configs/eval/experiment1_long_delay.json

{
 "evaluations": [
  {
   "batch": 12,
   "intervention": "none",
   "run": "PNSR_K4_fin",
   "script": "eval/evaluate.py",
   "split": "val_4k"
  },
  {
   "batch": 12,
   "intervention": "none",
   "run": "TX768_fin",
   "script": "eval/evaluate.py",
   "split": "val_4k"
  },
  {
   "batch": 12,
   "intervention": "none",
   "run": "TXE_fin",
   "script": "eval/evaluate.py",
   "split": "val_4k"
  },
  {
   "batch": 12,
   "intervention": "none",
   "run": "PNSR_K1_fin",
   "script": "eval/evaluate.py",
   "split": "val_4k"
  }
 ],
 "purpose": "horizon extrapolation: L=4096 is 16x the trained horizon. Per-delay-bucket accuracy comes from the same runs (delay is recorded per row)."
}