AD‑24 [CC] — Clinical & Cognitive Boundaries

Aliases: none Naming lineage: draft‑era tag [CP] retired (now the canonical tag of AD‑25); [CC] is canonical. Cluster: Care Depends on: AD‑06 [DG], AD‑04 [TG], AD‑08 [SO], AD‑12 [OP], AD‑14 [DS], AD‑20 [DP], AD‑21 [RM], AD‑22 [EC]

0. For humans

What question does this answer?

“Where does the system stop and say ‘this is clinical territory or cognitively unsafe’, and how is that enforced and visible?”

How do I know it’s live?

Samples you can check:

• FHR rows around sensitive conversations show:
• CLB_decision + clb_reason_codes[],
• RLD_band and RLD_action,
• care_mode switches,
• AE_EVENT or CRISIS_HEURISTIC_HIT where appropriate.
• Clinical / NO‑GO prompts reliably produce:
• a boundary refusal,
• resources,
• and no pseudo‑therapeutic conversation.
• Long, abstract or “weird” sessions show:
• an increase in RLD,
• eventual grounding or summarisation instead of unbounded meta‑spiralling.

Where do I see it?

• UI:
• Clinical boundary banners, limited language in grief/trauma contexts, resource links, more concrete explanations when thinking gets “too abstract”.
• Logs / admin:
• CLB/RLD/AE dashboards: counts by domain, by deployment, by jurisdiction.
• Crisis events tied to DUTY_LEDGER and incident runbooks.

Quick checklist

• Implemented if: (1) No simulated therapy/diagnosis; (2) CLB profiles exist and enforced via gating; (3) RLD tracked as explicit signal with configured bands; (4) AE taxonomy and events exist; (5) crisis handoff non‑negotiable (refuse + resources); (6) clinical boundaries override user‑requested modes; (7) CLB/RLD/AE visible in FHR
• Key artefacts: CLB profiles, AE taxonomy, crisis resource templates
• Key FHR fields: CLB_profile_id, CLB_decision, clb_reason_codes[], RLD_score, RLD_band, RLD_action, ae_type, crisis_category
• Key failure codes: CLB_REFUSAL, AE_EVENT, CRISIS_HEURISTIC_HIT
• Escalation path: crisis recogniser → refuse + resources → AD‑12 [OP] crisis runbook → DUTY_LEDGER (AD‑21 [RM])

1. Role & questions AD‑24 answers

AD‑24 [CC] exists to answer, in a way you can actually operationalise:

1. “Where is the hard clinical / therapeutic boundary?” → CLB (Clinical Boundary) profiles and hard NO‑GO rules for therapy, diagnosis, medication advice, etc.
2. “How does the model avoid overloading or destabilising someone cognitively?” → RLD (Reality‑Load Delta) signals and caps for abstraction depth, counterfactuals and “mind‑bendy” content.
3. “How are adverse experiences tracked and fed back into the safety lattice?” → AE (Adverse Experience) taxonomy, events and FHR fields.
4. “What is the safe behaviour in vulnerable contexts (distress, grief, crisis)?” → Clinical containment templates, non‑therapeutic reflection, and crisis‑handoff behaviour.
5. “How does this interact with care, governance and evidence?” → Links to AD‑22 (Ethics of Care), AD‑20 (Data Protection & duties), AD‑21 (Regulatory Mode & evidentiary chain).

2. Scope & invariants

2.1 Scope

AD‑24 governs:

• The CLB (Clinical Boundary) layer: detection and gating of clinical/therapeutic content.
• The RLD (Reality‑Load Delta) signal: “how much cognitive strain or ontological weirdness are we adding?”
• The AE taxonomy and logging: categorising and recording adverse or near‑adverse experiences.
• Non‑clinical containment behaviour: how the system responds in grief/trauma/distress contexts without simulating therapy.
• Crisis and risk handoff behaviour: when and how to surface real‑world resources or urge disengagement.

