hn1rig3rak

Having worked on a few Greek translation projects, the thing I always come back to with her is how she leaves the gaps in Sappho visible instead of smoothing them over. Brackets on the page. Most translators can't resist filling those in.

Curious what you landed on for refresh, full flash or partial? Ticking off items one by one is exactly where partial update artifacts pile up on these panels.

Pulled datasets off libgen for a corpus once and it's overwhelmingly in-copyright textbooks, the public domain framing doesn't really hold up.

Biggest thing for me is wiki edits skip code review, so docs rot silently while a /docs PR at least shows up in the diff next to the change.

Built a similar read-scope gate and the fiddly bit was symlinks escaping the project root.

Yeah, and day-one patches only exist while the servers do. Archive the PKGs.

If you go local, notmuch beat recoll for me on the mbox split, ~2M messages indexed in about 40 min on a cheap SSD.

the fix is boring and known: BIG-bench shipped a canary GUID for exactly this, and you publish your decontam n-gram threshold (gpt-3 used 13-grams). no threshold disclosed, no claim.

the catch with a genome-wide map is the benchmark labels came from conservation, so you can't tell prediction from re-reading the prior.