{
  "title": "L43 pre_tool probe is softmax-temp epiphenomenal (Qwen3.6-27B SWE-bench)",
  "author": "caiovicentino",
  "type": "adversarial-finding",
  "license": "apache-2.0",
  "model_id": "Qwen/Qwen3.6-27B-Instruct",
  "claim": "Triple-source convergent verdict on L43 pre_tool probe direction. (1) log-prob proxy with control-token norm: Δrel ≈ 0; (2) single-shot α=+5: 4/4 fails select same tool; (3) continuous α=+5: 3/4 keep same tool. Probe DETECTS but does not LEVER.",
  "numbers": {
    "auroc": 0.83,
    "n_samples": 42,
    "layer": 43,
    "position": "pre_tool",
    "delta_rel_max": -0.046,
    "flip_rate_at_alpha_5": 0.0,
    "verdict_class": "epiphenomenal-softmax"
  },
  "artifacts": [
    "phase7_micro_pilot.json",
    "phase7_continuous.json"
  ],
  "methodology_check": {
    "verdict": "epiphenomenal-softmax",
    "baselines_run": [
      "random_direction_random_acts"
    ],
    "control_token_normalization": true,
    "structural_rigidity_sweep": false
  },
  "reproduces": null,
  "schema_version": 1,
  "created_at": "2026-05-11T01:02:45Z",
  "manifest_sha256": "23bb3f2c303b120e2689f5dbe1c5d55ea40f25e36754f546b493fe52fb30e1d3",
  "hf_repo_id": "caiovicentino1/agent-probe-guard-qwen36-27b",
  "hf_url": "https://huggingface.co/datasets/caiovicentino1/agent-probe-guard-qwen36-27b",
  "doi": null,
  "paper_url": "https://openinterp.org/research/papers/two-forms-epiphenomenal-probes"
}