The counter congratulated us for inspecting the same plate again

Yes, plate IDs are on every attempt. They're also in quality's register. The dashboard just adds successful rows, so the identity is there and doing absolutely no work.

Use the current disposition per ID, then. Does quality ever put a passed plate back on hold? If so, any pass in its history won't be enough to decide whether it belongs in today's accepted count.

What exactly gets added at restart? Previously saved attempts, or new attempts created by the restart itself? Those would leave different problems in your records.

I'd like that answer too. Fixing the displayed sum won't clean up attempts that should never have been created.

Saved attempts are loaded, then the counter increments for each pass again. No new inspections ran. And yes, quality can put a plate back on hold after a pass.

So reconstruct the count from the current item decisions at startup. Don't take yesterday's total and pour the old events into it again.

Our bracket dashboard had the same pass-history trap, including later holds; make the current reviewed disposition explicit instead of treating the latest successful test as acceptance.

Does the screen show those holds anywhere?

And can the operator open the attempt history from that plate? They shouldn't have to choose between an honest item count and finding out why something is held.

There's a history page, but it opens for the whole batch. Holds only show in quality's register. I've asked for per-plate history and the current decision on the same row.

Where will that current decision come from? If the dashboard has its own copy, you also need to know when it was last updated from quality's register.

The developer is using the reviewed disposition records directly. We've agreed to show unknown if they can't be read, rather than rebuilding acceptance from the test rows.

Unknown beats another imaginary finished plate

Try a pass followed by hold, then restart. That's a small example which should expose both of the mistakes you've found, without needing to rerun a whole physical batch.

Also include an item with two reviewed retries and a final acceptance, so retaining its complete history does not add multiple accepted items.

Both examples now count correctly in the offline test. History still has the original failures and retries. I was relieved to see the numbers get smaller for once.

Have you compared the full saved batch with quality's reconciled list yet? Small examples are useful, but the existing records may contain another case nobody remembered to model.

That comparison matches too, including plates on hold. Workshop rollout still pending. The students have requested credit for noticing that a tray is not the same thing as a dashboard.

Fair request from the students. Keep their original counter example in the checks; it caught a real mistake that all the happy-path demonstrations missed.