Check prices survive translation before burning live-selling captions

Before you burn translated live-selling captions, compare the digit groups of each source cue with its translation: if the multiset of numbers differs, hold the cue for review. It catches a dropped price or a changed quantity, and it costs nothing, while each re-burn on Sume is another $0.20 caption job.

Bilibili's Index-Translate README says the models support translation instructions such as terminology and formatting preservation, but a model is not a guarantee. Treat the check as a gate, read 2026-10-05.

What the check does and does not prove

It compares numbers, not meaning.

The check

It strips thousands separators and decimal marks only for comparison, so 1,299 and 1.299 are both 1299. Some languages may write digits differently, so a hold is a flag for a person, not proof of an error.

import re
from collections import Counter

def digits(text):
    return Counter(re.sub(r"(?<=\d)[.,](?=\d)", "", t) for t in re.findall(r"\d[\d.,]*", text))

def needs_review(source, translated):
    return digits(source) != digits(translated)

pairs = [("Only 12 left at $24", "Solo quedan 12 a $24"),
         ("Only 12 left at $24", "Solo quedan a $24"),
         ("Total $1,299", "Total 1.299 $")]
for src, out in pairs:
    print("HOLD" if needs_review(src, out) else "ok", "|", out)

Where this sits in the pipeline

Run it after translation and before the caption job. A held cue goes to a person, and only passed cues become authored cues for a Sume caption job. The Video captions docs describe the cue fields.