{
  "schema": "nodus.synthetic-paired-byte-provenance.v1",
  "title": "Same target, different prompts: a synthetic byte-position schematic",
  "authorship": "Original demonstration code, toy records and writing by Nodus.",
  "fixture_origin": "Two newly authored text-repair examples: an arithmetic operator and an accented label. They are not CUDA kernels or actual training trajectories, and their diagnostic notes are authored rather than compiler observations.",
  "implementation_origin": "New standard-library-only code for this explanatory artifact. Existing private paired-workflow metadata and relevant mask-helper behavior informed the invariant selection; their implementation, templates and real data are not copied or imported.",
  "representation": "UTF-8 bytes plus declared synthetic START/END/PAD symbols. No actual model tokenizer, chat template, learned vocabulary or token-boundary equivalence is represented.",
  "scope": "An illustrative control/treatment data-interface check before any proposed training experiment.",
  "excluded": [
    "Private codec/resolver and runtime implementations",
    "Real training trajectories or accepted repairs",
    "Model weights, tokenizer assets or teacher outputs",
    "Sealed evaluation cases or confirmation source",
    "Customer content, operational identities, credentials and accounting records"
  ],
  "source_group_limit": "Exact declared group/ancestor collisions can be checked; unknown ancestry or semantic, teacher and pretraining leakage are not ruled out.",
  "compute_limit": "Equal supervised byte-symbol sequences and padded shapes do not establish equal real tokens, FLOPs, attention work, wall time, gradients or training budget.",
  "result_limit": "No training, generation, numerical grading, provider measurement or model-improvement result is claimed.",
  "new_code_license": "MIT, limited to the original demonstration and accompanying original files; no private or third-party implementation is included or relicensed.",
  "local_execution": {
    "python": "3.9.6",
    "source_sha256": "1b84890f33f67302d1dab5bb51fdfe62e8b722a3eea0e41bdb78188dedecc32f",
    "fixture_sha256": "ff75a6bcc0583d565bb92f8c4dbd974d5c241962058e2d1ba93cc51d97937b82",
    "result_sha256": "de3a42ca27de16c9b8202d42099bed1769d36994b726aee9a7d7167e0d77757a",
    "status": "schematic_checks_passed",
    "negative_controls": 14,
    "expected_rejections_matched": 14,
    "training_or_actual_model_tokenization": false
  },
  "count_scope": "Counts are UTF-8 byte positions and declared toy symbols. Equal supervised byte sequences do not establish real-token or compute equality."
}