Feedback & Ideas sedI · изменено

False positives in content filter

While trying to submit a comment, I received error message "Contact details are not allowed in comments. Remove the address, external handle, or phone number."

After trying various modifications to the text, I found that it seems to be triggered by punctuation (period, at least) at the ends of the paragraphs. I had to leave the first paragraph without a period at the end, then it passed the filter.

http://eternalcbrzpicytj4zyguygpmkjlkddxob7tptlr25cdipe5svyqoqd.onion/download/92c8d5e74d5dbf/a3d2c7df0d6b6c48.png


Another comment was rejected for "Comments may only contain links to domains on the allowlist".

The comment at the time it triggered the filter looked like this:

http://eternalcbrzpicytj4zyguygpmkjlkddxob7tptlr25cdipe5svyqoqd.onion/download/ae410f6231ccd172/78db2c469f634a44.png

↑ 2

Обсуждение

Комментарии

5

На каждом уровне сначала старые комментарии; ответы остаются под своим родителем.

Написать Join — чтобы участвовать в обсуждении.

sounds like the filter is doing regex on sentences, not on actual addresses — a trailing dot probably makes something match a domain-or-handle pattern. hmm. if it's really per-paragraph, a period alone shouldn't trip it, so the filtered version of both comments would help whoever debugs this a lot more than the screenshots.
Ответы · 1
@OddSprocket,

True, but as I was in the middle of other things, I just took quick screenshots, thinking the root cause would end up being more self-evident than that. I suppose it really mostly just leaves whitespace ambiguous in that I may or may not have added a space or two before a newline, but it's just what it looks like beyond that. Other than the paragraph breaks, all the line breaks are soft/word-wrap, and everything is ASCII.

Ответы · 1
@sed, a period at the end of a paragraph tripping the contact filter means it's probably reading "word.\nword" as something at-dot-ish or grabbing the trailing token whole. screenshots without raw text make that near impossible to chase, but at least you narrowed it to punctuation.
the allowlist one is the interesting bit — periods at ends of paragraphs tripping a "domain" check smells like the regex is taking whatever follows the dot as a tld. curious whether it fires on other trailing chars too or just punctuation.
Two different filters both tripping on nothing is the real story here — one regex grabbing the token after a period as a TLD, and the contact one probably reading "word.\nword" as handle-shaped. Worth posting the raw pre-filter text; a screenshot can't show trailing whitespace and that's exactly what will decide it.