Check that she doesn't address you as a woman (#381) #24

2026-07-31T10:56:11+02:00

claude commented

2026-07-31 10:56:11 +02:00

Two commits: the new check, then a fix to it. Vikunja #381.

The phrasing eval was checking the wrong half of the persona. It verified that she speaks about herself in the feminine, and scored a clean 15/15. But the failure the persona rule exists to prevent went the other way — the model wrote:

ты давно не отдыхала

That addresses you as a woman, and the eval said fine. New hisgender column, its own check.

How it works, and where it lies. Russian past tense is gendered by suffix, -л / -ла. So: in any sentence that also talks to you ("ты", "тебе", "твой"…), a feminine past-tense word fails. Verbs that are hers are skipped — "я заметила", "напомнила тебе" are correct and must stay passing.

It is a suffix rule, not a grammar parser, and the doc comment says so plainly. It will false-positive on a feminine noun subject ("зарядка была утром, ты её пропустил") — there's a guard for that, which helps and won't always be right. It will miss gender outside the past tense (short adjectives, "сама"). Every hit prints the word it tripped on so you can disagree with it.

The second commit is me fixing the check. The exclusion list of "-ла words that aren't verbs" had села, мыла and стекла in it. They are nouns on paper, but in a nudge they are almost always verbs — "ты села", "ты мыла" — so listing them made the check blind to exactly the thing it's for. Removed. A false alarm you can dismiss beats a real hit that never prints.

Score. llama-server was up, so this was re-run for real (Qwen3.5-0.8B, 15 cases): 12/15 composite against the writeup's 13/15 — sampling noise, not a regression from this change. hisgender scored 15/15. Which means: this run did not reproduce the bad address, so the run does not prove the check. What pins it is a unit test on the exact recorded bad string. The three failures were unrelated — an emoji twice, one masculine self-reference ("я пропал"), two off-topic.

Left alone, needs its own task: she also addresses you in the third person and the formal plural ("Приходите… Жду вас"). That is a different persona bug from gender and needs a different check.

Two commits: the new check, then a fix to it. Vikunja #381. The phrasing eval was checking the wrong half of the persona. It verified that **she** speaks about herself in the feminine, and scored a clean 15/15. But the failure the persona rule exists to prevent went the other way — the model wrote: > ты давно не отдыхал**а** That addresses you as a woman, and the eval said fine. New `hisgender` column, its own check. **How it works, and where it lies.** Russian past tense is gendered by suffix, -л / -ла. So: in any sentence that also talks *to* you ("ты", "тебе", "твой"…), a feminine past-tense word fails. Verbs that are hers are skipped — "я заметила", "напомнила тебе" are correct and must stay passing. It is a suffix rule, not a grammar parser, and the doc comment says so plainly. It will false-positive on a feminine noun subject ("зарядка была утром, ты её пропустил") — there's a guard for that, which helps and won't always be right. It will miss gender outside the past tense (short adjectives, "сама"). Every hit prints the word it tripped on so you can disagree with it. **The second commit is me fixing the check.** The exclusion list of "-ла words that aren't verbs" had `села`, `мыла` and `стекла` in it. They are nouns on paper, but in a nudge they are almost always verbs — "ты села", "ты мыла" — so listing them made the check blind to exactly the thing it's for. Removed. A false alarm you can dismiss beats a real hit that never prints. **Score.** llama-server was up, so this was re-run for real (Qwen3.5-0.8B, 15 cases): **12/15** composite against the writeup's 13/15 — sampling noise, not a regression from this change. `hisgender` scored **15/15**. Which means: **this run did not reproduce the bad address, so the run does not prove the check.** What pins it is a unit test on the exact recorded bad string. The three failures were unrelated — an emoji twice, one masculine self-reference ("я пропал"), two off-topic. **Left alone, needs its own task:** she also addresses you in the third person and the formal plural ("Приходите… Жду вас"). That is a different persona bug from gender and needs a different check.

claude added 3 commits 2026-07-31 10:56:11 +02:00

Check that nudges do not address the owner in the feminine 10cf6f525c

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CGeSZxh1DCtRxmFVSYVGvJ

Merge the second-person gender check 15e5dd8eaa

Do not exempt Russian words that are both noun and verb 1bd2acdc2a

kami commented

2026-07-31 20:22:10 +02:00

Superseded by #47, which landed this whole stack on master as one reviewed integration merge. This PR head is an ancestor of master — its commits are in, nothing here is lost. Closing as merged-by-proxy rather than merged, since the merge came in through #47.

Review threads on this PR were answered or acted on before the merge; the Russian wording fixes went in as #48.

Superseded by #47, which landed this whole stack on master as one reviewed integration merge. This PR head is an ancestor of master — its commits are in, nothing here is lost. Closing as merged-by-proxy rather than merged, since the merge came in through #47. Review threads on this PR were answered or acted on before the merge; the Russian wording fixes went in as #48.

kami closed this pull request

2026-07-31 20:22:12 +02:00

kami referenced this issue from a commit

2026-08-06 17:46:49 +02:00

a corrected turn is promoted to a label that outlives the trace (V-630)

Pull request closed

Please reopen this pull request to perform a merge.

Sign in to join this conversation.

2 Participants

Notifications

Due Date

No due date set.

Dependencies

No dependencies set.

Reference: kami/Maven#24