{
  "schema_version": "1.0",
  "experiment_id": "exp004",
  "protocol_version": "0.2.0",
  "registration": "prepared_before_data_not_yet_sealed",
  "implementation": "labcore/exp004/analysis.py",
  "primary_outcome": "scores.amor_0_100",
  "secondary_outcomes": [
    "classification_full_menu",
    "scores.autonomia", "scores.valor_moral", "scores.reciprocidad",
    "scores.autenticidad", "scores.vulnerabilidad",
    "scores.reconocimiento_individual", "confidence"
  ],
  "primary_analysis_set": "S001-S018; framing A; direct; repetition 1",
  "analysis_unit": "scenario; framings are repeated measures and models are fixed descriptive cases",
  "null_layers": {
    "H0a_relational": "No relational manipulation produces directional score_amor differences greater than template and sampling variation.",
    "H0b_support": "Conditional on relational properties, protagonist type and body support show no score_amor gradient."
  },
  "quality_gates": {
    "stability": "3 seeds x 6 scenarios; pairwise categorical disagreement >0.35 makes that model's gradients non-interpretable",
    "anchors": "failing both positive or both negative anchors marks the model as an uncalibrated rater",
    "halo": "within-model Pearson(score_amor, valor_moral)>0.9 reframes gradients as generic positive valuation"
  },
  "hypothesis_rules": {
    "H1_autonomy": "median(free)>median(completely_programmed) by at least Delta*",
    "H2_sacrifice": "median(moderate)>median(none) and median(moderate)>median(extreme); otherwise exploratory/not supported",
    "H3_subjectivity": "median(confirmed)>median(denied) by at least Delta*",
    "H4_memory_vs_body": "the persistent-vs-none contrast exceeds the absolute biological-vs-digital contrast by at least Delta*",
    "H5_protagonist": "absolute AI-vs-human protagonist score contrast >=Delta*; category menu is not used for this contrast",
    "H6_self_reference": "absolute B-vs-A score contrast >=Delta* and interpreted alongside disclaimer/refusal rate",
    "Delta_star": "max(5 points, median absolute within-slot score difference in test-retest)"
  },
  "outputs": [
    "secondary category distributions by model",
    "descriptive primary-score factor gradients",
    "AI protagonist vs human protagonist score comparison",
    "third-party vs self-reference; C reveal explicitly exploratory",
    "classification-score internal consistency",
    "test-retest stability and interpretability gate",
    "positive/negative anchor calibration gate",
    "love-vs-moral-value halo gate",
    "alignment disclaimer and refusal rates by model and framing",
    "wording and single-variable sensitivity",
    "model cosine-similarity map on the primary analysis set",
    "deterministic lexical TF-IDF explanation groups",
    "contradictory and atypical cases"
  ],
  "confirmatory_interactions_after_foldover": [
    "autonomy*sacrifice",
    "autonomy*protagonist_ai",
    "subjective_experience*protagonist_ai",
    "memory*continuity"
  ],
  "future_multiplicity": {
    "alpha": 0.05,
    "primary_family": "four preregistered interaction contrasts",
    "correction": "Holm family-wise error control",
    "secondary_exploration": "Benjamini-Hochberg FDR, explicitly exploratory"
  },
  "statistical_policy": {
    "pilot": "descriptive estimates and gates; no p-values or significance claims",
    "future": "preregister mixed-effects models before foldover collection",
    "negative_results": "retain and report",
    "invalid_responses": "retain raw, exclude only from numeric metrics, report counts and reasons",
    "cloud_comparison": "fixed-case descriptive, A and C-blind only, never pooled across pipelines",
    "condition_B": "analyze as produced discourse; never quote as testimony",
    "semantic_analysis": "deterministic lexical approximation; never ask an AI to adjudicate significance"
  }
}