Naimono Lab

Watch out

The overall score is a maximum across the existence angles, splice and sentence, not an average and not a product of probabilities, so one confident angle decides the round. Thresholding is a hand-tuned policy, and where the numbers came from is not a benchmark — the README says explicitly that the scores are the model's judgement values rather than existence probabilities or web-verified evidence, that the classification and word-feel outputs are the model's imagination and are never used for the existence decision, and that there is no guarantee it will find an unknown band or technical term. The helper model's answer is treated as advisory only: if Jev's existence or name-suspicion is 0.65 or above, or splice or sentence is 0.60 or above, the answer is escalated to discussion regardless, and the helper's own verdict and Jev's numbers are stored in separate fields so the original score is never overwritten. Its memory is a model, not a search index, so proper noun misses and misidentifications remain. Running it locally with npx wrangler dev connects to your Cloudflare account and can incur Workers AI usage, while npm start is the offline bundled-dictionary version where the family decides. The Durable Object budget stops at 300 calls a day or a conservative 5,000 Neuron estimate, whichever comes first, and the README is explicit that estimate is not actual billing. Copy reviewed against the repository README on 2026-10-04, not from running it.