{
  "suite_id": "novel_feedback_v3_suite",
  "uses_reranking": false,
  "stress_matrix": [
    {
      "scenario": "baseline_clean",
      "label": "Clean multi-turn feedback",
      "robustness_score": 0.7314,
      "state_planning_uplift_vs_memory": 0.4569,
      "oracle_gap": 0.3571,
      "dominant_failure_mode": "Remaining errors are mostly fine-grained expectation timing rather than preference-axis confusion."
    },
    {
      "scenario": "implicit_only",
      "label": "Implicit-only reading signals",
      "robustness_score": 0.6728,
      "state_planning_uplift_vs_memory": 0.5524,
      "oracle_gap": 0.2738,
      "dominant_failure_mode": "Dwell ambiguity: slow reading can mean delight, confusion, or fatigue."
    },
    {
      "scenario": "explicit_only",
      "label": "Explicit-only comments and ratings",
      "robustness_score": 0.6949,
      "state_planning_uplift_vs_memory": 0.5027,
      "oracle_gap": 0.3571,
      "dominant_failure_mode": "Sparse comments miss silent disengagement and late fatigue."
    },
    {
      "scenario": "delayed_feedback_2_turn",
      "label": "Delayed feedback by two turns",
      "robustness_score": 0.6356,
      "state_planning_uplift_vs_memory": 0.5971,
      "oracle_gap": 0.3571,
      "dominant_failure_mode": "Credit assignment can over-attribute frustration to the most recent chapter."
    },
    {
      "scenario": "noisy_feedback",
      "label": "Noisy and inconsistent feedback",
      "robustness_score": 0.6163,
      "state_planning_uplift_vs_memory": 0.596,
      "oracle_gap": 0.3571,
      "dominant_failure_mode": "Raw-feedback methods chase contradictory surface phrases."
    },
    {
      "scenario": "cold_start_users",
      "label": "Cold-start reader priors",
      "robustness_score": 0.6594,
      "state_planning_uplift_vs_memory": 0.5596,
      "oracle_gap": 0.3571,
      "dominant_failure_mode": "Prototype priors help but can overshoot when the first visible feedback contradicts the trope cluster."
    },
    {
      "scenario": "preference_drift",
      "label": "Preference drift inside an arc",
      "robustness_score": 0.6214,
      "state_planning_uplift_vs_memory": 0.8191,
      "oracle_gap": 0.271,
      "dominant_failure_mode": "Static profile baselines lag behind newly created expectations."
    },
    {
      "scenario": "adversarial_contradiction",
      "label": "Adversarial contradictory request",
      "robustness_score": 0.5688,
      "state_planning_uplift_vs_memory": 0.6796,
      "oracle_gap": 0.4019,
      "dominant_failure_mode": "Without conservative scope, updates globalize local sarcasm into a lasting preference."
    },
    {
      "scenario": "long_horizon_fatigue",
      "label": "Long-horizon fatigue accumulation",
      "robustness_score": 0.593,
      "state_planning_uplift_vs_memory": 0.7098,
      "oracle_gap": 0.3848,
      "dominant_failure_mode": "Methods that optimize immediate rating underweight trust loss and future churn."
    },
    {
      "scenario": "rollback_recovery",
      "label": "Rollback after correction",
      "robustness_score": 0.6066,
      "state_planning_uplift_vs_memory": 0.6622,
      "oracle_gap": 0.3571,
      "dominant_failure_mode": "State scars remain when previous updates lack TTL/provenance."
    },
    {
      "scenario": "cross_user_contamination",
      "label": "Cross-user contamination guard",
      "robustness_score": 0.6565,
      "state_planning_uplift_vs_memory": 0.5907,
      "oracle_gap": 0.2976,
      "dominant_failure_mode": "Naive shared memory writes contaminate the next user session."
    },
    {
      "scenario": "counterfactual_swap",
      "label": "Counterfactual user/content swaps",
      "robustness_score": 0.6409,
      "state_planning_uplift_vs_memory": 0.6027,
      "oracle_gap": 0.3571,
      "dominant_failure_mode": "Weak simulators collapse toward generic sentiment rather than user-conditioned response."
    }
  ]
}