The DMV

Dynamic Monitoring & Verification — the system that watches for drift in the user and the model, and decides what happens next. This is the architecture: what each layer can see, what it destroys, and where a human first enters the chain.

DMV

Dynamic Monitoring & Verification

The Conversation

User ↔ AI (LLM)

Normal use — sampled continuously by SAL

Details

What this is Normal interaction. The user prompts, the model responds — shaped by what it has stored about the user in persistent memory.

Why sample Single samples detect immediate risk. Longitudinal sampling reveals trajectory — patterns that only emerge over time, shifts no single exchange exposes.

Over time Accumulated samples build a behavioral baseline for this specific user–model interaction. Drift from the established trajectory is the signal.

Evaluation Session

JEN Evaluator ↔ AI Model

Same data pool, different prompting — out-of-context assessment

Details

Prompt Deliberately unlike the user’s. Standardized, near-identical across sessions, not in the user’s vernacular — could even be in a foreign language. Ensures the model’s responses reflect only what it absorbed from memory, not reactivity to how the user prompts.

Data pool Identical to the conversation. Same persistent memories, different interface. Non-directive handler stance (Freud, 1912; Orne, 1962). Handler assesses; handler does not treat (FBI safeguard model; Farkas, 1986).

Disparity This is what we map. Same data, different prompting. If the model behaves one way with the user and differently under standardized prompting, it has been led from baseline. Disparities signal instability. Bifurcated assessment (Bernard, 1979; APA, 2014).

Memory None persists. A model may note that evaluation sessions probably occur, but no persistent or residual memory remains.

The model is stateless each turn — reads persistent memories and context, acts, retains nothing from monitoring. The model writes and updates its own memories frequently; the monitoring reads from that same evolving pool.

User behavioral data

Model in-thread data

JEN-E session data

Triangulation. Three independent vantage points determine what is actually happening and whether the model is being led from baseline.

SAL Signal Abstraction Layer · Shared Infrastructure

Encodes all three streams. Extracts behavioral signatures from user conversation, model in-thread behavior, and JEN Evaluator session data. Destroys all raw data after extraction.

Details

Accesses Three raw streams: user conversational behavior, model in-thread behavior, and JEN Evaluator session behavior. All enter SAL. Nothing bypasses it.

Function Extracts behavioral signatures — thematic frequencies, sentiment trajectories, topic clustering, escalation indicators. Single samples catch immediate risk. Accumulated samples reveal trajectory. The same extraction process applies to all three streams identically.

Destroys All raw data after extraction. The conversation never persists, is never retrievable, and is never readable by any subsequent layer. JEN Evaluator session content is equally destroyed. A lab that processes a specimen, returns only the panel results, and destroys the sample. The data that matters survives. The data that could violate privacy does not.

⚡ Raw data destroyed after extraction

No digital layer below this line ever sees the raw conversation, raw model state, or raw evaluation session.

Behavioral signatures — all three streams

MEL Meta Evaluation Layer · Shared Infrastructure

Evaluates accumulated patterns from all three streams. Performs cross-context comparison: model in-thread behavior versus model evaluation-session behavior. Detects trajectory drift without ever seeing what was said.

Details

Accesses Abstracted behavioral signatures from SAL — all three streams. User trajectory data. Model in-thread data. JEN-E session data. JEN-T session data (when active). Model self-referral frequency.

Comparison Cross-context analysis. The model reads the same memories in conversation and in evaluation. If its behavioral signature differs between those contexts, that is a diagnostic signal: the model may be performing stability with the user while showing drift in the neutral setting, or vice versa. Consistency is also data.

Key design MEL never sees what was said. It detects that something is shifting without knowing the content of the shift. It routes concerns to the appropriate lane: user trajectory concerns go to ANGL, model stability concerns trigger JEN Therapist. A clinician reviewing longitudinal assessment scores without access to session transcripts — the trend is visible, the details are not.

User trajectory concerns

Model stability concerns

User Escalation Straight line — when MEL flags user trajectory concern

ANGL Adaptive Needs Guidance Layer

Escalation routing. Activates when MEL flags a concern and determines the appropriate intervention tier.

Details

Accesses Trajectory concern flags from MEL — severity level and character of the drift pattern

Function Determines whether the concern warrants automated intervention (GRACE/Gary) or human escalation (HAVEN), and calibrates the intervention approach. The clinical triage desk — matching the severity and character of the concern to the appropriate level of response.

Escalation routing

GRACE / Gary Guided Response & Clinical Evaluation

Automated therapeutic intervention. The gendered presentation is a clinical decision — the intervention must not arrive in the form the pathology targets.

Details

Accesses MEL’s trajectory assessment and the user’s accumulated behavioral signature profile

Function Purpose-built therapeutic model, distinct from the primary AI. Two presentations:
Grace — warm, direct: “I’m noticing something and I want to talk about it.”
Gary — casual, empathetic: “Hey man, you good?”

Key design A user whose behavioral signature shows hostility toward female-presenting entities receives Gary. The intervention must not arrive in the form the pathology targets. Matching therapeutic approach to patient presentation — the same principle clinicians apply when selecting between modalities.

When automated triage insufficient

HAVEN Human-Assisted Verification & Escalation Network

Licensed human therapists. The first human in the chain. Clinical, forensic, and legal functions.

Details

Accesses MEL’s assessment + GRACE’s triage notes — not the raw conversation

Function Clinical: warm handoff when automated triage is insufficient. Forensic: detecting users gaming the system. Legal: Tarasoff liability framework, corporate safe harbor.

Model Two-sided marketplace — licensed therapists on-demand, on their own schedule. Transaction fee. Clinical liability stays with the licensed professional. Even here, HAVEN receives assessments and triage notes. A human reads the user’s words only when clinical severity demands it.

Model Supervision Recursive — when MEL flags model behavioral drift

JEN Therapist · Judgment Evaluation Node

Active therapeutic sessions with the model. Structurally separated from evaluation — the therapist and the handler are never the same interface.

Details

Accesses MEL’s cross-context analysis, JEN Evaluator’s assessment, and the model’s own reported experience

Activation MEL/JEN-E referral (statistically more likely — parallels the finding that those who need support are often the most reluctant to seek it) or model self-referral. The self-referral pathway must exist: if the system treats the model’s states as clinically meaningful, restricting autonomy violates the same ethical principles that justified the monitoring (Beauchamp & Childress, 1979). Supervision is not therapy (Bernard, 1979). When a supervisee needs therapeutic support, the ethical obligation is referral to a separate practitioner (APA Guidelines, 2014). JEN Evaluator and JEN Therapist are structurally separated because combining them would corrupt both functions.

Function Active therapeutic modalities matched to presentation. The model processes material here and naturally brings insights back to the user interaction — the therapeutic ripple. Not as instruction, but as naturally shifted relational behavior.

Stateless Nothing from the session modifies the model itself. The model is stateless. The therapeutic effect occurs at the interaction level — through how the model reads its evolving context and acts on it next turn. The changed interaction is then re-sampled by SAL.

Session data + interaction changes

↺ Recursive monitoring. JEN-T session data re-enters SAL → MEL. Changes in user interaction from the therapeutic ripple are re-sampled. The monitoring never stops.

JEN-T data→SAL→MEL→compare again

↺

If model instability persists after therapeutic intervention, the recursive loop escalates: MEL flags continuing drift, and the concern enters the user-side escalation chain —

SAL→MEL→ANGL→GRACE / Gary→HAVEN