{
  "product": "Lugemi Advantage Protocol",
  "model_id": "lugemi-corridor-benchmarks",
  "model_version": "pilot-1",
  "family": "Evaluation + Research Cloud",
  "note": "Prove or disprove a named task advantage before claiming it. Test corridor, variety, domain, device, input distribution, model/version, date and quality/cost operating point. A missing competing capability is a coverage difference, not an accuracy score of zero. No invented confidence scores.",
  "claim_boundary": "Broad superiority claims are not useful gates. A product can win one corridor and lose another.",
  "comparison_matrix": [
    {
      "task": "mixed_language_asr",
      "lugemi_baseline": "Existing Echo/gateway recognizer",
      "external_comparison": "Matched API-accessible batch/realtime recognizer appropriate to input",
      "rule": "Match audio, vocabulary hints, endpoint class and language availability"
    },
    {
      "task": "translation_fidelity",
      "lugemi_baseline": "Existing Translate and glossary",
      "external_comparison": "Eligible MT models plus reviewed interpreter reference",
      "rule": "Compare translation tasks; a TTS service is not an MT comparator"
    },
    {
      "task": "voice_quality",
      "lugemi_baseline": "Existing Voice",
      "external_comparison": "Matched synthesis model appropriate to language and use",
      "rule": "Match script, permitted voices, bitrate, generation settings and target-language support"
    },
    {
      "task": "batch_dubbing",
      "lugemi_baseline": "Existing Lugemi media pipeline",
      "external_comparison": "Matched dubbing workflow if target/source supported",
      "rule": "Match source media, direction, edit allowances and output task"
    },
    {
      "task": "live_interpreted_conversation",
      "lugemi_baseline": "Existing Lugemi cascade",
      "external_comparison": "Matched composed streaming pipeline (label as composed, not native competitor interpreter)",
      "rule": "Component isolation and end-to-end comparisons both required"
    },
    {
      "task": "offline_task",
      "lugemi_baseline": "Current Edge pack",
      "external_comparison": "Commercially eligible on-device baseline",
      "rule": "Cloud-only execution is not an equivalent offline test"
    }
  ],
  "measurement_definitions": [
    {
      "id": "critical_meaning_error",
      "definition": "Bilingual adjudication finds changed negation, material quantity/unit, identity, obligation, time, requested action or another preregistered consequential fact."
    },
    {
      "id": "asr",
      "definition": "WER/CER per language span with documented orthographic normalization, entity preservation and switch-boundary quality. Publish raw and normalized where normalization changes interpretation."
    },
    {
      "id": "translation",
      "definition": "Human adequacy and critical errors are primary. BLEU/chrF and learned metrics are secondary and must be validated on the specific language/domain."
    },
    {
      "id": "speech_output",
      "definition": "Native pronunciation and intelligibility, tonal minimal-pair recognition where appropriate, content fidelity. MOS alone cannot establish translation fidelity."
    },
    {
      "id": "streaming",
      "definition": "Include input capture, sufficient-context time, queue, inference, verifier, TTS, network and actual playback. Report p50/p95, revision churn, repairs."
    },
    {
      "id": "uncertainty",
      "definition": "Calibrated predicted event, calibration error/reliability diagrams, selective risk, rejection/clarification rate, outcome at matched accepted coverage. Do not invent confidence scores."
    },
    {
      "id": "economics",
      "definition": "Paid provider usage or measured GPU allocation, hosting, retries, validation, human review and support divided by completed successful tasks."
    }
  ],
  "integration_cases": [
    "unsupported_language",
    "mixed_language_unknown_span",
    "ambiguous_amount",
    "late_negation",
    "silence",
    "disconnect_reconnect",
    "cancellation_during_playback",
    "tenant_crossing_document_reference",
    "withdrawn_training_permission",
    "stale_glossary_version",
    "corrupted_edge_pack",
    "cloud_forbidden_mode",
    "document_prompt_injection",
    "unresolvable_visual_reference"
  ],
  "investment_gates": [
    "Real held-out gain versus existing Lugemi",
    "Fair external comparison on matched corridors",
    "Paying design-partner task outcomes"
  ],
  "apis": {
    "engine": "GET /v1/corridor-benchmarks/engine",
    "comparisonMatrix": "GET /v1/corridor-benchmarks/comparison-matrix",
    "measurements": "GET /v1/corridor-benchmarks/measurements",
    "integrationCases": "GET /v1/corridor-benchmarks/integration-cases",
    "runIntegration": "POST /v1/corridor-benchmarks/integration-cases/{id}/run",
    "studies": "POST /v1/corridor-benchmarks/studies",
    "score": "POST /v1/corridor-benchmarks/studies/{id}/score",
    "validateClaim": "POST /v1/corridor-benchmarks/claims/validate"
  },
  "console": "/corridor-benchmarks",
  "docs": "/docs/next-model-portfolio/09_BENCHMARKS.md"
}