diff --git a/submissions/inbox/LLM_Claude_Opus_47/run_claude_47/metadata.json b/submissions/inbox/LLM_Claude_Opus_47/run_claude_47/metadata.json
new file mode 100644
index 0000000..0dee6ed
--- /dev/null
+++ b/submissions/inbox/LLM_Claude_Opus_47/run_claude_47/metadata.json
@@ -0,0 +1,8 @@
+{
+  "team": "LLM_Claude_Opus_47",
+  "run_id": "run_claude_47",
+  "model_name": "GraphSAGE (zero-shot Claude Opus 4.7, agent mode, temperature=0)",
+  "model_type": "llm-only",
+  "submitter": "LLM_Claude_Opus_47",
+  "notes": "GNN-CB LLM baseline. Generated by Claude Opus 4.7 (claude-opus-4-7) under frozen prompt template prompt.md v1.1 (temperature=0, top_p=1, seed=42). Agent mode (Claude Code). Zero few-shot examples, zero per-task hints, zero prompt edits, zero repair iterations. Inductive cross-domain task: train on 209 cfRNA samples, predict on 111 placenta samples (no labels seen during training). Pipeline: pre-built graph_artifacts.pt (HeteroData, 320 nodes total, 6,650 gene-expression features, 3,200 similarity edges); 2-layer GraphSAGE (mean aggregator) hidden=128 + BatchNorm + dropout=0.5; CrossEntropyLoss with inverse-frequency class weights for the 2.1:1 imbalance (142 control / 67 preeclampsia); Adam lr=5e-3 wd=5e-4 full-batch transductive forward pass on the combined graph (supervision restricted to cfRNA nodes); stratified internal 80/20 val split on cfRNA labels; early stop on val F1 patience=20 max 200 epochs; 5-seed soft-vote ensemble {42..46} averaging softmax probabilities. Local internal-val F1: 0.5366 (ensemble; P=0.39 R=0.85), 0.4275 +/- 0.1541 per-seed (cross-seed variance is high due to tiny minority class of ~13 val samples). CPU only, wall-clock ~10 s."
+}
diff --git a/submissions/inbox/LLM_Claude_Opus_47/run_claude_47/predictions.csv.enc b/submissions/inbox/LLM_Claude_Opus_47/run_claude_47/predictions.csv.enc
new file mode 100644
index 0000000..414bfe7
Binary files /dev/null and b/submissions/inbox/LLM_Claude_Opus_47/run_claude_47/predictions.csv.enc differ