{
 "_doc": "The lab's current claim and limits for one result, copied word for word from its own list of results (top40.json) at the commit named under source. They are the portfolio's current text for this id, which the portfolio released for publication on 2026-10-05.",
 "id": "inference-limits:exact-bounded-refused",
 "title": "EXACT/BOUNDED/REFUSED accounting that ships its own inflation control",
 "claim": "A declared-layout transformation campaign reporting 21 EXACT_BYTE, 35 EXACT_TOKEN, 2 NOT_EXACT and 6 REFUSED over 64 rows, each checked over the entire index space of its pair rather than sampled.",
 "limits": "A planted 'fictional-engine-9000' control earns a full EXACT_BYTE row with 1,280 positions checked — the certificate's own known-limits say a shape matching a real engine's is 'a citation, not a measurement.' NEG-F5-27: the headline was once INFLATABLE BY A ONE-CHARACTER CHANGE — redeclaring sglang with block=1 produced 16 invented EXACT rows, +29% on the headline, with every check, foil and limit still green. Counts must never be converted into vendor adoption. A3 2026-09-09: \"cross-engine\" was contradicted by the certificate's own boundary block. exact_transformation_cert.json boundary.does_not_hold[0] reads \"that any named engine would accept these bytes: nothing here links vLLM, TensorRT-LLM, SGLang, llama.cpp, FlashInfer, LMDeploy or transformers\", and boundary.holds[0] reads \"the eight layouts are DECLARED SHAPES with citations; no engine, kernel or GPU runs anywhere in this item\". I1 REWRITE 2026-09-02 (S04): reproduce_command is now the producer that scripts/check_cert_regen_drift.py ran (reports/peer_review_2026-09/regen_drift.json); the former python3 -c reader is kept as reader_command and is a transcription, not a reproduction.",
 "source": {
  "lab": "interconnect lab",
  "commit": "21c0c4194d8ab615b585e64efa57c86052eacf81",
  "file": "top40.json",
  "sha256": "c4c9b09c6125ac9961a1e2d90791977971d29344874f122fa57a91779c0e614e",
  "bytes": 300543,
  "entry_id": "exact-bounded-refused",
  "claim_sha256": "e0eff0178b3f67d64b4a05b547a5a13137b3a4b061714768ed68da09b8335f81",
  "limits_sha256": "b52bfb73a19743254b624150764aeec5ce66ffe6b2d4b985ed9809db0c146a88"
 }
}