Decision Models

The decision component family evaluates discrete questions (choose) and assertion verifications (verify) against context using non-autoregressive, calibrated discriminative models.

This allows Daimon pipelines to perform deterministic leaf-node disambiguation, clinical coding matching, and hallucination checks without token generation overhead.

---

Configuration

Declare decision models under components: in your YAML config:

1. HTTP Webhook (decision/http)

Forwards requests to any external service implementing /choose and /verify:

components:
  - name: verifier
    type: decision/http
    metadata:
      base_url: http://localhost:8000
      api_key: optional-bearer-token
      timeout_ms: "800"

2. TypeSafe / Jev (decision/jev)

Connects to the TypeSafe Jev calibrated discriminative decision engine:

components:
  - name: verifier
    type: decision/jev
    metadata:
      api_key: ${JEV_API_KEY}
      base_url: https://api.typesafe.com/v1
      default_model: jev-v1
      timeout_ms: "800"

---

Auto-Generated LLM Tools

Declaring a decision model registers two tools for LLM agent use:

1. {name}_choose

• Description: Selects the single best match among a list of candidate strings given context.
• Parameters:
• question (string, required): The target question.
• choices (array of strings, required): Candidate list (up to 255 items).
• state (string, optional): Context excerpt. Defaults to active conversation history if omitted.

2. {name}_verify

• Description: Returns calibrated truth probability for a boolean assertion.
• Parameters:
• statement (string, required): Statement to verify against state.
• state (string, optional): Context excerpt.

---

HTTP REST API

POST /v1/decision/{name}/choose

curl -s http://127.0.0.1:3500/v1/decision/verifier/choose \
  -H "Content-Type: application/json" \
  -d '{
    "state": "Patient presents with right knee pain following a sports injury.",
    "question": "Which ICD-10 code matches laterality?",
    "choices": ["M25.561", "M25.562", "M25.569"]
  }'

Response:

{
  "selected": "M25.561",
  "index": 0,
  "probabilities": {
    "M25.561": 0.97,
    "M25.562": 0.02,
    "M25.569": 0.01
  }
}

POST /v1/decision/{name}/verify

curl -s http://127.0.0.1:3500/v1/decision/verifier/verify \
  -H "Content-Type: application/json" \
  -d '{
    "state": "Patient presents with right knee pain. Examination reveals joint effusion.",
    "statement": "Right knee joint effusion is documented."
  }'

Response:

{
  "probability": 0.98,
  "supported": true
}

---

SDK Usage

PythonTypeScriptJava

import daimon

client = daimon.Client()
verifier = client.decision("verifier")

# 1. Choose best matching code
choice = verifier.choose(
    question="Which ICD-10 code matches laterality?",
    choices=["M25.561", "M25.562", "M25.569"],
    state="Patient presents with right knee pain following a sports injury.",
)
print(f"Selected: {choice.selected} (p={choice.probabilities[choice.selected]:.2f})")

# 2. Verify boolean assertion
verify = verifier.verify(
    statement="Right knee joint effusion is documented.",
    state="Patient presents with right knee pain. Examination reveals joint effusion.",
)
print(f"Supported: {verify.supported} (p={verify.probability:.2f})")

import { Client } from 'daimon-client';

const client = new Client();
const verifier = client.decision('verifier');

// 1. Choose best matching code
const choice = await verifier.choose(
  'Which ICD-10 code matches laterality?',
  ['M25.561', 'M25.562', 'M25.569'],
  { state: 'Patient presents with right knee pain following a sports injury.' }
);
console.log(`Selected: ${choice.selected}`);

// 2. Verify boolean assertion
const verify = await verifier.verify('Right knee joint effusion is documented.', {
  state: 'Patient presents with right knee pain. Examination reveals joint effusion.',
});
console.log(`Supported: ${verify.supported} (p=${verify.probability})`);

import io.github.sonicboom15.daimon.ChoiceResult;
import io.github.sonicboom15.daimon.Client;
import io.github.sonicboom15.daimon.DecisionClient;
import io.github.sonicboom15.daimon.VerifyResult;
import java.util.List;

Client client = new Client();
DecisionClient verifier = client.decision("verifier");

ChoiceResult choice = verifier.choose(
    "Which ICD-10 code matches laterality?",
    List.of("M25.561", "M25.562", "M25.569"),
    "Patient presents with right knee pain following a sports injury."
);

VerifyResult verify = verifier.verify(
    "Right knee joint effusion is documented.",
    "Patient presents with right knee pain. Examination reveals joint effusion."
);