# CLAUDE.md labeling: the AIMsight displacement labels and the knowledge to place them, in a single copy on the workstation (`README.md`). Diagnostic-App reads and writes there, Optimizer only reads, both through a link `labeling -> ~/labeling`; commits are made here. - **The labels**: `manual/.jsonl` (the expert, for him alone), `auto/.jsonl` (the agents, reviewed by `review`), one per line, `{"id", "start", "end", "family", "description", ...}`. The Diagnostic-App analyses: `auto/analyses/__/`. - **The knowledge**, one source per kind (`README.md`), maintained by hand: `families.md` (the families, their boundaries, the catalogue of phenomena), `labeling-rules.md` (bounds, a single event, description, stated cause), `reading-rules.md` (what the data mean), `stations.md` (the installations); the cases are the labels, through `precedents.py`; what the data are: `tools/DATA.md`. A piece of knowledge lives in one of them only. - **The tools**, standard library: `python3 scripts/validate.py [STATION | file.jsonl]` (the rules), `python3 scripts/ids.py [STATION]` (the stable id of new labels), `python3 scripts/precedents.py "" [--exclude STATION START END] [--family F] [--k N]` (the RAG; it starts the Ollama container `ollama-bge-m3` by itself if it does not answer; `--check` without it). - **Verify**: `python3 scripts/validate.py` (today 19 expected errors: the labels submitted by the app without `review`) and `python3 scripts/precedents.py --check`. Rules: - **`manual/` is the expert's, for him alone**: no agent and no process reads or writes there; the whole project (analyst, precedents, benchmark, optimization) uses only `auto/`. - A label added to `auto/` goes through `ids.py` then `validate.py`; an id never changes. - Every time is in UTC; descriptions in words, without numeric values; written in English from now on. - No rule specific to a station, a target or a date in the rules and the families. - Commit by path, never `git add -A`: Diagnostic-App writes here at the same time. - Language: English.