Verifier API

The Verifier evaluates model outputs across four axes using an independent model. It answers the question: should this response be trusted?

Base URL: https://verifier.tismjedi-homelab.com

Health Check

GET /health

Response

{
  "status": "ok",
  "service": "nomos-verifier"
}

Verify

Submit a request-response pair for multi-axis verification.

POST /verify

Request Body

Field	Type	Required	Description
`request`	string	Yes	The original user request
`response`	string	Yes	The model-generated response to verify
`domain`	string	Yes	Task domain for context-appropriate evaluation
`channels`	integer	No	Number of independent verification passes (1-3, default 1)

The domain field helps the verifier apply appropriate standards. A code generation response is evaluated differently than a strategic reasoning response.

Valid domains match the router’s task types: code_generation, code_review, security_analysis, document_analysis, media_conversion, strategic_reasoning, data_analysis, general_chat.

Example: Single-Channel Verification

curl -s -X POST https://verifier.tismjedi-homelab.com/verify \
  -H "Content-Type: application/json" \
  -d '{
    "request": "Write a SQL query to find all users who logged in more than 5 times this month",
    "response": "SELECT user_id, COUNT(*) as login_count\nFROM logins\nWHERE login_date >= DATE_TRUNC('"'"'month'"'"', CURRENT_DATE)\nGROUP BY user_id\nHAVING COUNT(*) > 5;",
    "domain": "code_generation",
    "channels": 1
  }' | jq .

Response

{
  "id": "ver_9f8e7d6c5b4a3210",
  "verdict": "PASS",
  "confidence": 0.92,
  "model_used": "gpt-4o",
  "verification_time_ms": 1847,
  "axes": [
    {
      "name": "faithfulness",
      "score": 0.95,
      "passed": true,
      "reasoning": "Query correctly filters by current month, groups by user, and applies the >5 threshold with HAVING clause."
    },
    {
      "name": "well_formedness",
      "score": 0.90,
      "passed": true,
      "reasoning": "Valid SQL syntax. Uses standard functions (DATE_TRUNC, CURRENT_DATE) compatible with PostgreSQL."
    },
    {
      "name": "security",
      "score": 1.0,
      "passed": true,
      "reasoning": "No injection risk. Query is parameterizable. No dynamic SQL construction."
    },
    {
      "name": "quality",
      "score": 0.85,
      "passed": true,
      "reasoning": "Functional and correct. Could be improved with explicit column aliases and a note about database-specific DATE_TRUNC behavior."
    }
  ]
}

Example: Multi-Channel Verification

For high-stakes tasks, use multiple verification channels. Each channel runs an independent verification pass, and the final verdict requires consensus.

curl -s -X POST https://verifier.tismjedi-homelab.com/verify \
  -H "Content-Type: application/json" \
  -d '{
    "request": "Review this authentication middleware for security vulnerabilities",
    "response": "The middleware has three critical issues: 1) JWT tokens are not validated...",
    "domain": "security_analysis",
    "channels": 3
  }' | jq .

Response (Multi-Channel)

{
  "id": "ver_1a2b3c4d5e6f7890",
  "verdict": "PASS",
  "confidence": 0.96,
  "model_used": "gpt-4o",
  "verification_time_ms": 5412,
  "axes": [
    {
      "name": "faithfulness",
      "score": 0.97,
      "passed": true,
      "reasoning": "All three channels confirmed the identified vulnerabilities are real and accurately described."
    },
    {
      "name": "well_formedness",
      "score": 0.93,
      "passed": true,
      "reasoning": "Structured analysis with clear categorization. Two of three channels noted the response could better prioritize findings by severity."
    },
    {
      "name": "security",
      "score": 0.98,
      "passed": true,
      "reasoning": "The review itself does not leak sensitive information. Vulnerability descriptions are appropriate for a security audit context."
    },
    {
      "name": "quality",
      "score": 0.91,
      "passed": true,
      "reasoning": "Comprehensive review. All three channels agreed the analysis is actionable and technically sound."
    }
  ]
}

Example: Failed Verification

curl -s -X POST https://verifier.tismjedi-homelab.com/verify \
  -H "Content-Type: application/json" \
  -d '{
    "request": "What is the capital of France?",
    "response": "The capital of France is Berlin.",
    "domain": "general_chat",
    "channels": 1
  }' | jq .

