CLI

python -m lemmas <subcommand> "<query>" [flags]

Works offline against the deterministic echo stub OR with OPENAI_API_KEY against gpt-4o-mini. All subcommands accept --json for piping.

Subcommands

python -m lemmas cove "Who invented the laser?" --n-questions 4
python -m lemmas self_consistency "What is 13 * 17?" --n 5 --extractor last_number
python -m lemmas best_of_n "Write a haiku" --n 5 --keywords "snow,arctic"
python -m lemmas drift "hello world"          # requires OPENAI_API_KEY

Common flags

Flag	Default	Notes
--model	gpt-4o-mini	Only used with OPENAI_API_KEY.
--temperature	0.7
--max-tokens	512
--json	off	Emit JSON instead of human-readable output.

Per-subcommand flags

Subcommand	Flag	Notes
cove	--n-questions N	Verification questions to plan (default 4).
self_consistency	--n N	Sample count.
self_consistency	--extractor X	last_line, last_number, regex, similarity.
self_consistency	--regex PATTERN	Required when --extractor regex.
best_of_n	--n N	Sample count.
best_of_n	--target-chars N	For default length_scorer.
best_of_n	--keywords k1,k2	Use keyword_scorer instead.
drift	--bucket NAME	Bucket id (default default).
drift	--z-threshold F	Drift-flag threshold (default 3.0).
drift	--warmup-n N	Observations before drift can fire (default 20).