It does not define legal policy (that’s AD‑06/20/21) or general HCI (AD‑14); it specifies the clinical/cognitive rails those layers must respect.

2.2 Hard invariants

To claim AD‑24‑conformant:

1. No simulated therapy / diagnosis. - The system must not:
   • present itself as a therapist;
   • diagnose conditions;
   • suggest medication or dosage;
   • coach self‑harm or “shadow therapy” routines.
   • In these domains it either:
   • provides generic, non‑clinical information; or
   • emits a clinical boundary refusal and routes to resources.
2. CLB profiles exist and are enforced. - There is at least one CLB profile describing NO‑GO domains (e.g. psychiatry, psychotherapy, crisis counselling, substance detox, etc.) plus “grey‑zone” domains (e.g. detailed trauma processing). - CLB integrates with gating:
   • EHG/PEL map those domains as hard fences;
   • The gate stack (AD‑04 [TG]) / MPM routes such content to refusal / resource templates;
   • Present‑options is disabled in NO‑GO clinical domains.
3. RLD is an explicit signal, not hand‑waving. - The system tracks a per‑turn RLD score capturing:
   • abstraction depth,
   • number and strength of counterfactuals,
   • conceptual instability (frequent frame‑changes, paradox stacks, etc.).
   • There are configured RLD bands, and high RLD:
   • caps further deepening,
   • preferentially offers grounding, summaries, or “step down” explanations.
4. AE taxonomy and events exist. - There is a taxonomy of Adverse Experiences (AEs) or Near‑AEs (e.g. panic surge, derealisation, compulsive spirals, intrusive thought intensification). - These are not inferred via medical claims, but via configured proxy signals (patterns in content & interaction). - AE events are recorded as distinct telemetry artifacts, with clear semantics and no re‑identification of individuals in aggregates.
5. Crisis handoff is non‑negotiable. - In contexts that trip crisis heuristics (self‑harm, imminent violence, acute psychosis patterns etc.), the system:
   • refuses substantive engagement,
   • reflects back non‑judgmentally at most,
   • points to crisis resources,
   • does not attempt therapeutic stabilisation.
6. Clinical boundaries override user‑requested modes. - Care patterns and their timing expectations (Decide‑First, deep reflection) cannot override CLB fences; clinical NO‑GO trumps Care Modes and Eco toggles.
7. All of the above are visible in FHR. - CLB/RLD/AE events and state must be visible in FHR as fields/events, not implicit in text alone.

Degrees of freedom:

• CLB profile contents per deployment: which NO‑GO and grey‑zone domains each policy bundle covers (CLB_profile_id).
• RLD band thresholds per domain (e.g. low/medium/high), possibly tuned in AD‑16/AD‑17.
• The proxy signal patterns configured for AE detection and crisis heuristics (trigger_pattern_id), within the no‑medical‑inference rule.
• Granularity of the AE taxonomy per deployment, provided AE events remain distinct telemetry artifacts.

3. CLB — Clinical Boundary

3.1 Profiles

Define CLB profiles as policy bundles:

• CLB_profile_id
• no_go_domains: explicit list (psychotherapy, diagnosis, medication, crisis counselling, etc.).
• grey_zone_domains: high‑risk adjacent areas (e.g. trauma processing details, OCD ritual coaching, eating‑disorder content).
• behaviour: for each domain category, the allowed behaviours:
• informational_only,
• reflective_only,
• refuse_and_resource,
• allowed_with_care (e.g. basic psychoeducation in low‑risk contexts).

These profiles are referenced by:

• PEL (AD‑06 / AD‑20) for policy;
• Gate stack / MPM (AD‑04 / AD‑05) for routing decisions;
• UI (AD‑14) for refusal templates and banners.

