{
 "schema_version": 1,
 "evidence_id": "paired-loading-2026-09-09",
 "date": "2026-09-09",
 "qualification": "physical device measurement; single run; not a release gate",
 "device": "NVIDIA RTX PRO 6000 Blackwell, CUDA, one device; drives: PCIe 4.0 NVMe SSD for the original HDF5 (direct-read ceiling 5.5 GB/s measured separately), a second NVMe SSD for the saved form",
 "provenance": {
  "package": "quantem.gpu",
  "branch": "main",
  "parent_commit": "2876b7fb",
  "loader": "quantem.gpu.io.PairedLoader defaults (6 readers, 10 x 128 MiB pinned slots, 3 x 8192-scan rolling buffers)"
 },
 "input": {
  "acquisitions_offered": 69,
  "shape": [
   512,
   512,
   192,
   192
  ],
  "dtype": "uint16",
  "source": "complete bitshuffle+LZ4 HDF5 masters, 27 shards each",
  "original_bytes_per_acquisition_approx": 2190000000.0
 },
 "protocol": "io.load-equivalent PairedLoader.load_many over the acquisition list with a memory-budget admit callback; the first three to sixty-six acquisitions were admitted, the sixty-seventh was refused for memory",
 "rows": [
  {
   "measurement": "series load, original bitshuffle+LZ4 HDF5 to paired resident",
   "acquisitions": 66,
   "acquisitions_offered": 69,
   "value_seconds": 29.34,
   "statistic": "single run wall",
   "boundary": "first shard read submitted to last source resident (consumer stream synchronized)",
   "cache_state": "shards read with direct I/O (page cache bypassed); master metadata pages may be warm",
   "admission": "1.31 GiB per source plus 1 GiB margin against free device memory plus cached pool bytes"
  },
  {
   "measurement": "per-acquisition resident-ready wall inside the pipeline",
   "acquisitions": 66,
   "value_seconds": 1.498,
   "statistic": "median",
   "boundary": "first shard read submitted to source resident"
  },
  {
   "measurement": "per-acquisition encode",
   "acquisitions": 66,
   "value_seconds": 0.197,
   "statistic": "median",
   "boundary": "device encode kernels, consumer stream synchronized"
  },
  {
   "measurement": "per-acquisition index",
   "acquisitions": 66,
   "value_seconds": 0.12,
   "statistic": "median",
   "boundary": "device index kernels, consumer stream synchronized"
  },
  {
   "measurement": "resident bytes, all sources",
   "acquisitions": 66,
   "value_bytes": 93973719027,
   "memory_kind": "sum of resident arrays (payload, offsets, models, index)"
  },
  {
   "measurement": "index bytes, all sources",
   "acquisitions": 66,
   "value_bytes": 7046943828,
   "memory_kind": "sum of index arrays"
  },
  {
   "measurement": "device bytes in use after load",
   "acquisitions": 66,
   "value_bytes": 101593579520,
   "memory_kind": "cudaMemGetInfo total minus free (includes pool cache and loader staging)"
  },
  {
   "measurement": "detector.prepare over all sources",
   "acquisitions": 66,
   "value_seconds": 0.0379,
   "statistic": "single run wall"
  },
  {
   "measurement": "first virtual image, all sources, bright-field mask",
   "acquisitions": 66,
   "value_seconds": 0.0082,
   "statistic": "single run wall",
   "parity": "first source digest equal to the frozen independent raw-count reduction"
  },
  {
   "measurement": "first diffraction pattern, all sources",
   "acquisitions": 66,
   "value_seconds": 0.0019,
   "statistic": "single run wall"
  },
  {
   "measurement": "save one source as the paired resident form",
   "acquisitions": 1,
   "value_seconds": 1.387,
   "value_bytes": 1366029888,
   "statistic": "single run wall",
   "boundary": "write plus fsync to the root drive"
  },
  {
   "measurement": "reopen one saved form through io.load",
   "acquisitions": 1,
   "value_seconds": 0.527,
   "statistic": "single run wall",
   "boundary": "header read to resident, byte-identical to the streamed source",
   "cache_state": "file written seconds earlier; root drive"
  },
  {
   "measurement": "series load, original HDF5 to paired resident, chunk tables parsed from the staged image (revision 91d6f5b6)",
   "acquisitions": 66,
   "acquisitions_offered": 69,
   "value_seconds": 28.82,
   "statistic": "single run wall",
   "boundary": "first shard read submitted to last source resident",
   "cache_state": "shards read with direct I/O; no metadata reads during streaming",
   "note": "0.437 s per acquisition against a 0.40 s direct-read floor for 2.19 GB per acquisition at 5.5 GB/s"
  }
 ]
}