Hayrullah Kar

Yasin abi, taslakları göndermeden önce bir şey sormak istiyorum. Sizin ekip tasarımı hangi programda yapıyor, Figma mı? Bize sonunda renk kodları, yazı tipi ve ikonlarla birlikte bir dosya olarak dönerseniz yazılıma çok daha hızlı aktarırız.

The SUM versus + comparison explains the 'total is too low' symptom better than most error guides. One catch in fix 2: REGEXEXTRACT with [0-9]+ keeps only the first run of digits, so '2.5 kg' becomes 2 and '1,200 pcs' becomes 1, with no error to warn you. A pattern like [0-9]+(?:[.,][0-9]+)? helps, though the separator still has to match the sheet's locale.

Thanks, the amount threshold is a sensible way to keep that queue from getting noisy. It puts the review effort on the invoices where a missing due date actually costs something.

Thanks, Charlie. Sending the empty-notes bookings straight to the unsure list with a plain check takes one more decision away from the model, the same move the post made with the rule. The disagreement log is the one I like most, because the emails where the three runs split are exactly the cases I would never think to write as tests myself.

Yes, failed has the same gap, since the run has to live long enough to write it. That is why the sweep is the piece I'd keep even with leases in place: it is where a stuck event finally becomes visible to a person.

Good to hear the gate already keeps a hash per chunk, since that is the part that actually points at the bad piece. Thanks for taking the correction so openly.

Thanks for explaining the PDF to base64 step in your Firebase backend, that was the piece I couldn't see from the post. With that conversion in front of it, the image_url path makes complete sense.

Agreed on doing that one first, Rishabh. A null tells the pipeline someone needs to look, while an invented date just flows through looking valid.

A small correction on the opening story. A SHA-256 check can't fail "by one character": flip one hex digit in the file and about half the bits of the digest change, so the two hashes look unrelated. The hash tells you the file differs, a byte diff tells you where. Since the file moved in pieces, listing a hash per piece in the manifest would point straight at the bad chunk.

One state in the claim table can strand an event: claimed. If the run dies after the claim row is written but before the AI step finishes, say a timeout on the model call or a scenario error, every retry of that event hits Path A and stops, because the row still says claimed. The duplicate gate then turns a crash into a lost dispute, and nothing fails visibly. A lease handles it. Treat a claimed row as live only while claimed_at is younger than the longest a run can reasonably take, and let a retry take over a stale claim by updating it. The failed status has the same weakness, since writing it needs the run to survive long enough to write it. A small scheduled step that lists claims older than the lease gives a human a short list to check.