Data-quality: in-app /bug → LLM-titled pipeline yields ~1.5% content-empty issues naive metrics can't detect (measured, reproducible)

Sharing a measured, reproducible data-quality finding about the in-app /bug → issue-title pipeline. I'm an independent researcher (no affiliation with Anthropic; I'm a Claude Code user).

Measuring this repo's public issues over one week (2026-05-30 to 2026-06-06):

• ~22.7% of new issues come through the in-app /bug flow, whose titles are LLM-generated.
• When the Bug Description is empty/off-topic, the titler emits a refusal and that becomes the issue title (one title literally leaks the titler's own task instruction). ~1.5% of new issues (95% CI 0.8-2.8%) carry no human bug content.
• They pass naive emptiness checks (the template auto-fills environment/error fields), and github-actions closes them NOT_PLANNED in scheduled sweeps. So Pulse / triage / "issues closed" metrics count a few % of pure scaffolding. The same measurement on microsoft/vscode, golang/go, and kubernetes/kubernetes over the identical window = 0% — it is specific to this pipeline.

Two fixes that would remove most of it:

1. Don't promote a titler refusal to an issue title — detect the refusal pattern and fall back to a neutral title (or block submission when the Bug Description is empty).
2. Flag/label pipeline-origin issues with an empty Bug Description so triage and metrics can exclude them.

Full mechanism, de-identified data, and reproduction scripts (no third-party handles or issue numbers — aimed at the pipeline, not at users): https://doi.org/10.17605/OSF.IO/JCDG5

Disclosure: AI-assisted analysis (Claude).