{
  "status": "verified_runtime_regression",
  "previous_revision": "a5dd79f4a25d68b942e0df2ab674db8c8866a802",
  "change": "Materialize each encoder cache rebase before constructing the next layer rebase. Same model weights, cache dtype, context length and inference profile.",
  "runtime_sources": {
    "__init__.py": "687ad18a08c3dd374e55113a43b33ac1b27e208b01ce99a827187af4244ccd47",
    "__main__.py": "cf7390a2508a0fcad4777a4a629fc5212ab655cb45845975995f835676a834e3",
    "cli.py": "c8d7029a9e77436ba50739499b2132ce02d45e16c148640f3d3824bf61883158",
    "frontend.py": "fcfa93161c7c431fe5918fb559c52faa587c9eb0e790fae4057caed278f7f383",
    "gptq_q4_provenance.py": "461163cd7c68fcf5c4ff0061be1d895636f91ff31cfb12ba6c89f218aec0c418",
    "math.py": "839517e331f76bb3b000998e8368761c18575e67c07cc6be49cdd5029e634851",
    "model.py": "b2544d52e5dc2ff00e5f92364e9500d01028d3da9f5d0f5ce2acaa50e75229e8",
    "native_model_format.py": "600b957bb9f6ed38b1078d2d2d870d201b120774415d4b5d8f4121784f4be258",
    "native_weights.py": "4e518da4e765a90a3e419ad0c9827651f2a5829c55052bcef9ec855cf2ea2ffd",
    "silero_input.py": "8d36f9cfc8aa92ce34afd1702113cec17f688a5bb0e0cd5ca73f387821cc620d",
    "silero_stream.py": "1406e2e8972c638dedfd0dcf9b8e72d06296aed503f1bef28d86a1b9d8aa595e",
    "stream.py": "60bc5d7b5c6946ad63df284feb8d04bc99163bf15eab2f75d23f5a4ca7a97a55",
    "weights.py": "4c34bb6e77e41cdaa268498d80a65dc595a8c46e273d2414e2f122e3f0a04c1d"
  },
  "source_archive": "sources/mlx-q3-sequential-cache-rebase-v1.tar.gz",
  "source_archive_sha256": "96946f72aaaf8ec09e7a45fcd80ddb8d2378f5e23d73df740d4f6b73f7e345ae",
  "regression_report": "evaluation/mlx-memory-regression-11/verification.json",
  "regression_report_sha256": "09dcfba1bace7545f360c8cb810bd23e429810a8d20a8bd28222c01a0dc7b02f",
  "inputs_previously_evaluated": true,
  "no_new_heldout_claim": true,
  "zero_copy_loader_not_selected": true,
  "encoder_compression_not_selected": true
}