Commit ea7d48e7 by PLN (Algolia)

docs: README section for buvard fix — the local-only loop, the guards, the two-liner

parent 19f0fc05
...@@ -98,6 +98,32 @@ catalog as an agent skill: load it, run the linter, apply the findings by hand. ...@@ -98,6 +98,32 @@ catalog as an agent skill: load it, run the linter, apply the findings by hand.
Hits are review candidates, never verdicts — the agent edits prose like an Hits are review candidates, never verdicts — the agent edits prose like an
editor, not like a filter. editor, not like a filter.
## Fix — the loop with a local model
`buvard fix article.md` closes the loop: it runs the linter, hands the flagged
excerpts — with each rule's one-line rationale — to a **local** language model,
then checks the suggestions before believing any of them.
- **Local only.** The endpoint is LM Studio's OpenAI-compatible server at
`http://127.0.0.1:1234/v1` by default. `BUVARD_FIX_URL` points at another
local endpoint, and that is all: the text never leaves the machine.
- **Dry run by default.** You get the diff, a per-suggestion table (applied /
KEEP / dropped, with the reason) and before/after scores. Nothing is written
until `--write`, and `--json` gives the whole structured result.
- **The guards outrank the model.** A suggestion that alters a number or a
name is dropped. A suggestion whose excerpt does not match the text exactly
once is dropped. A suggestion the linter cannot confirm — score up, or hits
down — is dropped. KEEP is always honored: the voice wins.
- **Three rounds maximum.** Each round feeds the remaining flags back; the
loop stops as soon as a round stops improving the score.
Setup, two lines:
```
LM Studio → Developer → Start Server # serves http://127.0.0.1:1234
buvard fix article.md # BUVARD_FIX_MODEL picks the model
```
## Science and references ## Science and references
The catalog is data with provenance, and the barème is a ratified decision — The catalog is data with provenance, and the barème is a ratified decision —
......
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment