{
  "overlap_offset_s": 0.16666666666666705,
  "gap_offset_s": -0.22999999999999918,
  "backchannel_offset_s": 0.0,
  "engine_id": "maya-research/Veena",
  "status": "COMPLETE \u2014 overlap n_valid=6/10 (mean_bias=-0.167s, std=0.075s), gap n_valid=10/10 (mean_bias=+0.230s, std=0.046s). backchannel NOT calibrated (identity/0.0, deferred per plan M3b NOTE -- dropped from 20-dialog M4 gate anyway). Measured via fd_bench.extract_timings (Silero VAD, resampled to 16kHz) -- the SAME primitive the M4 gate scores with, not stitch.py's activity_mask.",
  "calibrated_at": "2026-07-11T02:13 IST",
  "device": "cuda",
  "n_target": 10,
  "acceptance_check": {
    "gap": {
      "target_s": 6.3,
      "placed_s": 6.07,
      "measured_s": 6.2,
      "rel_err": 0.016,
      "passed": true
    },
    "overlap": {
      "target_s": 0.2,
      "placed_s": 0.367,
      "measured_s": 0.1,
      "rel_err": 0.5,
      "passed": false
    },
    "overall_passed": false,
    "note": "overlap correction FAILED the +-20% acceptance bar on a fresh non-circular draw (target=0.2s, short-duration regime); likely censoring/survivorship bias in calibration -- short overlap draws had a 40% VAD-nan rate, so the fitted correction is based only on the subset VAD happened to detect, which may not represent the true population (esp. near VADs detection floor). gap correction passed cleanly (1.6% rel_err)."
  },
  "ladder_test": {
    "status": "NEVER RAN \u2014 0 draws",
    "reason": "GPU-3 continuously occupied by a rotating queue of other jobs for the full 60-min timebox and beyond (observed: u4_yield_ft 18000s, then demo_launcher 1200s, then personaplex_gate_launcher 5400s back-to-back) -- run_safe.sh lock is host-wide (one PID file), not per-GPU, so no idle-GPU workaround exists for watchdogged jobs. 60-iteration/10min retry-loop exhausted without ever acquiring the lock.",
    "floor_vs_compression": "UNRESOLVED \u2014 deferred to Sunday per team-lead ruling"
  }
}