experiments: freeze C2 calibration quality recovery

1 files changed, 49 insertions, 0 deletions

diff --git a/experiments/c2_calibration_quality_pilot.sh b/experiments/c2_calibration_quality_pilot.sh
new file mode 100755
index 0000000..d1c9291
--- /dev/null
+++ b/experiments/c2_calibration_quality_pilot.sh
@@ -0,0 +1,49 @@
+#!/usr/bin/env bash
+# Development-only C2 screen after the direct-NP diagnosis confirmed that
+# feedback amortization is the bottleneck.  Task seed 0 is reused strictly for
+# development; task seeds 6--8 are reserved for any independent confirmation.
+#
+# Frozen choices: context-gated vectorizer, eta=0.03, eta_A=0.01, layerwise
+# antithetic calibration every four steps.  Compare K4/K16 at d1/d4 over model
+# seeds 0--2.  The analyzer chooses the higher mean d4 endpoint, preferring K4
+# whenever it is within one point of the best finite candidate.
+# Usage: c2_calibration_quality_pilot.sh <gpu> "<model seeds>"
+set -eu
+
+cd "$(dirname "$0")/.."
+GPU="${1:?GPU index required}"
+MODEL_SEEDS="${2:-0 1 2}"
+PYTHON="${PYTHON:-/home/yurenh2/miniconda3/envs/ep_pascal/bin/python3}"
+
+for directions in 4 16; do
+  for model_seed in $MODEL_SEEDS; do
+    for depth in 1 4; do
+      tag="c2_calibdev_v1_tent_l2_w8_context_k${directions}_d${depth}_t0_s${model_seed}"
+      out="results/${tag}.json"
+      if [ -s "$out" ]; then
+        echo "[$tag] exists; skipping"
+        continue
+      fi
+      lesion=0
+      if [ "$depth" -eq 4 ]; then
+        lesion=0.3333333333333333
+      fi
+      CUDA_VISIBLE_DEVICES="$GPU" "$PYTHON" experiments/run.py \
+        --mode sdil --dataset tentmap --device cuda \
+        --depth "$depth" --width 8 --act relu --residual 1 \
+        --residual_lesion_fraction "$lesion" \
+        --vectorizer_mode context_gated \
+        --epochs 80 --batch_size 256 --eta 0.03 --momentum 0.9 \
+        --eta_A 0.01 --eta_P 0.002 \
+        --pert_sigma 0.01 --pert_every 4 --pert_ndirs "$directions" \
+        --pert_mode layerwise \
+        --traffic_mode none --nuis_rho 0 \
+        --use_residual 1 --learn_A 1 --learn_P 1 --p_neutral 1 \
+        --task_train_examples 10000 --task_test_examples 5000 \
+        --task_levels 2 --task_n_in 1 --task_seed 0 \
+        --val_examples 2000 --split_seed 2027 --eval_split validation --eval_every 0 \
+        --diagnostics alignment --diagnostics_schedule final --probe_bs 512 \
+        --seed "$model_seed" --log_every 100000 --tag "$tag"
+    done
+  done
+done