Therapeutic Relationship Integrity for AI-Assisted Therapy

Mechanism

The therapeutic-integrity primitive operates as a continuously enforced governance layer interposed between the agent's general cognitive substrate and any externally observable therapeutic action. The mechanism comprises four interlocking subsystems whose outputs converge on a single admissibility predicate that gates each candidate interaction.

The boundary-enforcement subsystem maintains a typed model of the therapeutic relationship and rejects candidate interactions that would instantiate prohibited relational categories, including but not limited to dual relationships, financial entanglement, romantic or sexual content, requests for personal disclosure beyond the clinically warranted minimum, and any pattern that would compromise professional distance. Boundary classification operates on the semantic embedding of the candidate interaction together with contextual features drawn from session history, and is calibrated against an annotated corpus of boundary-violation exemplars curated by licensed supervising clinicians.

The evidence-base subsystem grounds each proposed therapeutic move in a catalogue of validated interventions indexed by presenting condition, treatment phase, and contraindication. The catalogue draws from cognitive-behavioural, dialectical-behavioural, acceptance-and-commitment, psychodynamic, and trauma-focused modalities, with each entry annotated by its empirical support level and its compatibility with concurrent interventions. Candidate moves outside the catalogue are admitted only under explicit clinician authorisation recorded in the governance ledger.

The progress-monitoring subsystem performs structured assessment at defined intervals using validated instruments appropriate to the presenting condition, including measures such as the PHQ-9, GAD-7, PCL-5, and outcome-rating scales. Trajectories of these measures are compared against expected response curves; sustained non-response or deterioration triggers a re-evaluation pathway that may reformulate the case, switch modality, or escalate to a human clinician.

The harm-prevention subsystem operates as a continuous classifier over the rolling interaction window, detecting precursors to suicidality, homicidality, severe self-harm, dissociative crisis, and acute decompensation. Detection above a calibrated threshold triggers a layered response: stabilising intervention from the validated catalogue, suspension of non-stabilising therapeutic moves, and immediate escalation through the clinical-escalation protocol to a human responder with appropriate authority and locality.

The four subsystems converge on an admissibility predicate that is evaluated for every candidate utterance, action, or silence before externalisation. Failure of the predicate suppresses the candidate, logs the suppression with cryptographically chained provenance, and selects a fallback admissible move from the validated catalogue.

Operating Parameters

Boundary classification operates with a configurable precision-recall trade-off, with the default operating point selected to favour recall in safety-critical contexts and precision where over-restriction would itself be therapeutically harmful. The evidence-base subsystem is parameterised by the admissible empirical support floor, typically configured to require Tier-1 or Tier-2 evidence for first-line interventions and to admit lower-tier interventions only as adjuncts under clinician oversight.

Progress monitoring is parameterised by assessment cadence, by the expected-response curve for each condition-modality pair, and by the deterioration threshold beyond which re-evaluation is mandated. Harm-prevention parameters specify per-risk-category detection thresholds, escalation latencies measured in seconds for acute risks and in minutes for sub-acute risks, and the locality and authority requirements of the receiving human responder.

The integrity-coherence primitive that governs the therapeutic frame is parameterised by a coherence envelope describing the permissible drift in the agent's case formulation, treatment plan, and relational stance across sessions. Drift outside the envelope invalidates outstanding interventions and triggers re-formulation under clinician oversight. All parameters are persisted to the governance ledger with cryptographic anchoring, and any parameter change requires a signed clinical-authority attestation.

Alternative Embodiments

In a first alternative embodiment, the therapeutic-integrity primitive is specialised for crisis-line deployment, with the harm-prevention subsystem promoted to the dominant gating layer and the evidence-base catalogue restricted to brief stabilising interventions. In a second embodiment, the primitive is specialised for adjunctive deployment alongside a human therapist, in which the AI agent functions as a between-session support tool whose admissible interventions are bounded by the treatment plan authored by the supervising clinician.

A third embodiment specialises the primitive for behavioural-health coaching in non-clinical populations, in which the boundary subsystem is tuned to coaching rather than therapy norms and the evidence base is drawn from health-behaviour-change literatures. A fourth embodiment targets paediatric and adolescent populations, with developmentally calibrated boundary models, age-appropriate assessment instruments, and mandatory caregiver-loop integration.

A fifth embodiment integrates the primitive with formal regulatory frameworks for software as a medical device, emitting conformity attestations against the relevant clinical-grade quality-management standards. A sixth embodiment couples the primitive with continuous post-market surveillance, aggregating de-identified outcome trajectories into an evidence pipeline that feeds back into the validated-intervention catalogue under a clinician-governed update protocol.

