Northbeam Precision / Operations Watchtower

ON WATCH 00:00:00 ET NEXT SWEEP 0:00 SWEPT 0 OPEN 0 watching
Conventional dashboard what everyone else is looking at
Verified today
0
orders through the station
Pass rate
100%
verdict PASS over total
Errors
0
pipeline executions failed
Stations up
2 of 2
last report just now
All green. Nothing here is wrong, and that is the problem.
Findings flagged, then passed anyway
Quiet.
This is the normal state, and it is the product.
Let the sweep run, or break something below.
Break somethingthe failures that leave a dashboard green

Neither of these turns a dashboard red. A dark station reports nothing at all, which reads as no failures, and a dead component keeps returning a green status while producing no output.

Calibrationthe other half of a real catch rate
Candidate rule under calibration: label lacks the product family prefix. Sound in theory. Over a 197 execution window it fired on almost every order, because internal family prefixes never appear on a physical label at all.
0 false fires
Retired. A monitor that cries wolf gets muted, and a muted monitor is worse than none. From the same window: zero cases of a hard failure that was nonetheless passed, and confidence landing exactly at the floor on 23 percent of traffic, which is house style rather than a defect and is only meaningful when the product has a look-alike sibling. That is why the sibling rule survived and this one did not.
How this works

Why it exists. The verification system was working. It was reading labels correctly and naming discrepancies correctly. At least four wrong shipments went out anyway. In three of them the discrepancy was raised in a way that could be clicked past, and nothing downstream ever asked whether a flag had been resolved or merely dismissed. On every dashboard those orders were green. They passed.

Shape. A read-only collector runs against the workflow engine's own execution database through a watermark, every five minutes, during working hours. It writes nothing to anything it observes and keeps its own findings store. Findings are triaged by a reasoning model, and the assessment sits beside the raw evidence. The model advises. A person rules.

Every rule records the incident that created it. That is a deliberate documentation practice: nobody may delete a rule without reading why it was written. Open the provenance line on any finding above.

Investigation. Click a finding, or type any order number, and get the attempts timeline in local time, who was signed in for each attempt, per line confidence against its floor, every flag raised, and the actual photographs the vision model judged, recovered from the execution record. Before this existed that reconstruction was a series of manual database queries and a hunt through cloud storage folders.

Two traps worth knowing. An order lookup that returns nothing usually means the execution history has aged out of retention, not that the record never existed, so queries are bounded by date. And order numbers appear quoted inside hundreds of unrelated execution payloads, so a plain substring search returns noise; photograph recovery is anchored on the finding's own execution identifiers.

A fault this watcher did not catch. An AI component feeding its training queue returned nothing for eighteen hours while re-billing the same three calls every thirty minutes. It looked healthy from outside. The morning check and an audit found it. The cause was a token budget that the model's own default reasoning consumed entirely, leaving no room for output. Repairing it surfaced two wrong-shipment risks that had been invisible, and the failure now has its own rule.

It runs on the server, not on a desk. This started as a compiled desktop program and that version was retired, because the floor has to be watched even when a particular workstation is off. On the day that decision was made, the machinist's workstation was asleep for the entire build session and nothing was blocked by it. Each desk now opens the same URL as a desktop-app-mode window, and rebuilding the container updates every desk at once.

Plain text for operators. Model prose is normalized before it reaches an operator. Characters are never deleted, because deleting welds neighbors together and silently corrupts arithmetic inside a note.

Measured. Sweep every five minutes, working hours, weekdays, for known failure classes. It makes no promise about every defect. Calibrated over a 197 execution window. Zero blunt hard-failure-but-passed cases found. One candidate rule retired after 215 false fires. A production machine caught after sitting powered off for an entire weekend, where the tell was subtle: the network controller still listed the client with a stale uptime for over half an hour and the switch port stayed link-up the whole time.