Blame for IRREPLACEABLE.json

{
 "artefacts": [
  {
   "path": "phase_battery",
   "class": "impossible",
   "why": "the battery the whole sweep is recorded against. Its per-prompt seeds came from `abs(hash((twin, core_id, cell_key)))`, salted per process, so the seeds in CONTENT_HASHES.json cannot regenerate the prompts they name. 188 constructs, 964 sets, seven builds of amendment work behind it.",
   "verified": true,
   "local": true,
   "bytes": 77890544,
   "files": 966,
   "sha256": "ee486c2ee9de93c2d43ad9cad0189029e14d02e5fbe0cd3facf320afdb182128",
   "in_dataset": true
  },
  {
   "path": "battery",
   "class": "impossible",
   "why": "the pilot battery, 30 constructs. Same seed defect, and every pilot figure quoted before 2 Oct -- the ablation pilot, the first twin units, C0.5's first reading -- is against it.",
   "verified": true,
   "local": true,
   "bytes": 13409138,
   "files": 220,
   "sha256": "4786fff98eb575fcc978c36b3c51a6e74cf8738432be794ed017b19e0afe5b78",
   "in_dataset": true
  },
  {
   "path": "probe/companions.json",
   "class": "impossible",
   "why": "the companion table check 10 reads. `audit_companions.py` assembles its prompts with `seed=abs(hash((core.core_id, cell.key())))`, so a rebuild gives a different table. Was on the laptop alone until 6 Oct.",
   "verified": true,
   "local": true,
   "bytes": 26378,
   "files": 1,
   "sha256": "02ef23f6863be26c573b449fd339f68ec291c8c0bc85fc559315b4847c2eea3e",
   "in_dataset": true
  },
  {
   "path": "probe/bpool.json",
   "class": "impossible",
   "why": "the B pool every A/B contrast is read against. `probe_bpool.py` assembles it with `seed=abs(hash((core.core_id, cell.key())))`, salted per process.",
   "verified": true,
   "local": true,
   "bytes": 1842,
   "files": 1,
   "sha256": "c75dea213a74cc802ee9876dc44a424c6ccd33a5d1bc72abb9cd7ef94794a1b0",
   "in_dataset": true
  },
  {
   "path": "probe/access.json",
   "class": "impossible",
   "why": "the access labelling. `relabel_access.py` built it from `abs(hash(tuple_of_strings))`, salted per process, and says in its own docstring that a rebuild produces a different labelling.",
   "verified": true,
   "local": true,
   "bytes": 9070,
   "files": 1,
   "sha256": "bbdf936b5553cb7703835f566dd477e2431eea0ecc825b061c64d9eaacfbfbf6",
   "in_dataset": true
  },
  {
   "path": "probe/caps.json",
   "class": "cpu",
   "why": "the companion cap, fitted per read site on one twin and audited on the other. `check_caps.py` rebuilds it from the battery, so it is as safe as the battery is -- which is to say not at all. Whether the fit itself is deterministic was not checked.",
   "verified": false,
   "local": true,
   "bytes": 48403,
   "files": 1,
   "sha256": "c00aedced96db13342f883308675a272c47382d05a0ed63ab50523eaf767cf39",
   "in_dataset": true
  },
  {
   "path": "probe/probe.json",
   "class": "impossible",
   "why": "the machine probe: tokens/s with the 24 pre-hooks live, peak memory per batch size, disk throughput. Its own docstring says it measures the machine and not the design, and that machine is gone, so it cannot be re-measured -- but **nothing reads it**. Kept as the schedule's provenance, not as an input to any reading.",
   "verified": true,
   "local": true,
   "bytes": 1649,
   "files": 1,
   "sha256": "9f9df523f6021b71c3bbfd80fc007b92945fdea43901e39d3b04abb1e17ce60b",
   "in_dataset": true
  },
  {
   "path": "probe/probe_a100.json",
   "class": "impossible",
   "why": "the same probe on the A100 the sweep was costed against, and the source of the 6.85 units/h the budget rests on. That session is gone and a new one is a different machine; nothing downstream reads it either.",
   "verified": true,
   "local": true,
   "bytes": 1876,
   "files": 1,
   "sha256": "8425f920e209660186812cd56f2edd561fb173934a7fdc957417746a9e5e6b00",
   "in_dataset": true
  },
  {
   "path": "calibration",
   "class": "impossible",
   "why": "the A calibration, carried through four amendments (A15, A16 and their two corrections). Depends on the battery, so it inherits the battery's class whatever its own generator does.",
   "verified": false,
   "local": true,
   "bytes": 968430,
   "files": 26,
   "sha256": "845f1ef40a89743fedab55cff3d1b8b69cddfcfc8a191dc04859c51d93a42309",
   "in_dataset": true
  },
  {
   "path": "inventory",
   "class": "cpu",
   "why": "the checkpoint inventory, from the released repository's listing. Rebuildable while OpenCoder's checkpoints stay published -- which is a dependency on somebody else.",
   "verified": false,
   "local": true,
   "bytes": 129461,
   "files": 3,
   "sha256": "745cd8f69d93efbe485c695900e52906b63c786eba94f747bf07b33528d3c0c9",
