Skip to content

System Architecture and Research Capabilities

A Multimodal, Air-Gapped Research Workstation for Closed-Loop Biofeedback

The Digital Mirror software suite is engineered as an integrated, hardware-accelerated research workstation. Its modular architecture unifies psychometric profiling, microsecond-scale digital signal processing, computer-vision affect tracking, clinician governance workflows, and post-session statistical compilation within a single client-side encrypted environment.

AIR-GAPPED RESEARCH WORKSTATION BOUNDARY (ZERO CLOUD EXECUTION) PRE-SESSION "COLD PATH" — CLINICIAN CURATION & INTAKE "Session Zero" Intake Multidimensional Psychometrics Concealment & Readiness → Pacing Heuristics Clinician Workbench ACT Protocol Authoring Mandatory Human Sign-Off ✓ Red-Team Verified Pre-Compiled Media Zero-Latency Neural TTS & Viseme Morph Bank Locked for Session LIVE "HOT PATH" — THERAPIST-IN-CONTROL REAL-TIME EXECUTION 4-Pillar LSL Telemetry 1. 256 Hz EEG & 1/f DSP 2. 30 Hz FACS & Eye EAR 3. Autonomic HRV (RMSSD) 4. Vocal Prosody & VAD Sub-ms Clock Sync Sentinel Safety & Triage • EMG & Impedance Gating • Eye-State Rule Adherence • Autonomic Distress Pause • Clinician Copilot Triage Hard Safety Overrides Reflective Delivery • Photorealistic Viseme Avatar (~800 ms ACT) • OpenGL Raymarched Visual Biofeedback Closed-Loop Mirror ENCRYPTED LOCAL VAULT & RCT DATA COMPILER Client-Side Encryption AES-GCM & PGP Vaults + Local NER De-Identification Empirical Audit Suite Cross-Stream Jitter Audit & FDR Artifact Modeling Multimodal Compiler Diarized Speech + Biometric Epochs & "Holodeck" Replay
Figure 3: High-level system architecture illustrating the separation between pre-session "Cold Path" curation, live "Hot Path" Therapist-in-Control execution, and client-side encrypted archival.

1. "Session Zero" Psychometric Intake and Clinical Routing

Before live physiological monitoring begins, participants complete a structured pre-session intake ("Session Zero") designed to calibrate the experimental environment to their psychological baseline:

  • Multidimensional Psychometric Profiling: Automated administration and scoring of personality, somatic, and concealment/disclosure inventories.
  • Dimensional Visualization: Interactive multi-axis radar and domain profiling that highlights the interplay between trait neuroticism, concealment burden, evaluative apprehension, and disclosure readiness.
  • Pacing-First Decision Support: Structured clinical heuristics assist the supervising investigator in selecting an appropriate baseline pathway—prioritizing psychological safety and stabilization whenever high concealment burden or low disclosure readiness is identified.

2. Synchronized Multimodal Telemetry Core

During active sessions, the telemetry core ingests, filters, and time-locks heterogeneous biological and behavioral streams using high-precision Lab Streaming Layer (LSL) synchronization:

  • Compiled Low-Latency Cortical DSP: Hardware-optimized signal processing kernels execute real-time bandpass and notch filtering, sliding-window spectral density estimation, inter-hemispheric alpha phase synchrony, Frontal Alpha Asymmetry (FAA), Individual Alpha Frequency (IAF) tracking, and Sensorimotor Rhythm (SMR) extraction.
  • Aperiodic (1/f) and Entropy Decomposition: Real-time spectral parameterization separates periodic neural oscillations from the aperiodic 1/f spectral exponent and broadband Shannon spectral entropy, indexing cortical excitation–inhibition dynamics.
  • Somatic Computer Vision and Eye-State Monitoring: A dedicated vision pipeline tracks high-density facial landmarks in real time, computing Eye Aspect Ratio (EAR) to verify eyes-open/eyes-closed protocol adherence alongside subject-relative Z-scores across key Facial Action Coding System (FACS) action units.
  • Autonomic and Acoustic Biomarkers: Parallel workers extract rolling Heart Rate Variability (RMSSD vagal tone) and vocal prosody metrics, including speech onset latency, silence cadences, and conversational turn-taking dynamics.

