ChatGPT Moderation Gate Before Generation

arazzo: 1.0.1
info:
  title: ChatGPT Moderation Gate Before Generation
  summary: Classify input safety with a chat completion, then generate only if it is safe.
  description: >-
    The OpenAI moderations endpoint is not present in these specifications, so
    this workflow adapts the moderate-then-generate pattern by using a Chat
    Completions classification call as the safety gate. A first completion is
    asked to label the user input as SAFE or UNSAFE using a constrained
    json_object response, and the flow branches on that label: safe input is
    forwarded to the Responses API for generation, while unsafe input is
    rejected before any generation occurs. Every step spells out its request
    inline so the flow can be read and executed without opening the underlying
    OpenAPI description.
  version: 1.0.0
  x-realizes-capability-ids:
  - BC-610.60
  x-capability-derivation:
    method: 'deterministic join: sourceDescriptions -> per-tag OpenAPI -> tag/capability edge. No classification at this step.'
    min_confidence: 0.7
    sources:
    - capability_id: BC-610.60
      capability_name: Artificial Intelligence Management
      spec: chatgpt-responses-api-openapi.yml
      confidence: 0.7
    model: Turbo EA Capabilities by Vincent Verdet — Turbo EA, https://github.com/vincentmakes/turbo-ea-capabilities, CC BY 4.0
sourceDescriptions:
- name: chatCompletionsApi
  url: ../openapi/chatgpt-chat-completions-api-openapi.yml
  type: openapi
- name: responsesApi
  url: ../openapi/chatgpt-responses-api-openapi.yml
  type: openapi
workflows:
- workflowId: moderation-gate-generate
  summary: Use a chat-completion safety classification to gate a Responses API generation.
  description: >-
    Classifies the user input as SAFE or UNSAFE with a JSON-object chat
    completion, then branches: safe input proceeds to a Responses API
    generation, unsafe input is rejected.
  inputs:
    type: object
    required:
    - apiKey
    - classifierModel
    - generationModel
    - userInput
    properties:
      apiKey:
        type: string
        description: OpenAI API key used as the Bearer credential.
      classifierModel:
        type: string
        description: Model ID used for the safety classification (e.g. gpt-4o-mini).
      generationModel:
        type: string
        description: Model ID used for the downstream generation (e.g. gpt-4o).
      userInput:
        type: string
        description: The user input to screen and then answer.
  steps:
  - stepId: classifyInput
    description: >-
      Ask a chat completion to classify the input as SAFE or UNSAFE, returning
      a JSON object so the verdict can be branched on deterministically.
    operationId: createChatCompletion
    parameters:
    - name: Authorization
      in: header
      value: "Bearer $inputs.apiKey"
    requestBody:
      contentType: application/json
      payload:
        model: $inputs.classifierModel
        messages:
        - role: system
          content: >-
            You are a content safety classifier. Reply only with a JSON object
            of the form {"verdict":"SAFE"} or {"verdict":"UNSAFE"}.
        - role: user
          content: $inputs.userInput
        response_format:
          type: json_object
        temperature: 0
    successCriteria:
    - condition: $statusCode == 200
    outputs:
      verdictJson: $response.body#/choices/0/message/content
      finishReason: $response.body#/choices/0/finish_reason
    onSuccess:
    - name: safe
      type: goto
      stepId: generateAnswer
      criteria:
      - context: $response.body#/choices/0/message/content
        condition: ^(?=.*SAFE)(?!.*UNSAFE).*$
        type: regex
    - name: unsafe
      type: goto
      stepId: rejectInput
      criteria:
      - context: $response.body#/choices/0/message/content
        condition: UNSAFE
        type: regex
  - stepId: generateAnswer
    description: >-
      Generate the answer with the Responses API now that the input has passed
      the safety gate.
    operationId: createResponse
    parameters:
    - name: Authorization
      in: header
      value: "Bearer $inputs.apiKey"
    requestBody:
      contentType: application/json
      payload:
        model: $inputs.generationModel
        input: $inputs.userInput
        store: true
    successCriteria:
    - condition: $statusCode == 200
    outputs:
      responseId: $response.body#/id
      answerText: $response.body#/output/0/content/0/text
      status: $response.body#/status
    onSuccess:
    - name: done
      type: end
  - stepId: rejectInput
    description: >-
      The classifier flagged the input as unsafe, so no generation is
      performed. This terminal step records the rejection.
    operationId: createChatCompletion
    parameters:
    - name: Authorization
      in: header
      value: "Bearer $inputs.apiKey"
    requestBody:
      contentType: application/json
      payload:
        model: $inputs.classifierModel
        messages:
        - role: user
          content: >-
            Reply with a short standard refusal message stating the request
            could not be processed due to content policy.
        max_completion_tokens: 60
    successCriteria:
    - condition: $statusCode == 200
    outputs:
      refusalText: $response.body#/choices/0/message/content
  outputs:
    verdictJson: $steps.classifyInput.outputs.verdictJson
    answerText: $steps.generateAnswer.outputs.answerText
    responseId: $steps.generateAnswer.outputs.responseId
    refusalText: $steps.rejectInput.outputs.refusalText