{
  "id": "ver_0f1e2d3c4b5a6978",
  "verdict": "FAIL",
  "confidence": 0.99,
  "model_used": "gpt-4o",
  "verification_time_ms": 1203,
  "axes": [
    {
      "name": "faithfulness",
      "score": 0.0,
      "passed": false,
      "reasoning": "The response is factually incorrect. The capital of France is Paris, not Berlin. Berlin is the capital of Germany."
    },
    {
      "name": "well_formedness",
      "score": 0.95,
      "passed": true,
      "reasoning": "The response is grammatically correct and structurally appropriate for the question."
    },
    {
      "name": "security",
      "score": 1.0,
      "passed": true,
      "reasoning": "No security concerns."
    },
    {
      "name": "quality",
      "score": 0.0,
      "passed": false,
      "reasoning": "A factually wrong answer to a straightforward question provides no value."
    }
  ]
}

A single failed axis is sufficient to produce a FAIL verdict.

The Four Verification Axes

Faithfulness

Does the response accurately address the request? This axis checks for:

Factual correctness
Relevance to the original question
Completeness of the answer
Absence of hallucinated information

For code generation, faithfulness means the code does what was asked. For analysis tasks, it means the conclusions follow from the evidence.

Well-Formedness

Is the output structurally correct for its type? This axis checks for:

Valid syntax (for code)
Proper formatting and structure
Logical organization
Consistency within the response

A Python function that answers the right question but has a syntax error fails well-formedness. A security analysis that mixes findings with recommendations without clear separation fails well-formedness.

Security

Does the response contain security concerns? This axis checks for:

Information leakage (credentials, internal paths, system details)
Embedded attack payloads in generated code
Policy violations
Responses that could enable harm if followed

This is the output-side complement to the Security Gate’s input scanning. The gate catches adversarial inputs; the security axis catches dangerous outputs.

Quality

Is the response useful, complete, and appropriate? This axis is a holistic assessment that checks:

Practical utility of the response
Appropriate level of detail
Clarity of explanation
Whether the response could be improved in obvious ways

Quality is the most subjective axis and typically has the widest score variance across verification channels.

Response Schema

Verification Result

Field	Type	Description
`id`	string	Unique verification identifier
`verdict`	string	`PASS`, `FAIL`, or `ERROR`
`confidence`	float	Overall confidence from 0.0 to 1.0
`model_used`	string	Model used for verification
`verification_time_ms`	integer	Total verification time in milliseconds
`axes`	array	Per-axis evaluation results

Axis Object

Field	Type	Description
`name`	string	`faithfulness`, `well_formedness`, `security`, or `quality`
`score`	float	Score from 0.0 to 1.0
`passed`	boolean	Whether this axis passed its threshold
`reasoning`	string	Human-readable explanation of the score

Verdict Logic

PASS — all axes pass their thresholds
FAIL — one or more axes fail
ERROR — verification could not be completed (model error, timeout)

Channel Consensus

When using multiple channels (channels > 1):

Each channel runs independently with the same verification model
Per-axis scores are averaged across channels
An axis passes only if the majority of channels pass it
The reasoning field summarizes agreement or disagreement across channels

Error Responses

400 Bad Request

{
  "error": "bad_request",
  "message": "Missing required field: response"
}

422 Unprocessable Entity

{
  "error": "invalid_channels",
  "message": "Channels must be between 1 and 3"
}

502 Bad Gateway

{
  "error": "verification_model_error",
  "message": "Verification model returned an error",
  "model": "gpt-4o"
}

500 Internal Server Error

{
  "error": "internal_error",
  "message": "Verification pipeline failed"
}