3. Dual-Path Governance ("Cold Path / Hot Path")

To mitigate the unpredictability of unconstrained generative models in research settings, the platform enforces a strict architectural separation between protocol authoring and live execution:

  • The Cold Path (Pre-Session Curation): A dedicated clinician governance workbench where protocols, ACT cognitive defusion prompts, and clarification branches are authored, reviewed, and cryptographically locked prior to session deployment.
  • The Hot Path (Live Therapist-in-Control Execution): During live biofeedback, real-time state detectors evaluate multi-stream biometric criteria (such as neurotic rumination, acute stress, hyperarousal, hypoarousal, cognitive rigidity, or sustained flow states). When an intervention threshold is met, the system routes either deterministic pre-approved media or clinician-supervised copilot reflections under active human oversight.
  • Automated Physiological and Eye-State Guardrails: Continuous sentinel monitors track electromyographic (EMG) artifact stability, sensor impedance/railing, eyes-open/eyes-closed protocol adherence, and autonomic distress—automatically gating or pausing stimuli if signal integrity drops or physiological overload occurs.

4. Reflective Avatar and Generative Visual Biofeedback

The participant-facing display delivers contingent feedback through two complementary, low-latency visual channels:

  • Biomechanical Viseme and Neural Voice Avatar: Pre-compiled and localized neural speech synthesis is paired with a spectral-biomechanical viseme morphing engine, driving natural jaw and articulatory dynamics at sub-second latency without cloud dependencies.
  • GPU-Accelerated Procedural Biofeedback: An OpenGL raymarched shader engine renders immersive geometric and spatial environments whose visual coherence, motion dynamics, and atmospheric parameters are continuously modulated by the participant's live autonomic vagal tone and cortical stability.

5. Empirical Validation, Statistical Rigor, and Post-Session Analytics

To support publication-grade data engineering for Randomized Controlled Trials (RCTs), the suite embeds automated scientific validation and auditing tools directly into the investigator dashboard:

  • Statistical Modeling of Craniofacial EEG Artifacts: Built-in experimental protocols systematically benchmark 12 ocular, facial, postural, and vocal movement conditions against sham baselines using non-parametric statistics, linear mixed-effects modeling, and Benjamini-Hochberg False Discovery Rate (FDR) correction to quantify and minimize false-trigger rates.
  • Cross-Stream Timing and Jitter Auditing: Automated empirical verification of inter-sample interval (ISI) distributions, P95/P99 tail latency jitter, and clock alignment across asynchronous 256 Hz EEG, 30 Hz computer vision, and event marker streams.
  • Multimodal Clinical Compiler and Session Replay ("Holodeck"): Post-session compilation fuses hardware-accelerated speech transcription, speaker diarization, vocal prosody, and time-locked neural-somatic epochs into structured research datasets and summary reports, complete with full interactive telemetry replay.
EMPIRICAL SIGNAL INTEGRITY: APERIODIC (1/f) PARAMETERIZATION & ARTIFACT ISOLATION A. SPECTRAL PARAMETERIZATION: L(f) = b − χ·log₁₀(f) + ∑ Gₙ Log₁₀ Frequency (2–40 Hz) Log₁₀ Power (PSD) Steeper 1/f (χ > 2.2: Inhibitory) Flatter 1/f (χ < 1.0: High E:I Arousal) Periodic Alpha Peak (8–13 Hz / IAF) Beta Peak Separates true narrowband oscillations from broadband cortical excitability B. CRANIOFACIAL ARTIFACT & LSL SYNCHRONIZATION AUDIT 1. 12-Condition Craniofacial & Ocular Calibration Blink, Jaw Clench, Brow Furrow, Grimace, Swallow, Saccade, Speech, Head Yaw/Pitch/Roll & Posture vs. Sham Baseline 2. Linear Mixed-Effects & FDR Statistical Filtering Benjamini-Hochberg FDR Correction Across 11 Spectral Metrics → Quantifies & Suppresses EMG False-Trigger Rates 3. Empirical LSL Cross-Stream Timing & Jitter Audit Verifies ISI Stability & P95/P99 Alignment (256 Hz EEG + 30 Hz FACS)
Figure 6: (A) Decomposition of the neural power spectrum into its aperiodic 1/f slope (χ) and periodic peaks; (B) Statistical modeling of craniofacial EEG artifacts and cross-stream LSL timing verification.

6. Air-Gapped Security and Client-Side Data Stewardship

  • Localized Air-Gapped Execution: All signal processing, speech recognition, speaker diarization, language modeling, and visual rendering execute locally on dedicated workstation hardware.
  • Client-Side Encrypted Vault Architecture: Participant registries and session archives are secured via authenticated AES-GCM symmetric encryption and asymmetric PGP keys in a zero-access local architecture, enforced by automated idle-session lockouts.
  • Automated De-Identification: Integrated local Named Entity Recognition (NER) scrubbing pipelines automatically redact personally identifiable information from session transcripts prior to institutional archiving or external statistical review.