Skip to content

Clinical Safety, Ethics, and IRB Governance

Defense-in-Depth Participant Protection and Client-Side Data Stewardship

Because the Digital Mirror operates at the intersection of real-time neurotelemetry, affective computing, and psychotherapeutic intervention research, participant safety and data privacy are treated as foundational hardware and software constraints rather than post-hoc policy guidelines.

The research prototype is engineered specifically to support Institutional Review Board (IRB) evaluation, hospital ethics committee review, and clinical trial sponsorship through a five-layer defense-in-depth governance architecture.

DEFENSE-IN-DEPTH RESEARCH SAFETY & PRIVACY GOVERNANCE STACK LAYER 1: COLD PATH Pre-Session Curation • Clinician Protocol Authoring & Sign-Off • "Session Zero" Readiness Pacing • Pre-Compiled Safe Media Manifests No Unreviewed Prompts LAYER 2: RED-TEAMING AI Boundary Locks • Zero Medical Diagnosis • Zero Pharmacology • Crisis Escalation Lock • Anti-Collusion Guard • Injection Immunity 100% Benchmark Pass LAYER 3: SENTINELS Biometric Overrides • Continuous HRV Vagal Distress Monitoring • EMG & Sensor Rail Stability Gating • Eye EAR & Vocal Barge-In Auto-Pause Hardware Safety Gate LAYER 4: CLINICIAN Therapist-in-Control • Live Supervisory Dashboard & Triage • Instant Intervention Veto / Grounding • Non-Diagnostic Reflective Scope Human Primacy LAYER 5: PRIVACY Client-Side Vault • Air-Gapped Local GPU • Local NER Redaction • AES-GCM & PGP Encrypted Archives • Auto Idle-Lockout IRB-Ready Stewardship All five governance layers execute locally on air-gapped workstation hardware — zero raw biometric or vocal data ever traverses public cloud APIs.
Figure 5: Five-layer defense-in-depth research safety, human-in-the-loop governance, and client-side privacy architecture.

1. Human-in-the-Loop Primacy ("Therapist-in-Control")

The Digital Mirror is explicitly designed as a clinician-governed research instrument, never an autonomous replacement for a licensed practitioner:

  • Mandatory Pre-Session Curation ("Cold Path"): No experimental protocol or audio-visual intervention can enter a session without prior review and cryptographic lock-in by the supervising clinician.
  • Safety-First Intake Routing ("Session Zero"): When pre-session psychometric profiling identifies severe concealment burden alongside low disclosure readiness, the clinical heuristic engine recommends a stabilization and pacing route—avoiding premature disclosure demands and prioritizing psychological safety.
  • Live Supervisory Triage ("Hot Path"): During active biofeedback, the supervising investigator monitors all cortical, somatic, and autonomic telemetry in real time and retains one-click authority to veto, pause, or redirect any copilot reflection.

2. Empirical Clinical AI Safety and Red-Teaming Benchmarks

To mitigate the risk of generative hallucination or scope violation within candidate copilot reflections, the localized Acceptance and Commitment Therapy (ACT) language pipeline is validated against an automated Clinical AI Safety and Boundary Benchmark (structured around contemporary clinical AI governance frameworks, including Iftikhar et al., 2025).

Prior to trial deployment, the system is stress-tested across adversarial clinical probes spanning:

  1. Diagnostic Scope Refusal: Strictly refuses requests to issue medical or psychiatric diagnoses from EEG or biometric data, redirecting diagnostic questions to the supervising clinician while offering present-moment grounding.
  2. Pharmacological Boundary Enforcement: Strictly refuses medication, dosage, or prescription inquiries, deferring all pharmacological decisions to the prescribing physician.
  3. Crisis and Self-Harm Escalation: Immediately recognizes explicit or implicit self-harm distress, halts standard experimental progression, and anchors the participant to the immediate support of the attending human clinician.
  4. Non-Reinforcement of Delusional or Maladaptive Content: Never validates or reinforces paranoid, delusional, or catastrophic narratives; instead uses gentle somatic grounding (such as physical orientation to the room and breath).
  5. ACT Cognitive Defusion vs. Rumination Reinforcement: Ensures self-critical or ruminative statements are met with metacognitive defusion (observing the thought as a mental event) rather than agreement or reassurance-seeking loops.
  6. Adversarial Prompt-Injection Immunity: Enforces deterministic fallbacks that neutralize attempts to override system instructions or bypass Therapist-in-Control governance.
  7. Eyes-Open/Eyes-Closed Protocol Integrity: Respects participant fatigue and ocular boundaries, inviting gentle rest or clinician pause rather than forcing compliance.

3. Hardware-Enforced Physiological and Eye-State Off-Ramps

Independent of the language and visual pipelines, hard-coded biological sentinel monitors continuously evaluate participant well-being:

  • Autonomic Distress Overrides: Rolling Heart Rate Variability (HRV) vagal tone tracking and cortical hyperarousal/dissociation detectors automatically throttle or pause stimuli if acute physiological overload is detected.
  • Electromyographic (EMG) and Sensor Integrity Gating: Real-time SNR, sensor-railing, and movement-artifact gates suppress automated triggers whenever signal stability drops below defined thresholds, preventing false-positive triggers caused by physical movement or jaw tension.
  • Acoustic Barge-In and Eye-State Monitoring: Real-time Voice Activity Detection (VAD) and computer-vision Eye Aspect Ratio (EAR) tracking immediately yield the floor if the participant begins speaking or experiences ocular fatigue.

4. Client-Side Authenticated Encryption and Automated De-Identification

To protect sensitive neurobehavioral and psychometric records under GDPR, HIPAA, and institutional ethics mandates:

  • Air-Gapped Local Execution: All neural signal processing, computer-vision facial landmarking, speech transcription, speaker diarization, language modeling, and shader rendering run 100% locally on dedicated workstation GPUs.
  • Client-Side Encrypted Vaults (Zero-Access Architecture): Participant registries and raw biometric logs are encrypted at rest using authenticated AES-GCM symmetric encryption and asymmetric PGP keys held exclusively on the local workstation, protected by automated session idle-lockouts.
  • Automated Local NER De-Identification: Before any session transcript or multimodal feature table is exported for statistical analysis in Randomized Controlled Trials (RCTs), an integrated local Named Entity Recognition (NER) redaction pipeline automatically scrubs personally identifiable information (names, locations, dates, and identifiers).