sync EmbedEd artifacts (phase01-v8)

artifacts/phase4_decisions.json

@@ -0,0 +1,22 @@
+{
+  "recorded_utc": "2026-10-21 04:44 UTC",
+  "version": "phase01-v8",
+  "decisions": {
+    "D4-1": {
+      "choice": "three panels inline (benchmark / generalisation / FN rates) + non-comparability note; C4-C6 inline, flagged as the recorded factorial extension",
+      "rationale": "Part 7 names C0-C3; the bouncer trio joined by the recorded 2026-10-03 addendum - visible, not smuggled in."
+    },
+    "D4-2": {
+      "choice": "headline condition = C1; C5 only if audit_c5 lands in band",
+      "rationale": "C1 is the best *trained* condition with its gate closed; C5's audit is the open one (section 3)."
+    },
+    "D4-3": {
+      "choice": "the ~0.95 published number is framed as a pair-regime difference (reconciliation as a limitation)",
+      "rationale": "Corpus-wide natural prevalence (ceiling 0.2406) vs the near-clone-heavy published splits; MAP@R carries the prevalence-robust companion. Not chased as a harness defect."
+    },
+    "D4-4": {
+      "choice": "the denoising delta-ladder on C3, ONLY after a completed draft AND a SCOPE amendment",
+      "rationale": "SCOPE's own 'natural next experiment' (Part 6 rule 7 keeps the bar: a clear 9-page write-up beats a half-finished extension)."
+    }
+  }
+}

report/measurements.md

@@ -640,8 +640,8 @@ Generated from `artifacts/runs/<condition>_<seed>/eval_metrics.json` by `python
 - C0 is the untuned baseline (one evaluation, not a trained condition) at mean F1 0.2572.
 - Among the trained conditions the highest mean F1 is C1 (0.7464); per-seed values above, not the best seed.
 - Sanity reference: published CodeXGLUE fine-tuned-CodeBERT F1 ~ 0.95 under a different model/protocol. Observed F1 range 0.2408–0.7623. A large discrepancy is a harness/protocol check (metric implementation, pair order, labels, threshold handling), not a reason to tune against the test split.
-- 3 smoke-run result(s) are excluded from this table (harness checks, not measurements).
-- 2 result file(s) with conditions outside ['C0', 'C1', 'C2', 'C3', 'C4', 'C5', 'C6'] (Phase 3 sprime runs) are excluded from this table.
+- 4 smoke-run result(s) are excluded from this table (harness checks, not measurements).
+- 37 result file(s) with conditions outside ['C0', 'C1', 'C2', 'C3', 'C4', 'C5', 'C6'] (Phase 3 sprime runs) are excluded from this table.
 
