The two machines were taught to take turns, then measured properly for the first time: 200 jobs, no failures, 4,000 an hour each. Along the way they turned up a stowaway nobody had logged, and a spelling accident that has been mislabeling your live job board.
Solid = working cleanly · dashed red = starved or crashing · hatched = measurements that turned out to be lies
Seven moments, matching the numbered flags on the chart.
First real test: 20 jobs on each machine. Talos said 4.7 seconds a job, Nami said 11.5, and Nami's graphics driver kept crashing outright. Two identical machines shouldn't disagree by that much. That gap was the clue.
A production job-classifying service, running a model roughly five times heavier than ours, started up on both machines and never stopped. It was deployed months ago, appears in no schedule, and answered to nothing. Each machine's graphics chip can really only serve one program at a time, so ours stood in line behind it: Talos timed out politely, Nami's driver fell over.
Your instinct, one model per machine and one job at a time, is now enforced in the machines' settings rather than just in policy.
With the machines to ourselves: 100 jobs each, zero failures, 89.9 seconds on the clock, or 4,004 jobs an hour. Nami's "hardware problem" was never hardware, it was the roommate: the twins finished within a tenth of a second of each other and produced identical results, column for column.
Only 0.33s of each job is real work; the rest is fetching and filing. That plumbing is the next easy win.
The stowaway turned out to have been built with a yield switch already inside it, waiting years for something to connect to. We connected it. It now takes the machine, does a batch, hands it back, and waits its turn like everything else. Proof from the ledger: our test asked for a machine while the worker held it, waited, took over 5.3 seconds after the worker let go, finished, and handed it straight back.
An NVIDIA networking job came back categorized as nursing. The cause: the category matcher looks for keywords anywhere inside a title, and "RN", the abbreviation for Registered Nurse, hides inside ordinary words. Intern. Journeyman. Harness. Government.
It isn't only "RN". "PM" hides inside equipment. "AE" hides inside aerospace.
You said you felt the live site was mislabeling. It is. Counted against production, read-only: 503 jobs wrongly in nursing, 1,027 equipment jobs wrongly in Product & Design, 31 aerospace jobs wrongly in sales. That's from three keywords out of nineteen suspects, so the true figure is higher.
Fixed in the test system tonight and verified: the fakes stopped, while genuine nursing titles still land in nursing. The production repair is staged and waiting on you.
Fixing the fakes exposed something larger. 79,759 jobs, 24.5% of your board, carry no category, and some of what looked like coverage was noise wearing a label. That's the pool your marine biologist lives in.
The design that answers it, from your own question: let the system say "I don't know" instead of guessing. A blank is honest; a wrong label is damage. Unknown jobs pool up, similar ones cluster, and when enough of them agree they graduate into a proposed new category for you to approve. Research found the free government job dictionaries (they already contain "marine biologist") and a national statistics agency running this exact three-step design in production, so most of this is adoption rather than invention.
All of it in a sandbox. Nothing on your live site was changed.
A shared calendar plus a pass-along lease, so the machines take turns instead of colliding. DSF owns 4:00–7:00 every morning on both; outside that they belong to job parsing. Hands back early when it finishes early, keeps going past 7 if a render is mid-flight, exactly as you described it. A new project gets a block by adding one row. If anything crashes holding the machine, it's released automatically after ten minutes.
Per-company extraction rather than per-platform. That was your call, and clustering proved you right: Walmart and NVIDIA are both "Workday" and look nothing alike. Alongside it, a trust system watches each company's results and quietly stops double-checking once a company has earned it. That's where the speed came from.
Built and loaded: 100 jobs per machine where the AI reads each job cold and extracts everything itself, with no pattern-matching hints, scored field by field against the current method. It runs this morning after the DSF block.