2026.09.15 / MODEL SERVER

VisionEye · Model Server

Images were sent to a local server with YOLO26n loaded in memory. Actual detection responses were recorded, then the inference worker was stalled or terminated to test automatic recovery.

Server code ↗Run summary ↗Raw responses ↗Event log ↗

Normal inference: 20 / 20
HTTP response p95: 116.4 ms
Response after worker exit: 3.29 s
Inference device: CPU

01 / SYSTEM

Server architecture

01HTTP · Bearer token
021 inference slot · no queue
03YOLO26n · CPU worker
04Detection boxes · response log

The HTTP server and model process are separate. Readiness follows model loading and warmup. A failure disposes of the old process and pipe, then starts a new generation.

02 / INFERENCE

Actual responses

Generated footage · actual returned boxes · not privacy masking

Selected request

{
  "type": "result",
  "request_id": 2,
  "detections": [
    {
      "xyxy": [
        220.27630615234375,
        391.68316650390625,
        325.22369384765625,
        602.9668579101562
      ],
      "confidence": 0.8701023459434509,
      "class_id": 0
    },
    {
      "xyxy": [
        869.9039916992188,
        361.748779296875,
        962.1455688476562,
        580.7520751953125
      ],
      "confidence": 0.8581218719482422,
      "class_id": 0
    },
    {
      "xyxy": [
        468.11492919921875,
        232.8916015625,
        535.650146484375,
        410.51953125
      ],
      "confidence": 0.8497844338417053,
      "class_id": 0
    },
    {
      "xyxy": [
        555.414794921875,
        208.18496704101562,
        630.7957763671875,
        393.5848388671875
      ],
      "confidence": 0.8456692099571228,
      "class_id": 0
    },
    {
      "xyxy": [
        656.106201171875,
        273.41302490234375,
        721.9979248046875,
        458.52911376953125
      ],
      "confidence": 0.8320407867431641,
      "class_id": 0
    },
    {
      "xyxy": [
        777.20263671875,
        370.71307373046875,
        871.499267578125,
        600.5599975585938
      ],
      "confidence": 0.8298223614692688,
      "class_id": 0
    },
    {
      "xyxy": [
        114.68441772460938,
        374.2510986328125,
        220.70620727539062,
        586.3094482421875
      ],
      "confidence": 0.7722750306129456,
      "class_id": 0
    },
    {
      "xyxy": [
        744.9744262695312,
        0,
        774.9558715820312,
        55.1549072265625
      ],
      "confidence": 0.575564444065094,
      "class_id": 0
    }
  ],
  "width": 1280,
  "height": 720,
  "inference_ms": 89.661,
  "torch_threads": 2,
  "image_sha256": "e22ecbd82c9ac69d8f82a9819fe0ca256409160bca0237df6659204284c30ad1",
  "generation": 1,
  "model_sha256": "9b09cc8bf347f0fc8a5f7657480587f25db09b34bf33b0652110fb03a8ad4fef",
  "device": "cpu"
}

HTTP latency of 20 sequential requests

01 → 20p50 93.6 ms / p95 116.4 ms

A small functional smoke sample, not production throughput or accuracy.

03 / RECOVERY

Failures and recovery

Actual response after recovery

Check	HTTP	Behavior
Missing authorization	401	Rejected before inference
Invalid image / oversized body	400 / 413	Input validation
Admission limit	429	Additional request during an injected stall
Deadline	504	Terminate old worker → fresh pipe and model load
Worker exit	503 → 200	Actual inference succeeds after automatic restart

04 / EVIDENCE

Environment and scope

Pinned configuration

Python 3.12.14
PyTorch 2.11.0+cu128
Ultralytics 8.4.150
CPU · imgsz 640 · rect false · conf 0.1
Model SHA256
9b09cc8bf347f0fc8a5f7657480587f25db09b34bf33b0652110fb03a8ad4fef

Metrics ↗Input provenance ↗Reproduction ↗

Validation uses three frames from an AI-generated video. It excludes real CCTV, tracking-session APIs, GPU or LLM serving, long-running operation, HTTP-server/host recovery and model-version rollback. PyTorch is an existing CUDA build, but the actual model device and execution threads were verified as CPU / 2.

The rehearsal server was stopped after validation. This webpage displays recorded evidence and does not run a model.