶

Resource	Citation / URL
Butlin et al. (2023)	Butlin, C., et al. "Consciousness in Artificial Intelligence: Insights from the Science of Consciousness."
Butlin et al. (2025)	Butlin, C., et al. "Identifying indicators of consciousness in AI systems."
Global Workspace Theory	Baars, B.J. (2005); Shanahan & Baars (2005)
Attention Schema Theory	Graziano & Webb (2015)
Predictive Processing	Clark (2013); Friston (2010)
Recurrent Processing Theory	Lamme (2006)
OpenNeuro	https://openneuro.org
BIDS / EEG-BIDS	https://bids.neuroimaging.io
MNE-Python	https://mne.tools
MNE-BIDS	https://mne.tools/mne-bids/
Theta/Beta Ratio & Vigilance	Klimesch, W. (1999). "EEG alpha and theta oscillations reflect cognitive and memory performance."
N-Back & Working Memory	Kirchner, W.K. (1958). "Age differences in short-term retention of rapidly changing information."
Oddball Paradigm	Sutton, S., et al. (1965). "Evoked-potential correlates of stimulus uncertainty."