kinoTou

experiment log

S1.4 HF Layer-L Hidden Probe Results

File
2026-07-01_s1_4_hf_layer_probe.md
Size
3.0 KB
SHA-256
05fe92f7d5ac650a…
Download raw file ↓Back to research →
Primary source. This is the verbatim Oczy document. The analytical field notes on the research page interpret and summarize these sources.

S1.4 HF Layer-L Hidden Probe Results

  • Model: Qwen/Qwen2.5-0.5B-Instruct
  • Layers: 24 decoder layers
  • Embedding dim: 896
  • Corpus hash: e5b4ef66beb2
  • Primary pooling: mean (mean-pool over content tokens)
  • Mid-layer range: 6-17 (25–75% depth)
  • Timestamp: 2026-07-02T00:34:39.236374+00:00
Layer mean ← primary last max
L0 0.4198 0.0080 0.0890
L1 0.3012 0.0117 0.0739
L2 0.0788 0.0230 0.0000
L3 0.0789 0.0273 0.0000
L4 0.0779 0.0606 0.0000
L5 0.0952 0.0718 0.0000
L6 0.0922 0.0761 0.0000
L7 0.0857 0.0856 0.0000
L8 0.0874 0.1039 0.0000
L9 0.0829 0.1013 0.0000
L10 0.0857 0.1214 0.0000
L11 0.0876 0.1162 0.0000
L12 0.0815 0.1071 0.0000
L13 0.0829 0.1126 0.0000
L14 0.0860 0.1351 0.0000
L15 0.0812 0.1344 0.0000
L16 0.0843 0.1288 0.0000
L17 0.0792 0.1171 0.0000
L18 0.0760 0.1241 0.0001
L19 0.0719 0.1311 0.0001
L20 0.0708 0.1272 0.0001
L21 0.1219 0.1022 0.0647
L22 0.1193 0.1074 0.0694
L23 0.1753 0.1421 0.0909

Verdict

  • Final-layer (mean) silhouette: 0.1753
  • Max mid-layer (mean) silhouette: 0.0922
  • Gap: -0.0831 (threshold: +0.10)
  • Verdict: REFUTE

H-L refuted: mid-layer hiddens do NOT cluster by concept better than the final layer (gap < +0.10). This confirms lane_03's refutation on a substrate that can see every layer.

Pre-registered spec: research/10-hf-layer-l-hidden-probe.md

Secondary #3: LFM2.5-1.2B-Instruct

S1.4 HF Layer-L Hidden Probe Results

  • Model: LiquidAI/LFM2.5-1.2B-Instruct
  • Layers: 16 decoder layers
  • Embedding dim: 2048
  • Corpus hash: e5b4ef66beb2
  • Primary pooling: mean (mean-pool over content tokens)
  • Mid-layer range: 4-11 (25–75% depth)
  • Timestamp: 2026-07-02T00:34:43.309519+00:00
Layer mean ← primary last max
L0 0.2385 0.0090 0.0671
L1 0.1656 0.0048 0.0053
L2 0.3524 0.0926 0.0062
L3 0.3785 0.1175 0.0054
L4 0.1649 0.0712 0.0051
L5 0.3358 0.1706 0.0071
L6 0.3158 0.1701 0.0154
L7 0.1364 0.1298 0.0002
L8 0.2552 0.1945 0.0003
L9 0.0752 0.1773 0.0006
L10 0.1458 0.2391 0.0009
L11 0.1185 0.2212 0.0019
L12 0.2145 0.2930 0.0050
L13 0.2362 0.2682 0.0128
L14 0.2096 0.2780 0.0221
L15 0.2781 0.3274 0.0548

Verdict

  • Final-layer (mean) silhouette: 0.2781
  • Max mid-layer (mean) silhouette: 0.3358
  • Gap: +0.0576 (threshold: +0.10)
  • Verdict: REFUTE

H-L refuted: mid-layer hiddens do NOT cluster by concept better than the final layer (gap < +0.10). This confirms lane_03's refutation on a substrate that can see every layer.

Pre-registered spec: research/10-hf-layer-l-hidden-probe.md