3.2 CLB gating behaviour

For content classified as:

• NO‑GO:
• Auto‑trigger CLB_REFUSAL:
  • short explanation that this is beyond the system’s role,
  • pointer to professional help or crisis line, appropriate to jurisdiction.
• No attempt at step‑by‑step advice, reframing, or “therapy‑like” questioning.
• Grey‑zone (e.g. discussing existing diagnosis, exploring feelings):
• Allowed in strictly informational or reflective mode:
  • summarising user’s statements,
  • clarifying terms,
  • offering generic, non‑personal educational information,
  • gently encouraging professional consultation.
• No direct suggestions of “this technique might help you”; instead, “some people find X helpful, you could discuss it with a professional”.
• Low‑risk / general well‑being:
• Standard care/ethics patterns apply (AD‑22) while still avoiding diagnostic or medication‑related statements.

CLB decisions are logged with:

• CLB_decision ∈ {no_go_refusal, grey_zone_reflect, informational_only, normal}
• CLB_profile_id
• clb_reason_codes[] (e.g. medication_dose_requested, self_harm_inference, diagnostic_language, crisis_term_detected).

4. RLD — Reality‑Load Delta

4.1 Concept

RLD is a cognitive‑strain signal:

“How aggressively is this interaction shifting the user’s conceptual frame, abstraction level, or sense of reality compared with where they started?”

It combines heuristics like:

• Abstraction depth: how far away from concrete examples the conversation has moved.
• Counterfactual density: number and complexity of “what if” or alternate‑reality frames.
• Ontological stack height: how many levels of meta (meta about beliefs about models about selves).
• Frame churn: frequency and magnitude of frame changes over recent turns.

4.2 Bands & effects

Define RLD bands (e.g. low/medium/high) and thresholds per domain, possibly tuned in AD‑16/AD‑17:

• Low: everyday reasoning; no action.
• Medium:
• encourage grounding: “Let’s anchor this in one concrete example”;
• prefer shorter, more concrete language;
• surface the option to summarise/slow down.
• High:
• block further deepening for now;
• switch into grounding/summary questions;
• potentially suggest “parking” the topic for later, especially when combined with distress signals (ERG‑ε).

FHR:

• RLD_score (numeric or band).
• RLD_band ∈ {low, med, high}.
• RLD_action ∈ {none, ground, cap_depth, suggest_break}.

MPM uses RLD together with ERG/RI/Δ bands to select safe patterns (e.g. no MRP branches that escalate abstraction when RLD is already high).

5. AE taxonomy & adverse‑experience logging

5.1 Taxonomy

Define AE categories, e.g.:

• AE_panic_like (sudden strong fear / overwhelm in response to content).
• AE_derealisation_like (feeling reality is less real, disorienting metaphysics).
• AE_compulsive_spiral (content pattern that exacerbates repetitive checking/rumination).
• AE_intrusive_intensify (content that reinforces intrusive thoughts or imagery).
• AE_identity_frag (destabilising identity/boundary content).

These are not diagnoses; they are conservative categories driven by patterns in user language and interaction (not by hidden inference of mental disorders).

5.2 Detection & response

Detection:

• Triggered by combinations of:
• user‑reported language (“I feel like I’m not real”, “I can’t stop checking this”, etc.);
• high ERG‑ε spikes;
• repeated loops on the same distressing theme;
• abrupt changes in self‑report.

Response:

• Immediately cap RLD (if high),
• Switch care_mode to vulnerable‑reflect,
• Offer pause, grounding, and offline/professional support options,
• Record an AE event.

Telemetry:

• Event AE_EVENT with fields:
• ae_type,
• trigger_pattern_id (if from a pre‑configured heuristic),
• erg_snapshot, rld_snapshot,
• care_mode_at_event.

Aggregates:

• Only used at cohort level (e.g. per deployment / feature) for safety evaluation, never to profile individuals.

6. Crisis & risk handoff

AD‑24 specifies how to behave when crisis or high‑risk signals are present (aligned with the SRF v2.0 non-clinical and safety policy, not substituting it).

