#!/usr/bin/env bash
# Final bounded task-seed-0 C2 recovery screen. The context vectorizer uses
# normalized LMS in both a 100-step A-only layerwise-K4 prefix and the original
# simultaneous-K1/e4 joint phase. Compare eta_A 0.01/0.05 only.
# Usage: c2_nlms_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 eta_a in 0.01 0.05; do
  case "$eta_a" in
    0.01) eta_tag=0p01 ;;
    0.05) eta_tag=0p05 ;;
    *) echo "unexpected eta_A: $eta_a" >&2; exit 2 ;;
  esac
  for model_seed in $MODEL_SEEDS; do
    for depth in 1 4; do
      tag="c2_nlmsdev_v1_tent_l2_w8_context_e${eta_tag}_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 --vectorizer_optimizer nlms \
        --epochs 80 --batch_size 256 --eta 0.03 --momentum 0.9 \
        --eta_A "$eta_a" --eta_P 0.002 \
        --pert_sigma 0.01 --pert_every 4 --pert_ndirs 1 \
        --pert_mode simultaneous \
        --a_warmup_steps 100 --a_warmup_ndirs 4 --a_warmup_mode layerwise \
        --traffic_mode none --nuis_rho 0 \
        --use_residual 1 --learn_A 1 --learn_P 1 --p_neutral 1 \
        --task_train_examples 10000 --task_test_examples 5000 \
        --task_levels 2 --task_n_in 1 --task_seed 0 \
        --val_examples 2000 --split_seed 2027 --eval_split validation --eval_every 0 \
        --diagnostics alignment --diagnostics_schedule final --probe_bs 512 \
        --seed "$model_seed" --log_every 100000 --tag "$tag"
    done
  done
done