Five slides read on AWS, one per density band. The agent spends effort in proportion to the evidence, which is the whole claim:
| Slide | Fields read | Parasites/µL | What it shows |
|---|---|---|---|
| TF41_UT5 | 3 of 3 | 90,435 | stops on the first field. The parasites are everywhere. |
| TF124_S540 | 2 of 10 | 19,512 | stops at two. The expert read 12,800, the same band, so it reads high. |
| TF101_233 | 8 of 8 | 8,602 | the ordinary case. Reads until the interval sits inside one band. |
| TF22_C104P65 | 11 of 11 | ~700 | reads everything it has and still straddles a band edge. |
| TF150_R5 | 5 of 5 | 288 | runs out of slide before the interval settles. Goes to a person. |
All five are NLM held-out patients the fit never saw. TF124_S540 records
the OpenCV binary that read its fields, under "Read by", as the Batch job
reported it; the four older sessions ran before that was recorded.
Or leave the endpoint blank and “Replay a recorded session” reads
sample-session.json from this page’s own folder, so the demo
still works with nothing deployed and a judge can check it offline.
The call
What it read
| Fields counted | — |
| Parasites | — |
| White cells | — |
| Fields skipped | — |
| Agent calls | — |
| Stopped because | — |
| Read by | — |
The WHO method reads 100 fields before calling a slide negative, or counts against 200 white cells. Where this stopped earlier or later, the log says why.
The log: every field, and what the agent did about it
| # | Focus | Debris | Found | Running | Move | Why |
|---|---|---|---|---|---|---|
| Nothing loaded yet. | ||||||
This is the audit trail. A laboratory needs to see the working behind a result before it can use one. Rows with no reason came from the rule, which produces none; the agent gives a reason for every move, and that difference is why both were run.
What the agent looked at
The fields of the loaded session, in the order the
agent read them.
Green rings are parasites, blue are white cells. Both are
count_field's own detections at the shipped threshold,
drawn on the real capture. The running estimate and the move the agent
chose, with the reason it gave, are printed on each frame.
No slide moves under the lens here, and that is deliberate. NLM fields are separate captures taken minutes apart, so a sequence of them is what a technician sees down an eyepiece.
Verification: click a detection to reject it
Rejecting a detection recomputes the count and the density from the corrections. Nothing is sent anywhere from this page; the corrections are what a deployment would keep as training data.