scripts/run_v02_n60.sh

#!/usr/bin/env bash
# v0.2.0 n=60 harness on a CUDA GPU host. Runs the decisive cells sequentially:
#   1. v02_e6_plain   — full v0.2 ensemble, M4 feature-target ON  (max transfer)
#   2. v02_e6_noM4    — same, M4 feature-target OFF               (ablation: does steering earn its keep?)
#   3. v02_e6_perc    — M4 ON + LPIPS/DCT perceptual budget       (stealth arm)
# Each writes runs/<cell>/{adv,results.json}. Cross-arch held-out (dinov2 SSL +
# internvit VLM) is scored via M4 centroids, so held-out flips mean cross-architecture transfer.
set -u
cd ~/workspace/veil-pgd
. .venv/bin/activate
export HF_HOME=~/workspace/hf-cache
export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True

COMMON="--manifest examples/testset60.csv --images examples/testset60 --train v0.2 \
  --steps 120 --eps 6 --subset 6 --grad-norm --max-per-family 2 --min-feature 1 --metrics"

echo "=== [$(date +%H:%M:%S)] CELL 1/3 v02_e6_plain (M4 ON) ==="
python -u -m ensemble.run_attack $COMMON --out runs/v02_e6_plain

echo "=== [$(date +%H:%M:%S)] CELL 2/3 v02_e6_noM4 (M4 OFF ablation) ==="
python -u -m ensemble.run_attack $COMMON --out runs/v02_e6_noM4 --no-feature-target

echo "=== [$(date +%H:%M:%S)] CELL 3/3 v02_e6_perc (M4 ON + perceptual) ==="
python -u -m ensemble.run_attack $COMMON --out runs/v02_e6_perc \
  --lpips-weight 0.3 --lpips-tau 0.08 --dct-keep 0.6

echo "=== [$(date +%H:%M:%S)] ALL CELLS DONE ==="