Atlas / atlas/qwen3-0.6b-4bit/interventions/v1

qwen3-0.6b-4bit — interventions v1

Level 2 — method-robust

The map

agreement — causal direction-erasure BAND map (specific damage by layer)

mean of 6 fresh sourcesworst source (min)random-direction damage

The published claim is the BAND (layers 2–15): erasing the diff-of-means agreement direction at any single early-band layer destroys most of the grammatical margin (baseline 4.45), replicated across six fresh direction estimates on a fresh behavioral bank (Level 2). The late band is displayed but carries no claim — run #3's per-layer version was refused by the publication gate as estimator-unstable. Anti-correlates with the probes/v2 decodability ranking (survival ledger 0/2): decodability peaks ≠ causal joints.

Confidence

Confidence — qwen3-0.6b-4bit / interventions / v1

Level      : 2
Seeds      : six fresh direction sources (disjoint halves of two promptsets
             + two fresh bootstraps); fresh behavioral bank
Prompt sets: 3 (two estimation sets + held-out behavioral bank)
Methods in agreement : 2 (diff-of-means probing; direction erasure) — shared
                       estimator, hence Level 2 and not 3
Causal verification  : YES — rank-1 erasure with random-direction nulls

Pre-registered band claim (bar 2.5 on a +4.45 baseline margin):

Ahalf1: early-band mean +3.347
Ahalf2: early-band mean +3.228
bootA: early-band mean +3.302
Bhalf1: early-band mean +3.263
Bhalf2: early-band mean +3.326
bootB: early-band mean +3.270 Minimum early-band mean across sources: +3.228 — claim PASSES.

Granularity discipline: run #3's per-layer profile FAILED replication and was refused by this very gate; the published object is the BAND (layers 2–15). The late band (20–27) is displayed but carries no claim. This map is the causal counterpart of probes/v2, whose decodability ranking it contradicts (survival ledger 0/2) — both stay published, labeled by what they measure.

Provenance

{
  "author": "Simon-Pierre Boucher",
  "contact": "contact@spboucher.ai",
  "website": "https://modelmap.io",
  "model_id": "mlx-community/Qwen3-0.6B-4bit",
  "map_type": "interventions",
  "version": "v1",
  "commit": "f7c298d2859ed43654d388618b7274266e7e562b",
  "model_hash": "392e8d466d56100ada00eb82031fb854297fc9e389b7d303eba3af114e87bce2",
  "config": {
    "model": "mlx-community/Qwen3-0.6B-4bit",
    "sources": [
      "Ahalf1",
      "Ahalf2",
      "bootA",
      "Bhalf1",
      "Bhalf2",
      "bootB"
    ],
    "bar": 2.5,
    "early_layers": [
      2,
      15
    ],
    "late_layers": [
      20,
      27
    ],
    "n_random_dirs": 3,
    "n_pairs": 192,
    "seed": 780
  },
  "seed": 780,
  "hardware_manifest": {
    "author": "Simon-Pierre Boucher",
    "contact": "contact@spboucher.ai",
    "website": "https://modelmap.io",
    "chip": {
      "brand": "Apple M5 Max",
      "cores_total": 18,
      "cores_performance": 6,
      "cores_efficiency": 12
    },
    "memory": {
      "unified_gb": 48,
      "pagesize": 16384
    },
    "os": {
      "system": "Darwin",
      "version": "27.0",
      "arch": "arm64"
    },
    "software": {
      "python": "3.14.4",
      "numpy": "2.5.2",
      "mlx": "0.32.0",
      "torch": "2.13.0",
      "safetensors": "0.8.0"
    }
  },
  "created": "2026-08-12",
  "source_results": "results/expC_causal_verification/20260812T071954Z/results.json"
}

Files

confidence.md 1.4 KiB
map.json 7.7 KiB
mapcard.json 2.8 KiB
provenance.json 1.4 KiB