coord/dial_postdeadline_gradient_amendment.json

{
  "schema_version": 1,
  "study_id": "dial_lexical_differentiation_v1",
  "amendment_id": "postdeadline_gradient_exclusivity_v1",
  "amended_at_utc": "2026-07-20T13:11:15Z",
  "parent_amendment_sha256": "41e1f07f55bafc8df170f9d0250ec4f17b6bf81e9d808586f3e05c51bcd97009",
  "decision_reads_outcomes": false,
  "failure": "A gradient audit exhausted GPU memory while sharing one MI100 with a mechanism assay; no gradient result was produced or read.",
  "operational_change": "Run unchanged non-gradient analyses with two workers per physical GPU, then run unchanged gradient audits exclusively with one worker per physical GPU.",
  "analysis_workers": {
    "vavvi": {"physical_gpus": 2, "non_gradient_workers_per_gpu": 2, "gradient_workers_per_gpu": 1},
    "vavvi2-1": {"physical_gpus": 2, "non_gradient_workers_per_gpu": 2, "gradient_workers_per_gpu": 1},
    "vavvi2-2": {"physical_gpus": 2, "non_gradient_workers_per_gpu": 2, "gradient_workers_per_gpu": 1}
  },
  "scientific_invariants": [
    "Gradient batches, seeds, model checkpoints, outputs and analysis code are unchanged.",
    "The scheduler change used only the HIP OOM and process memory evidence.",
    "Completed non-gradient artifacts remain value-blind and are not recomputed."
  ]
}