sync EmbedEd artifacts (phase01-v5)

artifacts/hardness_diagnostics.json

@@ -0,0 +1,91 @@
+{
+  "n_anchors": 1600,
+  "k": 20,
+  "corpus_size": 6451,
+  "scale": {
+    "random_pair_cos": 0.9603,
+    "anchor_positive_cos_mean": 0.96512,
+    "anchor_positive_cos_sd": 0.02333
+  },
+  "conditions": {
+    "C1": {
+      "mean_cos": 0.96056,
+      "sd": 0.02585,
+      "gap_vs_C1": 0.0,
+      "d_vs_C1": 0.0,
+      "position_in_C1_C3_range": 0.0,
+      "percentile_mean": 50.02,
+      "percentile_median": 50.12,
+      "closer_than_positive_frac": 0.3957,
+      "n": 32000,
+      "in_anchor_top20_frac": 0.0029,
+      "in_anchor_top100_frac": 0.0147
+    },
+    "C2": {
+      "mean_cos": 0.97552,
+      "sd": 0.02126,
+      "gap_vs_C1": 0.01496,
+      "d_vs_C1": 0.632,
+      "position_in_C1_C3_range": 0.539,
+      "percentile_mean": 79.61,
+      "percentile_median": 91.92,
+      "closer_than_positive_frac": 0.715,
+      "n": 32000,
+      "in_anchor_top20_frac": 0.1537,
+      "in_anchor_top100_frac": 0.2825
+    },
+    "C3": {
+      "mean_cos": 0.98834,
+      "sd": 0.00774,
+      "gap_vs_C1": 0.02778,
+      "d_vs_C1": 1.456,
+      "position_in_C1_C3_range": 1.0,
+      "percentile_mean": 99.81,
+      "percentile_median": 99.83,
+      "closer_than_positive_frac": 0.9922,
+      "n": 32000,
+      "in_anchor_top20_frac": 0.9481,
+      "in_anchor_top100_frac": 0.9998
+    }
+  },
+  "negative_set_overlap_frac": {
+    "C1&C2": 0.0034,
+    "C1&C3": 0.0032,
+    "C2&C3": 0.1578
+  },
+  "label_noise": {
+    "C1": {
+      "n": 32000,
+      "train_labelled_clone_frac": 0.0,
+      "valid_labelled_clone_frac": 0.0002,
+      "jaccard_mean": 0.2253,
+      "jaccard_median": 0.2121,
+      "near_dup_frac_ge_0.5": 0.0093,
+      "near_dup_frac_ge_0.75": 0.0003
+    },
+    "C2": {
+      "n": 32000,
+      "train_labelled_clone_frac": 0.0,
+      "valid_labelled_clone_frac": 0.0004,
+      "jaccard_mean": 0.4166,
+      "jaccard_median": 0.3731,
+      "near_dup_frac_ge_0.5": 0.2799,
+      "near_dup_frac_ge_0.75": 0.0664
+    },
+    "C3": {
+      "n": 32000,
+      "train_labelled_clone_frac": 0.0,
+      "valid_labelled_clone_frac": 0.0005,
+      "jaccard_mean": 0.4306,
+      "jaccard_median": 0.3974,
+      "near_dup_frac_ge_0.5": 0.3138,
+      "near_dup_frac_ge_0.75": 0.0688
+    },
+    "positives": {
+      "jaccard_mean": 0.3024,
+      "jaccard_median": 0.2838,
+      "near_dup_frac_ge_0.5": 0.0512,
+      "near_dup_frac_ge_0.75": 0.0019
+    }
+  }
+}

report/measurements.md

@@ -2,7 +2,7 @@
 
 ## Phase 0 — measured numbers
 
-generated by `python -m embeded.data.prepare_data` — version `phase01-v2`
+generated by `python -m embeded.data.prepare_data` — version `phase01-v5`
 
 | quantity | spec value | measured | ok |
 |---|---|---|---|
@@ -16,20 +16,21 @@ generated by `python -m embeded.data.prepare_data` — version `phase01-v2`
 
 **Token lengths** (microsoft/graphcodebert-base): p50=474 p90=1535 p95=2233 p99=5944 max=36823 → **max_len = 512**.
 
+
 ## Phase 0.5 — throughput (measured, fixes BATCH / TRAIN_PAIRS_CAP / EPOCHS)
 
-generated by `python -m scripts.phase0_throughput --candidates 512:4 512:8 512:16 512:32 --minutes 0.5` — version `phase01-v5`, recipe `AMP_DTYPE=float16`, `GRAD_CHECKPOINT=False` (Colab Tesla T4, 14.5 GB, 2026-09-26)
+generated by `python -m scripts.phase0_throughput --candidates 512:4 512:8 512:16 512:32 --minutes 0.5` — version `phase01-v5`, recipe `AMP_DTYPE=float16`, `GRAD_CHECKPOINT=False`
 
 | max_len:batch (triples) | triples/s | peak GB | cap for one 40-min epoch |
 |---|---|---|---|
-| 512:4 | 11.9 | 4.37 | 28,648 |
-| 512:8 | 13.5 | 6.71 | 32,367 |
-| 512:16 | 14.1 | 11.39 | 33,951 |
+| 512:4 | 11.7 | 4.37 | 28,048 |
+| 512:8 | 13.0 | 6.71 | 31,268 |
+| 512:16 | 12.9 | 11.39 | 30,937 |
 | 512:32 | OOM on Tesla T4 | — | — |
 
-Suggested: **BATCH = 8, TRAIN_PAIRS_CAP = 32,367, EPOCHS = 1** (rule: fastest-within-10% with the lowest peak memory; 13.5 vs top 14.1 triples/s). `settings.py` at run time: BATCH = 8, TRAIN_PAIRS_CAP = 32,000, EPOCHS = 1.
+Suggested: **BATCH = 8, TRAIN_PAIRS_CAP = 31,268, EPOCHS = 1** (rule: fastest-within-10% with the lowest peak memory; 13.0 vs top 13.0 triples/s). `settings.py` at run time: BATCH = 8, TRAIN_PAIRS_CAP = 32,000, EPOCHS = 1.
+
 
-Frozen in `settings.py` as `TRAIN_PAIRS_CAP = 32_000` (= 1,600 anchors × k=20 = 4,000 steps of 8; measured cap rounded down). Context: the same test in plain fp32 with the original candidates OOMed on its first candidate (256:32 = 96 sequences/step, 14.3 GB), which is why the training recipe is fp16 AMP.
 
