PGAC Phase 2: 🔴 INSUFFICIENT — methodology needs revision (different SAE? different layer pair?)

FINAL_VERDICT.json

@@ -0,0 +1,36 @@
+{
+  "experiment": "PGAC Phase 2 \u2014 probe training cross-layer",
+  "layer_pairs": [
+    [
+      11,
+      31
+    ],
+    [
+      11,
+      55
+    ]
+  ],
+  "results_per_pair": {
+    "L11->L31": {
+      "recall_at_k": 0.4686281979084015,
+      "recall_at_k_std": 0.09659925103187561,
+      "recall_at_2k": 0.5964295864105225,
+      "auroc_mean": 0.784482320220573,
+      "auroc_p10": 0.6383487495239836,
+      "auroc_p90": 0.962664110737133,
+      "n_test": 449
+    },
+    "L11->L55": {
+      "recall_at_k": 0.265224814414978,
+      "recall_at_k_std": 0.09403426945209503,
+      "recall_at_2k": 0.3707892596721649,
+      "auroc_mean": 0.8048945357768886,
+      "auroc_p10": 0.768089443677679,
+      "auroc_p90": 0.8416996278760984,
+      "n_test": 449
+    }
+  },
+  "best_pair": "L11->L31",
+  "verdict": "\ud83d\udd34 INSUFFICIENT \u2014 methodology needs revision (different SAE? different layer pair?)",
+  "paper_status": "rejected"
+}

README.md

@@ -0,0 +1,10 @@
+---
+license: apache-2.0
+tags: [pgac, probe, sae, qwen36-27b, adaptive-compute]
+---
+# PGAC Phase 2 — Probe-Gated Adaptive Compute
+
+Verdict: 🔴 INSUFFICIENT — methodology needs revision (different SAE? different layer pair?)
+Best pair: L11->L31
+Recall@k: 0.4686
+AUROC: 0.7845

probe_L11_to_L31.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5e5f1f4c72f9f7dbb7a8aaebab0e62ddb307f6ff7b2a2fa188cf33b2f160a720
+size 1342441453

probe_L11_to_L55.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ad6a807c6c9dfa4e27a9b1fd25b41242f05eb1928012438bc6b856fd11d21a4c
+size 1342441453