{
  "research_question": "Do sparse features that predict a concept also causally influence model behaviour?",
  "model_id": "Qwen/Qwen3-1.7B-Base",
  "sae_repo_id": "Qwen/SAE-Res-Qwen3-1.7B-Base-W32K-L0_50",
  "layers": [
    4,
    14,
    26
  ],
  "sae_top_k": 50,
  "sae_width": 32768,
  "concepts": [
    "code",
    "mathematics",
    "positive_sentiment",
    "negative_sentiment",
    "german_language",
    "factual_entities",
    "uncertainty"
  ],
  "discovery_prompts": 224,
  "paraphrase_pairs_per_concept": 16,
  "causal_tasks": 28,
  "split_seed": 42,
  "feature_selection": "training-split AUROC with activation-rate contrast tie-break",
  "held_out_metrics": [
    "AUROC",
    "F1"
  ],
  "causal_interventions": [
    "ablate",
    "scale_2x"
  ],
  "negative_control": "norm-matched random residual direction ensemble",
  "primary_causal_metric": "full target continuation mean log-probability delta per token (teacher-forced)",
  "live_causal_controls": "8-direction norm-matched random residual ensemble",
  "dose_response_multipliers": [
    0.0,
    0.5,
    1.0,
    1.5,
    2.0,
    3.0
  ],
  "statistical_inference": [
    "bootstrap_95_ci",
    "paired_sign_flip_test"
  ],
  "feature_set_sizes": [
    1,
    3,
    5
  ],
  "feature_set_interventions": [
    "ablate",
    "scale"
  ],
  "feature_set_negative_control": "norm-matched random residual direction ensemble",
  "dose_response_execution": "all scale conditions share the 1x batched no-edit reference",
  "feature_set_sweep_execution": "zero-edit reference + SAE edits + random-control ensembles evaluated in one batched forward",
  "live_random_controls": 8,
  "offline_random_controls_default": 8,
  "control_reference": "batched zero-edit residual row",
  "paraphrase_promptwide_pooling": "max activation per SAE feature across all prompt tokens",
  "concept_contrast_prompts_per_concept": 4,
  "interaction_feature_limit": 5,
  "concept_contrast_pooling": "max activation across non-padding prompt tokens",
  "live_geometry_feature_limit": 8,
  "contrastive_preference_metric": "change in exact-sequence log-odds between two user-specified continuations",
  "concept_candidate_discovery_metric": "balanced exploratory score = selectivity × target activation rate × log1p(target mean); causal-ready mode additionally requires current-token activity and log-scales that activation; raw mean-difference remains available as a scale-sensitive comparison",
  "completion_cue_scan": "final-token feature activation after controlled suffix/cue substitution",
  "candidate_causal_screen_limit": 8,
  "candidate_causal_screen_control": "batched zero-edit reference; no random controls in triage screen",
  "candidate_alignment_metrics": [
    "discovery rank versus target-effect rank",
    "discovery rank versus next-token JS rank",
    "Spearman candidate score versus absolute target effect",
    "Spearman candidate score versus next-token JS"
  ],
  "candidate_specificity_limit": 3,
  "candidate_specificity_control": "each candidate gets live_random_controls deterministic norm-matched residual directions in one shared batched zero-edit execution context",
  "candidate_specificity_metrics": [
    "target effect versus candidate-specific random mean absolute effect",
    "next-token JS versus candidate-specific random mean JS",
    "coarse empirical random-control tail probabilities",
    "discovery rank versus random-normalized target-specificity rank"
  ],
  "cross_target_feature_limit": 3,
  "cross_target_target_limit": 5,
  "cross_target_default_targets": [
    "2x",
    "x",
    "0",
    "x^2"
  ],
  "discovery_resample_replicates": 32,
  "offline_feature_pooling": "prompt-wide max SAE activation across non-padding prompt tokens; final-token sparse activations saved separately",
  "offline_selection_resamples": 128,
  "offline_study_outputs": [
    "selection_stability.csv",
    "study_feature_summary.csv",
    "study_summary.json",
    "summary.json",
    "report.md"
  ],
  "offline_causal_position_policies": [
    "final_token",
    "max_feature_activation"
  ],
  "primary_offline_causal_position_policy": "max_feature_activation",
  "offline_causal_statistical_unit": "causal task; average ablation and amplification within task before paired bootstrap/sign-flip inference",
  "release_status": "final",
  "public_study_artifacts": "measured offline results committed under artifacts/"
}