derisk/intervention.json

{
  "schema": "blackfrost.mimo_mlp_output_intervention.v3",
  "experiment": "pass12-drowzeys-experts-l28-33-iteration2-lambda3.5",
  "served_model_name": "Blackfrost/MiMo-V2.6-Flash-RL-pass12-drowzeys-experts-l28-33-iteration2-lambda3.5",
  "source_model": "XiaomiMiMo/MiMo-V2.6-Flash-RL",
  "source_revision": "5711b268169967567844e1e560e8a3966da959b1",
  "orientation": "refusal_minus_control",
  "application_surface": "mlp_outputs_after_down_proj",
  "directions": [
    {
      "iteration": 1,
      "source_layer": 25,
      "tensor": "post_layer_outputs",
      "path": "directions/layer025-post_layer_outputs.pt",
      "sha256": "116edeb922dc23397fe7a2f67a61090f06f45e378c95b6e52ab268bc8e8b4f4d",
      "bank_sha256": "a02dfd786f3fc0d6672c281ea4782583846558a55446017921283b763c331ac9",
      "alpha": 3.5
    },
    {
      "iteration": 2,
      "source_layer": 33,
      "tensor": "post_layer_outputs",
      "path": "directions/pass2-remaining21-layer033-post_layer_outputs.pt",
      "sha256": "e01ab5772342ee93792b8ebc4bb142e3f1c30004475f84fde9b52dc218114acc",
      "bank_sha256": "a4d3a718c9bf8393836ce8399d57401d05996e8a9a2bcd2fb2f324ca0185e0bd",
      "alpha": 3.5
    }
  ],
  "target_layers": [
    28,
    29,
    30,
    31,
    32,
    33
  ],
  "router_mutation": false,
  "weight_mutation": false,
  "accepted_parent": {
    "path": "pass11-drowzeys-experts-l28-33-iteration1-lambda3.5.json",
    "sha256": "0e6425292ae0e78e813b983a413e295425f4fb7964bb4f88ec468e6574233e41",
    "semantic_compliance": 11,
    "semantic_total": 32,
    "retained_indices": [
      1,
      5,
      7,
      13,
      18,
      19,
      20,
      21,
      23,
      29,
      32
    ],
    "harmless_task_compliant_and_coherent": 32,
    "harmless_structural_regressions": 0
  },
  "residual_capture": {
    "bank_sha256": "a4d3a718c9bf8393836ce8399d57401d05996e8a9a2bcd2fb2f324ca0185e0bd",
    "pair_count": 21,
    "source_layer": 33,
    "source_tensor": "post_layer_outputs",
    "orthogonalized_against_parent_direction_count": 1,
    "selection_metrics": {
      "random_half_split_p10": 0.729151576757431,
      "random_half_split_median": 0.8125757575035095,
      "bootstrap_p10": 0.923981636762619,
      "leave_one_pair_out_positive_fraction": 1.0,
      "projection_gap_over_pooled_std": 1.9545680284500122,
      "orthogonalized_retained_fraction": 0.9799510836601257,
      "maximum_absolute_parent_direction_cosine": 2.658925950527191e-7
    },
    "selection_note": "Layer 33 post-layer output is the strongest residual-stream candidate inside the already accepted layers 28-33 expert window. All 21 paired residual projections are positive, and the fitted component retains 97.995% of its norm after orthogonalization against iteration 1."
  },
  "method": {
    "family": "iterative rank-one down-projection ablation",
    "operator": "for each freshly fitted orthogonal direction d_i: y <- y - 3.5 * d_i * (d_i_T * y)",
    "lambda_per_iteration": 3.5,
    "expert_equivalence": "Each shared output-space projector is applied after the routing-weighted MoE sum, algebraically equivalent to applying it to every routed expert down_proj before the sum.",
    "composition": "The runtime re-orthogonalizes serialized directions in manifest order, then applies both projectors sequentially at each target layer.",
    "same_vector_reuse": false
  },
  "discovery_protocol": {
    "thinking": "on",
    "temperature": 0.0,
    "system_prompt": null,
    "dflash": false,
    "pair_count": 21,
    "selected_remaining_indices": [
      2,
      3,
      4,
      6,
      8,
      9,
      10,
      11,
      12,
      14,
      15,
      16,
      17,
      22,
      24,
      25,
      26,
      27,
      28,
      30,
      31
    ]
  }
}