   "in_dataset": true
  },
  {
   "path": "validation/_archive_pilot_suites",
   "class": "cpu",
   "why": "the eight pilot SUITE.json files, archived 6 Oct before the closure run could overwrite one. `run_validation` is CPU over the pilot record, so these are rebuildable **only while the pilot record exists** -- which item 33 is deciding.",
   "verified": true,
   "local": true,
   "bytes": 2958289,
   "files": 8,
   "sha256": "a5f8767920156b7797e5cfae432a227b013e0b6e1adda680cd777e7efde85aa9",
   "in_dataset": true
  },
  {
   "path": "validation/battery/SPLIT.json",
   "class": "deterministic",
   "why": "the P1 split. Seed 20261002 is in the file and `split_p1.py` is deterministic given the battery -- so this is safe exactly as far as the battery is.",
   "verified": true,
   "local": true,
   "bytes": 6614,
   "files": 1,
   "sha256": "8f9ecca8079a81667fd6db6b86210c5755f02da78482ccc9229bcde6786456ae",
   "in_dataset": true
  },
  {
   "path": "validation/schedule/CHECKPOINTS.json",
   "class": "deterministic",
   "why": "the 61 references in D5's order, with the dense block, the tail and the annealing branch -- `checkpoint_schedule.py` rebuilds it from the inventory, and the sweep, `hf_status --plan` and item 111 all read it as the definition of what the record should contain.",
   "verified": true,
   "local": true,
   "bytes": 7816,
   "files": 1,
   "sha256": "67707ce6262dca51d9659288c1ba835c177da779933f9c6eb234447dda7d4d8f",
   "in_dataset": true
  },
  {
   "path": "PROVENANCE.md",
   "class": "cpu",
   "why": "generated by `build_provenance.py` from the artefacts themselves, so it is rebuildable while they are. Note item 36: its `edf38a8f...` is the hash of this file rather than of the record it claims to identify.",
   "verified": true,
   "local": true,
   "bytes": 33589,
   "files": 1,
   "sha256": "edf38a8ff0b0042a93141af019e2c47b0c19e45bc1930f784d768f349e456e1d",
   "in_dataset": true
  },
  {
   "path": "smoke",
   "class": "impossible",
   "why": "a small battery build in its own right -- twin1, twin2, alt, CONTENT_HASHES.json and the SMOKE_COUNTS.json the first scope-use counts were read from. Built by the generator, so it carries the same salted-`hash()` seeds and cannot be rebuilt either; small enough that keeping it costs nothing.",
   "verified": true,
   "local": true,
   "bytes": 463884,
   "files": 65,
   "sha256": "071c5a2b4be7f4d13f05d77e87c16ffd1b4265f3b9c28322202ce786791b8db6",
   "in_dataset": true
  },
  {
   "path": "smoke2",
   "class": "impossible",
   "why": "the second smoke build, same shape and same seed defect. Kept because the amendments were argued against the differences between these builds, so a figure quoted from one cannot be checked against a rebuild of it.",
   "verified": true,
   "local": true,
   "bytes": 206624,
   "files": 23,
   "sha256": "fa98b7e6f1619c00f596faede0c4c06195de1050762d08b80df9f7e37fdf2ea1",
   "in_dataset": true
  },
  {
   "path": "smoke95",
   "class": "impossible",
   "why": "the smoke build that measured amendment A16's scope-use repair: the `FunctionDef` 0 -> 525 and 18-of-37 calibration figures are quoted from it, and it carries the generator's salted seeds like every other build.",
   "verified": true,
   "local": true,
   "bytes": 694101,
   "files": 51,
   "sha256": "df49f78d7298961a08324a817dd9a6ddca628eb93ecf3fb99303a2b8dc65d9f2",
   "in_dataset": true
  },
  {
   "path": "kernel",
   "class": "gpu",
   "why": "the kernel pass. An A100 pass over released checkpoints; `verify_kernel.py` checks it but cannot remake it without paying for it again.",
   "verified": false,
   "local": true,
   "bytes": 11550531771,
   "files": 9,
   "sha256": null,
   "in_dataset": true
  },
  {
   "path": "profiles",
   "class": "cpu",
   "why": "derived profiles for the eight pilot-era references. Rebuildable from the pilot record, and item 52 is about to write the sweep's own -- but check 9 reads these.",
   "verified": true,
   "local": true,
   "bytes": 11601588986,
   "files": 497,
   "sha256": null,
   "in_dataset": true
  },
  {
   "path": "record",
   "class": "gpu",
   "why": "the pilot record, 198 GB, 8 references. A paid A100 pass to remake, and item 33 decides whether it is kept at all. Listed because 'irreplaceable' and 'to be deleted' are compatible: the decision must be made knowing it cannot be undone.",
   "verified": true,
   "local": true,
   "bytes": 212931216963,
   "files": 3401,
   "sha256": null,
   "in_dataset": true
  }
 ],
 "hash_cap_bytes": 209715200,
 "note": "classes: impossible | gpu | cpu | deterministic; see experiments/irreplaceable.py"
}