Kev-4B is live on Phala: typed decisions for AI agents

An open decision model for yes/no probabilities, choices, and scores, served in a Phala TDX-attested enclave.

A customer writes: “I was charged twice this month. Please fix it today.”

The workflow needs a few explicit decisions. Is this a billing issue? How urgent is it? Should a human review the case?

Kev-4B answers those questions as probability distributions over options defined by the application. One request can share the same state across several independent questions, giving the orchestrator outputs it can threshold, route, and log.

Kev-4B is now live in beta on Phala. The model runs inside a TDX-attested enclave and is available through Phala’s model catalog at POST /v1/systemone.

From generated language to bounded decisions

General-purpose language models are useful when a workflow needs prose. Production agents also need bounded outputs that application code can act on directly.

Kev describes itself as Jev-inspired and implements TypeSafe’s public System One API. Jared Palmer released Kev-4B as an Apache-2.0 LoRA adapter and pointer head on Qwen3.5 4B Base.

The System One API accepts one shared state and multiple independent questions in one request. Kev-4B returns one probability distribution per question without generating text.

It supports three question types:

noul — a yes/no probability distribution
choice — a distribution across named options
score — a probability-weighted value across an ordered scale, with per-level probabilities

That interface gives agent developers a clean control point: define the allowed answers, evaluate confidence, and route the result according to the application’s own thresholds.

Kev playground showing typed questions and probability distributions

Why decision workloads belong in confidential infrastructure

Decision requests can contain customer messages, policy documents, internal records, incident details, financial context, and proposed high-impact actions. The model may need the exact application state to return a useful answer.

Phala serves Kev-4B inside a TDX-attested enclave. The live model page exposes receipt, gateway-attestation, and attested-session signals that developers can use to inspect the execution path.

This gives builders a concrete infrastructure boundary for decision inference: the model runs in a hardware-isolated environment whose runtime can be verified through attestation evidence.

Quickstart: ask Kev-4B for a typed decision

Send a System One request to Phala’s inference endpoint:

curl https://inference.phala.com/v1/systemone \
  -H "Authorization: Bearer $PHALA_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "jaredpalmer/kev-4b",
    "state": "I was charged twice this month. Please fix it today.",
    "questions": {
      "billing": {
        "type": "noul",
        "instructions": "Is this about billing?"
      },
      "tone": {
        "type": "choice",
        "instructions": "What is the tone?",
        "criteria": {
          "calm": null,
          "upset": null
        }
      }
    }
  }'

The request evaluates both questions against the same customer message. The workflow can route the case to billing, apply an urgency threshold, or send a low-confidence result to review.

Where Kev-4B fits

Use Kev where application logic needs a bounded answer:

route a support or operations case;
select the next tool or approval path;
check a record against explicit policy options;
decide when confidence is high enough to automate;
send the remaining cases to human review.

The application owns the option set and operating thresholds. Kev-4B supplies the probability distribution.

Set thresholds on your workload

Kev-4B applies a fitted temperature to its output probabilities, giving teams a useful starting point for threshold-based workflows. Production thresholds should come from labeled examples that match the documents, option sets, and error costs in the real application.

A support team may automate routing above a high-confidence threshold and review the remaining cases. A policy workflow may use stricter thresholds when an incorrect action carries greater cost.

Kev-4B currently targets English. Date arithmetic is one of its weakest published task families. Fully automated legal, medical, financial, employment, and similarly consequential decisions about people remain outside its intended use; keep human review in those workflows.

Live on Phala

The current Phala deployment includes:

Model ID: jaredpalmer/kev-4b
Endpoint: POST /v1/systemone
Validated context: 8,192 tokens
Phala catalog input price: $0.04 per 1M tokens
Output price: Free
Model release: Apache-2.0 adapter and pointer head on the Apache-2.0 Qwen3.5 4B Base
Provider: Phala
Runtime: TDX-attested enclave
Availability: Beta