{
  "version": 1,
  "slug": "vllm-streaming-tool-call-parser-corruption-2026",
  "evidence_class": "private_local_fixture_check",
  "what_this_is": "A client-side, pure-stdlib check of the published gate against hand-written OpenAI-compatible streaming deltas. No vLLM package, model weights, GPU, tokenizer or vendor endpoint is involved. This record does not measure vLLM or any hosted API.",
  "run": {
    "command": "/home/ubuntu/blog-automation/.venv/bin/python toolcall_gate.py --selftest",
    "cwd": "research/evidence/vllm-streaming-tool-call-parser-corruption-2026",
    "started_at_utc": "2026-10-28T12:07:37.837575+00:00",
    "ended_at_utc": "2026-10-28T12:07:37.911143+00:00",
    "exit_code": 0,
    "stdout": "{\n  \"cases\": [\n    {\n      \"case\": \"clean_stream_split_at_every_byte_offset\",\n      \"offsets\": 760,\n      \"failures\": 0,\n      \"detected\": true\n    },\n    {\n      \"case\": \"merged_calls\",\n      \"detected\": true,\n      \"codes\": [\n        \"ARGUMENTS_NOT_JSON\",\n        \"CALL_COUNT_MISMATCH\"\n      ],\n      \"call_count\": 1\n    },\n    {\n      \"case\": \"truncated_arguments\",\n      \"detected\": true,\n      \"codes\": [\n        \"ARGUMENTS_NOT_JSON\"\n      ],\n      \"call_count\": 2\n    },\n    {\n      \"case\": \"retyped_arguments\",\n      \"detected\": true,\n      \"codes\": [\n        \"TYPE_MISMATCH\"\n      ],\n      \"call_count\": 2\n    },\n    {\n      \"case\": \"dropped_non_string_arguments\",\n      \"detected\": true,\n      \"codes\": [\n        \"MISSING_REQUIRED\"\n      ],\n      \"call_count\": 2\n    }\n  ],\n  \"failures\": [],\n  \"summary\": {\n    \"cases\": 5,\n    \"failed\": 0\n  }\n}\n",
    "stderr": "",
    "python": "3.12.13",
    "platform": "Linux-6.17.0-1019-oracle-aarch64-with-glibc2.39"
  },
  "inputs": {
    "toolcall_gate.py": "37f15a2846fc20ab05f1e4931af0ed489cf29fc72fcac9939bc31dc8d480247e",
    "fixtures/sse_parallel_toolcalls.json": "394222291fe501026cf5de245d0391f51e58b00ae1fedadfefb31efe400f5025",
    "fixtures/tool_schema.json": "e19f1adea05231ee8bb707a4eec642e3ef5ea1d3c76af601b9201bf512049716"
  },
  "method": [
    "A 5-event fixture stream carrying two parallel tool calls is serialised as OpenAI SSE frames and split into two byte chunks at every offset from 0 to the full length (760 splits); each split is reassembled and validated.",
    "Four constructed corruptions are derived from the same stream: two calls merged into one delta with concatenated argument text; arguments cut mid-string; integer, boolean and array values re-encoded as strings; and all non-string arguments dropped.",
    "The expected outcome is zero rejections for every clean split and at least one specific issue code for every corruption."
  ],
  "interpretation": "The gate is demonstrated to be boundary-safe and to catch the four constructed corruption classes in this fixture. vLLM parser behaviour itself was not exercised; the reports in the article remain third-party claims.",
  "not_claimed": [
    "no vLLM import or reproduction",
    "no model output was sampled",
    "no timing, throughput or accuracy number"
  ]
}