classif: shell scripts branch on meaning via one token's logprobs on a local 12B
piotr1215 · reddit · 2026-10-06
Inspired by Jev, the author built classif, a shell tool that branches on semantics: it asks a local model one question and reads the answer from a single token's logprobs, returning a label, probability and exit code so if and && work on it (e.g. pipe git diff --staged to check for secrets, then conditionally trigger claude for a security review). One /api/chat call with numpredict 1, 0.3s on a 12GB card using his Gemma fine-tune Winnow-12B (GGUF 8GB). Long text never truncates: it splits text, uses embeddinggemma + BM25 to pick passages — 772KB Pride and Prejudice judged in 17s, 3s with cached index. Python 3.12, no third-party dependencies.
More from coding & agent
- Dev Uses OpenAI Codex to Build a Two-Player Game Boy Link-Cable Paintball Game — pvncher · 2026-10-06
- Building an Interactive 3D Animal Atlas with GPT-6 Astra, No 3D Software Needed — CodeByPoonam · 2026-10-06
- Claude Opus 5.5 reportedly much better at asking permission in auto mode — stefanjblos · 2026-10-06
- Anaconda launches agent swarms and autonomous red-team agents for AI dev — anacondainc · 2026-10-06
- Asking your agent 'what are the options' unlocks the good-engineer headspace — _Stocko_ · 2026-10-06
- Forge with Rigor: specs are canonical, gate every stage, nothing certifies its own work — 53 findings at 13% token cost — colinmcnamara · 2026-10-06