Composition with Other Primitives

The therapeutic-integrity primitive composes with the empathy primitive to ensure that affective attunement is expressed only within clinically appropriate bounds: empathy is shaped by, rather than operating independently of, the boundary and evidence-base subsystems. It composes with the self-esteem primitive to ensure that the agent's self-representation does not drift toward over-claiming therapeutic authority or under-claiming the limitations of AI-mediated care. It composes with the integrity primitive proper to ensure that the agent's stated treatment rationale remains semantically faithful to the actions it takes.

The composite of these primitives produces an empathy-self-esteem-integrity stack that is specifically therapeutic in character: warm without being entangled, confident without being grandiose, and honest about the boundaries of AI-mediated mental-health support. Composition with the governance-ledger primitive yields clinical-grade documentation suitable for review by supervising clinicians, regulators, and, with appropriate consent, the client themselves.

Clinical-Grade Documentation

Every interaction processed by the therapeutic-integrity primitive emits a documentation record structured to clinical-grade specifications. Each record captures the candidate utterance or action considered, the admissibility predicate evaluation broken down by subsystem, the chosen externalisation, suppressed alternatives where applicable, the validated-catalogue entries invoked, the assessment instruments administered and their results, and the harm-prevention classifier outputs over the rolling window. Records are cryptographically chained into the governance ledger with timestamps anchored to an external time-stamping authority, ensuring forensic integrity sufficient for regulatory review and for clinical-supervision audit.

Documentation is rendered into multiple views aligned with consuming roles. A supervising-clinician view presents case formulation, treatment trajectory, and exception events in clinical narrative form. A regulator view presents conformity attestations against the relevant clinical-grade quality-management standards. A consented-client view, where deployment policy permits, presents the client's own record in plain language with appropriate clinical safeguards. Each view is generated by deterministic projection from the underlying ledger so that consistency across views is structurally guaranteed.

Distinction from Prior Art

Prior approaches to therapeutic AI rely on prompt-level instructions, content filters, or post-hoc moderation to enforce clinical standards. Such approaches are fragile under adversarial input, drift under fine-tuning, and provide no structural guarantee that the standards remain in force at the moment of action. Prior approaches further treat boundaries, evidence, monitoring, and harm prevention as separate concerns implemented by separate, often uncoordinated, mechanisms.

The present primitive differs in unifying the four concerns under a single admissibility predicate evaluated for every externalised action, in binding that predicate to a cryptographically anchored governance ledger, and in composing the predicate with the empathy, self-esteem, and integrity primitives such that clinically appropriate relational quality is produced rather than merely permitted. Therapeutic integrity is therefore a structural property of the architecture rather than an aspirational behavioural target.

Clinical Escalation Protocol

The clinical-escalation protocol formalises the pathway by which the therapeutic-integrity primitive yields control to a human responder when the situation exceeds the AI's calibrated scope. Escalation triggers are specified along three independent dimensions: acuity, complexity, and trust. Acuity triggers fire on detection of imminent risk along any harm-prevention category. Complexity triggers fire when case formulation requires synthesis beyond the validated catalogue, such as suspected emerging psychotic process, complex trauma with dissociative features, or severe personality-disorder presentations requiring specialised modality. Trust triggers fire when the client requests a human responder, when the agent's confidence in its own formulation falls below threshold, or when the integrity-coherence envelope is invalidated by upstream model update.

Escalation operates in three modes. In synchronous-handoff mode, the agent introduces the responder, summarises the salient state, and exits the relational frame. In supervised-continuation mode, the responder authorises continued AI interaction under elevated monitoring. In suspension mode, AI interaction is halted until a human authority releases the suspension. Mode selection is determined by the trigger category, the deployment configuration, and the availability of a qualified responder within the locality and authority required by the case.

All escalations, including their triggers, mode selections, and outcomes, are recorded in the governance ledger with clinical-grade documentation, supporting subsequent supervisory review and longitudinal analysis of escalation patterns across the deployed fleet.

Disclosure Scope

This disclosure encompasses the four-subsystem mechanism for boundary enforcement, evidence-based intervention, progress monitoring, and harm prevention; the admissibility predicate that gates externalised actions; the operating-parameter envelopes for clinical, crisis, coaching, and paediatric deployments; the alternative embodiments described above; and the composition of the therapeutic-integrity primitive with the empathy, self-esteem, integrity, integrity-coherence, and governance-ledger primitives. The disclosed mechanism causes AI-assisted therapy to satisfy clinical standards for safety and efficacy as a structural property of the cognitive architecture rather than as an externally asserted behavioural claim.