 ## Phase 3 — decisions (recorded before the runs)
 
@@ -764,3 +764,48 @@ Phase 2 checkpoints (C0 = base model) scored on the same s′ partition without
 | C6 | 13 | 0.6690 | 0.6338 | +0.0352 | 0.7480 | 600 |
 | C6 | 14 | 0.6701 | 0.6659 | +0.0042 | 0.6374 | 600 |
 | C6 | 15 | 0.6760 | 0.6592 | +0.0168 | 0.6970 | 600 |
+
+## False-negative audit — C2/C3 (blind; SCOPE P2-9)
+
+`python -m scripts.audit_sample score` — version `phase01-v8`, conditions C2/C3. Share of mined negatives judged to be clones (functional equivalence). Headline counts `unsure` as not_clone; the upper bound counts it as clone.
+
+| condition | n | clone | not_clone | unsure | FN rate [95% CI] | upper bound incl. unsure |
+|---|---|---|---|---|---|---|
+| C2 | 50 | 21 | 29 | 0 | **42%** [29%, 56%] | 42% [29%, 56%] |
+| C3 | 50 | 15 | 34 | 1 | **30%** [19%, 44%] | 32% [21%, 46%] |
+
+## Phase 4 — the one table
+
+Assembled from the run dirs by the Phase-4 notebook — version `phase01-v8`, recorded 2026-10-21 04:44 UTC. Sources: `runs/<run>/eval_metrics.json`, the audit sections below, and the s′ table. **The two panels are different rulers** (benchmark: natural 13.7% prevalence, all-negative ceiling 0.2406; s′: balanced slices) — absolute F1 is NOT comparable across panels; within a panel, orderings are the finding.
+
+### Panel A — benchmark (F1 mean ± sd over seeds 13-15; pinned = validation thresholds saturated at 1.0)
+
+| condition | F1 | pinned | audited/measured FN rate |
+|---|---|---|---|
+| C0 | 0.2572 ± 0.0000 | 0/1 | — |
+| C1 | 0.7533 ± 0.0073 | 0/3 | ≈0-2% (audited) |
+| C2 | 0.2542 ± 0.0108 | 2/3 | 42% (audited) |
+| C3 | 0.2728 ± 0.0136 | 3/3 | 30% [0.191, 0.438] (audited) |
+| C4 | 0.4208 ± 0.0284 | 0/3 | 30% [0.191, 0.438] (audited) |
+| C5 | 0.7288 ± 0.0029 | 0/3 | 0.27% supply (audit_c5: §3) |
+| C6 | 0.4688 ± 0.0278 | 0/3 | 13.6% supply (audit_c6: §3) |
+
+*(Phase-2 extension seeds 16-18 exist on the hub and are reported as extra evaluations, not table rows.)*
+
+### Panel B — s′ generalisation (F1_seen / F1_unseen / Δ, seeds 13-15)
+
+| condition | F1_seen | F1_unseen | Δ |
+|---|---|---|---|
+| C1 | 0.7570 | 0.7031 | +0.0539 |
+| C2 | 0.6781 | 0.6674 | +0.0107 |
+| C3 | 0.7004 | 0.7057 | -0.0053 |
+| C4 | 0.6731 | 0.6687 | +0.0044 |
+| C5 | 0.7606 | 0.7244 | +0.0362 |
+| C6 | 0.7169 | 0.6865 | +0.0304 |
+| C0 (untuned anchor, transfer) | 0.7118 | 0.6934 | — |
+
+**The floor:** no poisoned-mining condition beats the untuned encoder on unseen functionality; only clean training (C1, and C5 pending its audit branch) converts fine-tuning into gains.
+
+### Panel C — mechanism bookkeeping
+
+Trigger = an FN that is simultaneously a functional clone and a J ≥ 0.40 near-duplicate; rate = dose; margin 0.10 = amplifier (§9.4 FINAL). The 40% cut removes ~40% of C2 contamination and 0 audited clean pairs. audit_c4 PASSED (30% ≤ 42%); audit_c5/c6 outcomes: see section 3 — no C5/C6 headline before its branch is recorded.

sync_manifest.json

@@ -1,6 +1,6 @@
 {
   "version": "phase01-v8",
-  "synced_at_utc": "2026-10-21T04:43:59Z",
+  "synced_at_utc": "2026-10-21T04:44:49Z",
   "staged_files": [
     "artifacts/audit/audit_key.csv",
     "artifacts/audit/audit_labels.csv",
@@ -27,6 +27,7 @@
     "artifacts/pairs_train.tsv",
     "artifacts/pairs_valid.tsv",
     "artifacts/phase3_decisions.json",
+    "artifacts/phase4_decisions.json",
     "artifacts/runs/C0_13/eval_metrics.json",
     "artifacts/runs/C1_13/eval_metrics.json",
     "artifacts/runs/C1_13/metrics.json",
@@ -298,7 +299,8 @@
     "artifacts/triples_C3.jsonl",
     "artifacts/triples_C4.jsonl",
     "artifacts/triples_C5.jsonl",
-    "artifacts/triples_C6.jsonl"
+    "artifacts/triples_C6.jsonl",
+    "report/measurements.md"
   ],
   "excluded_patterns": [
     "sprime/*",