sync EmbedEd artifacts (phase01-v6)

artifacts/gate_runs.json

@@ -32,6 +32,38 @@
         }
       }
     },
+    {
+      "stamp": null,
+      "version": "phase01-v6",
+      "rule": "D1 scale-aware: C1 < C2 <= C3 and d(C1->C2) >= 0.5",
+      "ok": true,
+      "d_c1_c2": 0.614,
+      "ci": [
+        0.795,
+        1.202
+      ],
+      "recorded_source": null,
+      "per_condition": {
+        "C1": {
+          "mean_cos": 0.96068,
+          "std": 0.0257,
+          "n_samples": 20000,
+          "n_anchors": 1000
+        },
+        "C2": {
+          "mean_cos": 0.97526,
+          "std": 0.02162,
+          "n_samples": 20000,
+          "n_anchors": 1000
+        },
+        "C3": {
+          "mean_cos": 0.98845,
+          "std": 0.00722,
+          "n_samples": 20000,
+          "n_anchors": 1000
+        }
+      }
+    },
     {
       "stamp": null,
       "version": "phase01-v6",

report/measurements.md

@@ -503,6 +503,17 @@ Context recorded at the time: ordering C1 < C2 ≤ C3 holds; the C1→C2 gap is
 - legacy absolute-margin rule (0.02): FAIL — C1→C2 gap +0.0146 < margin 0.02 (mining produced equally easy negatives) (reported for the record; not the verdict since v6)
 - **verdict: PASS** — order C1 < C2 <= C3 holds (0.9607 < 0.9753 <= 0.9885) and d(C1→C2) = 0.614 >= 0.5 (absolute gap +0.0146)
 
+### run 2026-09-28 16:35 UTC — version `phase01-v6`, rule: D1 scale-aware: C1 < C2 <= C3 and d(C1->C2) >= 0.5
+
+- C1: mean cos(anchor, negative) = 0.96068 (sd 0.026, n = 20000, anchors = 1000)
+- C2: mean cos(anchor, negative) = 0.97526 (sd 0.022, n = 20000, anchors = 1000)
+- C3: mean cos(anchor, negative) = 0.98845 (sd 0.007, n = 20000, anchors = 1000)
+- corpus percentile of the negative in its anchor's similarity ranking, median: C1 50.1 / C2 91.9 / C3 99.8  (50 = random, 100 = nearest); inside the anchor's top-20: C1 0.3% / C2 15.4% / C3 94.8%
+- embeddings: microsoft/graphcodebert-base max_len=512 n=8063 version=phase01-v5
+- d(C1→C2) = 0.614, 95% CI 0.795–1.202 (2000 resamples over anchors, not negatives)
+- legacy absolute-margin rule (0.02): FAIL — C1→C2 gap +0.0146 < margin 0.02 (mining produced equally easy negatives) (reported for the record; not the verdict since v6)
+- **verdict: PASS** — order C1 < C2 <= C3 holds (0.9607 < 0.9753 <= 0.9885) and d(C1→C2) = 0.614 >= 0.5 (absolute gap +0.0146)
+
 ## Gate G1 — diagnostics (scale of the space; no verdict)
 
 `python -m scripts.hardness_diagnostics` — version `phase01-v6`, 1600 anchors × k=20, corpus 6451. Percentile = share of corpus fragments less similar to the anchor than the negative (50 = random, 100 = nearest).
@@ -528,6 +539,7 @@ Label-noise floors (correction 5; proxies, not the audit): share of negatives th
 | C3 negatives | 0.00% | 0.05% | 0.40 | 31.4% | 6.9% |
 | labelled positives | — | — | 0.28 | 5.1% | 0.2% |
 
+
 ## False-negative audit (50 + 50, blind; SCOPE P2-9)
 
 `python -m scripts.audit_sample score` — version `phase01-v6`. Share of mined negatives judged to be clones (functional equivalence). Headline counts `unsure` as not_clone; the upper bound counts it as clone.

sync_manifest.json

@@ -1,6 +1,6 @@
 {
   "version": "phase01-v6",
-  "synced_at_utc": "2026-09-28T15:32:51Z",
+  "synced_at_utc": "2026-09-28T16:35:50Z",
   "staged_files": [
     "artifacts/audit/audit_key.csv",
     "artifacts/audit/audit_labels.csv",