#!/usr/bin/env bash
# Run colony experiments back to back so the GPU never idles.
#   scripts/run_queue.sh   (edit the QUEUE below)
#
# Each entry: "<run_name>|<width>|<K>|<extra hydra overrides>"
# NOTE: deliberately not `set -e` — one failing run must not cancel the rest of
# the queue (a bad override once killed the whole queue in seconds).
set -uo pipefail
cd "$(dirname "$0")/.."

QUEUE=(
  # THE CONTROL that makes the flagship result interpretable: one unit, same
  # half-evidence, no colony, fully trained. If the colony's 9.8% is real
  # information pooling, this must land far below it.
  "solo-d64-evihalf|64|1|arch.blackboard=latent arch.standpoints=none arch.aggregate=mean arch.halt_rule=boss arch.view_fraction=0.5 arch.view_enforce_coverage=False"
  # Ablation: unit differentiation without distributed evidence.
  "colony-d64-k4-roles|64|4|arch.blackboard=latent arch.standpoints=sudoku arch.aggregate=mean arch.halt_rule=quorum arch.role_embeddings=True"
)

for entry in "${QUEUE[@]}"; do
  IFS='|' read -r name width k overrides <<< "$entry"
  echo "=== $(date -Is) starting ${name} ==="
  # shellcheck disable=SC2086
  PYTHONUNBUFFERED=1 PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True \
  .venv/bin/python -m dsem.pretrain \
    arch=colony \
    data_paths="[data/sudoku-extreme-1k-aug-1000]" \
    data_paths_test="[data/sudoku-testsub-12k]" \
    evaluators="[]" \
    epochs=50000 eval_interval=5000 \
    global_batch_size=768 \
    lr=1e-4 puzzle_emb_lr=1e-4 weight_decay=1.0 puzzle_emb_weight_decay=1.0 \
    arch.mlp_t=True arch.pos_encodings=none \
    arch.L_layers=2 arch.H_cycles=3 arch.L_cycles=6 \
    "arch.hidden_size=${width}" "arch.num_agents=${k}" \
    ema=True "+run_name=${name}" \
    $overrides > "runs/${name}.out" 2>&1 < /dev/null
  echo "=== $(date -Is) finished ${name} ==="
done
echo "=== $(date -Is) queue complete ==="