 ## s′ acquisition (Phase 0.7)
 
@@ -331,8 +332,8 @@ Frozen in `settings.py` as `TRAIN_PAIRS_CAP = 32_000` (= 1,600 anchors × k=20 =
         "label"
       ],
       "label_counts": {
-        "4": 20000,
         "5": 20000,
+        "4": 20000,
         "0": 20000,
         "3": 18317,
         "1": 15555,
@@ -468,3 +469,34 @@ Frozen in `settings.py` as `TRAIN_PAIRS_CAP = 32_000` (= 1,600 anchors × k=20 =
 - **verdict: FAIL** — C1→C2 gap +0.0146 < margin 0.02 (mining produced equally easy negatives)
 
 Context recorded at the time: ordering C1 < C2 ≤ C3 holds; the C1→C2 gap is ≈70 standard errors (n = 20,000) and d ≈ 0.6, but the full C1→C3 range of this untuned, anisotropic space is only 0.028, so the absolute 0.02 margin (set in `settings.py` before the spread was known) asks C2 to cover ~75% of it. Per SCOPE P6-3 training is blocked; next step is `python -m scripts.hardness_diagnostics` (percentile-rank view), then either a mining fix or an explicit, documented re-operationalisation of "visible gap" — decided before any training, with this run kept here either way.
+
+### run 2026-09-27 06:55 UTC — version `phase01-v5`, margin 0.02
+
+- C1: mean cos(anchor, negative) = 0.96068 (sd 0.026, n = 20000)
+- C2: mean cos(anchor, negative) = 0.97526 (sd 0.022, n = 20000)
+- C3: mean cos(anchor, negative) = 0.98845 (sd 0.007, n = 20000)
+- **verdict: FAIL** — C1→C2 gap +0.0146 < margin 0.02 (mining produced equally easy negatives)
+
+## Gate G1 — diagnostics (scale of the space; no verdict)
+
+`python -m scripts.hardness_diagnostics` — version `phase01-v5`, 1600 anchors × k=20, corpus 6451. Percentile = share of corpus fragments less similar to the anchor than the negative (50 = random, 100 = nearest).
+
+Scale: random corpus pair cos = **0.9603**; cos(anchor, positive) = **0.9651** (sd 0.023).
+
+| cond | mean cos | sd | gap vs C1 | d vs C1 | position C1→C3 | pct mean | pct median | in top-20 | in top-100 | closer than positive |
+|---|---|---|---|---|---|---|---|---|---|---|
+| C1 | 0.9606 | 0.026 | +0.0000 | 0.00 | 0.00 | 50.0 | 50.1 | 0.3% | 1.5% | 39.6% |
+| C2 | 0.9755 | 0.021 | +0.0150 | 0.63 | 0.54 | 79.6 | 91.9 | 15.4% | 28.2% | 71.5% |
+| C3 | 0.9883 | 0.008 | +0.0278 | 1.46 | 1.00 | 99.8 | 99.8 | 94.8% | 100.0% | 99.2% |
+
+Negative-set overlap per anchor: C1&C2 = 0.3%, C1&C3 = 0.3%, C2&C3 = 15.8%
+
+Label-noise floors (correction 5; proxies, not the audit): share of negatives that the valid split labels as a clone of the anchor, and share whose BM25 token set is near-identical to the anchor's (Jaccard). Positives shown with the same yardstick.
+
+| set | train-labelled clone (must be 0) | valid-labelled clone | Jaccard median | ≥ 0.5 | ≥ 0.75 |
+|---|---|---|---|---|---|
+| C1 negatives | 0.00% | 0.02% | 0.21 | 0.9% | 0.0% |
+| C2 negatives | 0.00% | 0.04% | 0.37 | 28.0% | 6.6% |
+| C3 negatives | 0.00% | 0.05% | 0.40 | 31.4% | 6.9% |
+| labelled positives | — | — | 0.28 | 5.1% | 0.2% |
+

sync_manifest.json

@@ -1,11 +1,12 @@
 {
   "version": "phase01-v5",
-  "synced_at_utc": "2026-09-27T06:54:55Z",
+  "synced_at_utc": "2026-09-27T06:55:56Z",
   "staged_files": [
     "artifacts/codexglue_data.jsonl",
     "artifacts/corpus_emb.meta.json",
     "artifacts/corpus_emb.npy",
     "artifacts/fragments.jsonl",
+    "artifacts/hardness_diagnostics.json",
     "artifacts/manifest.json",
     "artifacts/mining_summary.json",
     "artifacts/overlap.json",
@@ -2144,6 +2145,7 @@
     "artifacts/token_lengths.json",
     "artifacts/triples_C1.jsonl",
     "artifacts/triples_C2.jsonl",
-    "artifacts/triples_C3.jsonl"
+    "artifacts/triples_C3.jsonl",
+    "report/measurements.md"
   ]
 }