Skip to main content

Jev for routing and escalation in a counseling digital twin: viability

27 September 2026. Written in answer to the question "how viable is using TypeSafe's Jev for routing and human-escalation decisions in the counseling digital twin?"

The idea (from the 27 Sep session, 46:55–49:11)​

  1. Routing: a decision model classifies each patient message to pick the right agent.
  2. Escalation: a second decision model judges whether the counselor twin's answer is good enough to send, or whether the case goes to the human counselor with a summary and pointer. Everything is logged for the counselor's next session.

28 Sep implementation clarification. This preserves the original feasibility assessment. ../PLAYBOOK.md is the normative contract; the proxy follow-up defines routing, scoring and compatibility acceptance requirements. No runtime or clinical certification is established by this note.

Verdict​

The idea is architecturally sound and the model is a strong fit. As a hosted vendor dependency for this use case, it isn't viable today.

Why it fits​

  • Both calls are decision-shaped. Jev takes unstructured input plus a schema and returns typed choices with probabilities that the vendor claims are calibrated; target-population calibration remains unverified. It runs in one parallel pass, and TypeSafe markets it for "classify, route, score, extract, verify."
    • The route list is well under Jev's 255-option cap.
    • Escalation is a small fixed set, such as {send, send_and_flag, hold_for_clinician, crisis_protocol}.
    • Neither step needs free text, and Jev doesn't generate strings.
  • Escalation is mostly a threshold decision, and thresholds need calibrated probabilities. Jev's RLCD training targets calibration directly.
  • Latency isn't an issue. TypeSafe reports 70–500 ms, so every reply can be gated before the patient sees it.

What breaks it​

  1. Privacy architecture. Jev is proprietary and served only from TypeSafe's hosted API, with no published weights and no on-prem option. Counseling transcripts are among the most sensitive PHI, and no BAA is published. It contradicts the stated on-device design.
  2. Calibration is self-reported and not measured on this distribution. The benchmarks are "self-graded on a format the company invented" (TrueFoundry). Calibration on average says nothing about the rare tail (suicidal ideation), which is the only place that matters.
  3. "90% handled by the bot" is the wrong target. The metric that matters is the false-negative rate on escalations. Set recall floors first, using class-conditional conformal prediction, and measure the automation rate that results.
  4. Crisis handling can't depend on one probability. Run rule-based tripwires before any model and preserve an escalation across every later stage. A constrained model can still confidently pick the wrong option.
  5. Regulation. Illinois (2025) bars AI from therapeutic decision-making, and other states have similar bills. Frame the product as supervised between-session support with the counselor in the loop, and get counsel's review.

Resulting shape (later built out in this repo)​

  • Pre-generation: route the message, with crisis as a class.
  • Post-generation: a sufficiency gate over the conversation, the draft and context, choosing from the four actions above.
  • Tripwires run first and override both; required guards still run before release after any confident backend exit.
  • Every decision has a redacted audit record committed by the trusted host; the kernel does no I/O. Collect representative human labels across Act, Review and Escalate, including sampled automated decisions, rather than using only overrides. Audit failure must follow the playbook’s explicit release contract.
  • Rollout: shadow mode first, with the counselor reviewing every decision to get real recall numbers.

Self-hosting the same pattern (via the proposed LitJev port) can avoid sending inference inputs to a hosted provider, but does not by itself resolve storage, synchronization, telemetry, security or clinical obligations. Trusted policy and provenance authorize destinations; a sensitivity model can only veto, never grant egress. Unknown or internal data is not automatically eligible for hosted disclosure. The minimum calibration-positive count is a quantile requirement, not proof of deployment recall; synthetic cases do not count toward it. Evaluate the complete routed system and its subgroups under stated sampling assumptions. Plain Jev compatibility is advisory for approved nonclinical uses; clinical hosts require the native outcome contract, and the proxy must refuse any compatibility request whose required review or escalation cannot be safely represented. See open-decision-models-2026-09.md and ../PLAYBOOK.md.

Sources​