Etymica
Report an issue
Human-in-the-loop feedback for the curated corpus
The evaluation pipeline currently executes 30 deterministic assertions (DET) and 8 model-graded metrics (MOD) across every generated piece. Most issues are caught automatically.
But automated evaluation can’t catch everything a domain expert notices — contested etymologies, missing senses, factual errors against specific sources, surfaced quotes that look right but aren’t. That’s where this feedback channel comes in.
What to include
- The piece ID — visible in the URL of the dedicated views (e.g.
/scroll/abolish,/story/abolish) or in the subject line if you used the Report issue link on the piece view itself. - What looks wrong— factual error, missing form, wrong relation direction, contested claim presented as definitive, etc.
- A source if you have one— academic citation, Etymonline link, dictionary reference, etc. Even a pointer to a competing analysis is useful; we don’t need a finished argument.
What happens next
Reports flow into a curated review queue. Pieces flagged here may be regenerated through the pipeline with the corrected information, or the issue may be added to the documented list of known limitations alongside the piece.
This is the human-in-the-loop the automated evaluation can’t replicate. It’s also one of the active eval-methodology questions for the dissertation: what does human-grader feedback look like at scale, and what does the pipeline learn from it?