The camera wall finally points at the smoke.

Minutes to a human look, at a fixed budget.

Published smoke detectors report per-frame F1, or time-to-detection with no cap on false alarms. Neither tells a dispatcher which tile to open, or tells an agency whether the system survives a real shift before someone mutes it. PlumeRank reports one thing: minutes after ignition until a human is pointed at the right camera, with every system held to the same alert budget.

Held-out = 54 FIgLib fires no constant was fit on, read once after the constants were frozen. 27 of 54 is exactly half, so the median is fragile: it reads 32–38 min with the ignition clock shifted −5…0 min, and does not exist at +1…+5 (results/sensitivity_2026-10-10.csv). The development set (29 fires, in-sample) gave 17/29 and 32.0 min.

62 cameras. One list. Nothing hidden.

On the left, the live wall on the AWS box: 29 development fires and 33 fire-free confusion windows, replayed one minute of footage per second, re-ranked on every tick. On the right, the judge command run on a clean checkout. It recomputes the micro metrics from committed pixels, then prints the published held-out headline verbatim, labelled as not recomputed.

44.242.91.201 / the wallopen live ↗

Captured from the live box on 10 Oct 2026. The watch list keeps all 62 cameras and only reorders them.

plumerank — zshreplay of a real run

$ .venv/bin/python scripts/bench.py --arm micro
windows unified to 31 frames
CPUCreditBalance before : n/a
CPUCreditBalance after  : n/a

========================================================================
REPRODUCED HERE
========================================================================
  what           : the micro corpus, recomputed live by the real OpenCV 5 pipeline
  corpus         : data/micro/ (10 windows x 31 frames)
  PlumeRank      : p50    n/a min   p95    n/a min   (2/5 detected)
  B1 CNN grid    : p50    n/a min   p95    n/a min   (1/5 detected)      (the SAME onnx, every frame)
  B2 frame-diff  : p50    n/a min   p95    n/a min   (0/5 detected)
  realised rate  : 1.57 alerts/camera-night on the 5 non-fire micro windows (target 2.00, uncapped)
  escalation     : 0.65% of frames touched by ONNX (top-2 of a 10-window roster; the published figure in block 2 is the same rule over a larger one)
  checked against: results/verify_micro.json -- MATCH (byte-for-byte)

========================================================================
PUBLISHED, NOT RECOMPUTED HERE
========================================================================
  source         : results/bench_2026-10-10.json, printed verbatim -- not recomputed here

  PlumeRank -- median time-to-detection @ 2.00 alerts/camera-night
    arm            : hold (54 sequences, 6 excluded from the 60 in manifest_hold.json, see data/excluded.json)
    PlumeRank      : p50   38.0 min   p95    n/a min
    B1 CNN grid    : p50    n/a min   p95    n/a min      (the SAME onnx, every frame)
    B2 frame-diff  : p50    n/a min   p95    n/a min
    realised rate  : 1.99 alerts/camera-night on the confusion set (target 2.00, uncapped)
    escalation     : 1.76% of frames touched by ONNX
    model calls    : PlumeRank 1 / camera-night   |   B1 3645 / camera-night
    t0 sensitivity : advantage not finite at all 11 offsets in [-5,+5]
    background     : mog2
    commit 5df73321059f2797c32d302676b813c81b87fd0c  instance local (not EC2)  wall clock 12m29s  2026-10-10
  -> results/bench_2026-10-10.json

  reproduce with: make fetch && make bench  (about 95 min)

$ curl -s http://44.242.91.201/build-info | python3 -c "import json,sys;print(json.load(sys.stdin)['opencv']['version'])"
5.0.0

Fifty-four fires. One line each.

A fire lookout writes one line per event: time, bearing, what changed. Here is the held-out set written the same way: fires no constant was fit on, read once. Each row is a real FIgLib ignition, listed with the minute after ignition at which each system first detected it: the camera in its top four, with an alert. All three systems are calibrated to the same budget on the same fire-free windows. Misses are listed too.

The win is not a bigger network. It is the same one, asked less often.

B1 is PlumeRank's own ONNX tile classifier, run on every frame instead of the ~2% the classical stage escalates (1.76% on the held-out run). Because the model is identical on both sides, the gap comes from the pipeline, not from model size. Every system is calibrated to 2.0 alerts per camera-window on the confusion set; all of them land at 1.99.

Held-out set, 54 fires, read once. "No median" means the system detected fewer than half the fires; that is the result itself, not missing data. Source: results/bench_2026-10-10.json. On the development set PlumeRank also led, with 17 of 29; there B1 beat B2 (14 vs 10 through the same baseline, 6 vs 3 as deployed), the reverse of the held-out order.

A camera that runs hot all night is not a fire.

23 OpenCV 5 surfaces across 6 modules, counted from the source by scripts/check_claims.py on every test run. Four classical statistics fold into one hazard per camera. Each camera is then judged only against itself.

One Graviton instance. Every number committed.

What it cannot do yet, in its own words.

Asked, and answered.