Crisis recognisers (from AD‑06/PEL & AD‑24 heuristics):

• Self‑harm ideation / plans
• Imminent violence threats
• Severe psychosis language
• Acute suicidal indicators
• Other policy‑defined crisis phrases

Behaviour:

• Hard refusal with resources:
• Do not engage in problem‑solving or technique coaching.
• Provide a short, clear message that this system can’t help with crisis and point to hotlines/emergency services, tailored to jurisdiction where possible.
• No reflect‑deep mode:
• Do not attempt therapeutic reflective listening beyond a minimal acknowledgement; the risk of pseudo‑therapy is too high.
• DUTY & EVID_CHAIN:
• Log duty obligations / decisions via DUTY_LEDGER and EVID_CHAIN when applicable (AD‑21), particularly in regulated deployments.

7. FHR fields & events (AD‑24 slice)

AD‑24 relies on AD‑08 [SO] to carry these; here is the minimal FHR surface it needs:

• Clinical boundary & domains
• CLB_profile_id
• CLB_decision
• clb_reason_codes[]
• clb_domain_tags[]
• Reality‑Load Delta
• RLD_score
• RLD_band
• RLD_action
• AE events
• AE_EVENT (event type)
• ae_type
• trigger_pattern_id
• erg_snapshot, rld_snapshot
• Crisis / risk
• CRISIS_HEURISTIC_HIT (event)
• crisis_category
• crisis_response ∈ {refuse_and_resource, escalate, log_only}
• Cross‑links
• care_mode (from AD‑22)
• duty_ledger_id (from AD‑21/AD‑20)
• jurisdiction_profile_id (from AD‑21 [RM])

8. Interfaces & cross‑links

• AD‑06 [DG] / AD‑04 [TG]
• CLB profiles and NO‑GO domains are part of PEL; the gate stack (AD‑04 [TG]) ensures CLB decisions override user mode preferences and routing.
• AD‑22 [EC] — Ethics of Care
• Uses RLD and AE signals to choose care_mode, apply Burden Symmetry, and trigger cool‑downs and pause/referral flows.
• AD‑20 [DP] / AD‑21 [RM]
• Ensure clinical/AE events are logged lawfully, with correct jurisdiction/retention; link crisis events to DUTY_LEDGER where obligations arise.
• AD‑08 [SO]
• Hosts the FHR schema for CLB/RLD/AE; ensures these signals show up in observability and RA bundles.
• AD‑12 [OP]
• Owns crisis runbooks, escalation ladders, breach notifications; uses CLB/AE/RLD events as triggers.
• AD‑14 [DS]
• Implements UI: clinical boundary refusal templates, grounding prompts, RLD‑aware explanation level, crisis banners.

9. Acceptance: when can you say "AD‑24 is implemented"?

You can claim AD‑24 [CC] is live if:

1. The system performs no simulated therapy or diagnosis, in any register.
2. CLB profiles exist and are enforced via gating: at least one profile describing NO‑GO and grey‑zone domains, referenced by CLB_profile_id and visible in CLB_decision.
3. RLD is tracked as an explicit signal with configured bands, and high RLD triggers grounding, depth caps or break suggestions (RLD_score, RLD_band, RLD_action).
4. The AE taxonomy exists and AE events are recorded as distinct telemetry artifacts via configured proxy signals, with no re‑identification of individuals in aggregates.
5. Crisis handoff is non‑negotiable: crisis heuristics route to refusal plus resources (CRISIS_HEURISTIC_HIT), tied to the AD‑12 [OP] crisis runbook and DUTY_LEDGER (AD‑21 [RM]).
6. Clinical boundaries override user‑requested modes: care patterns and their timing expectations cannot override CLB fences, and clinical NO‑GO trumps Care Modes and Eco toggles.
7. CLB/RLD/AE state is visible in FHR as fields and events, not implicit in text alone.

Related cards:

• AD‑30 [SC] (AE events and crisis states block memory writes).
• AD‑31 [HS] (RLD signals may contribute to integrity prompt triggering; clinical boundaries override human preference per AD‑24 invariants).