README.md

simple-python — the ResearchForge launch demo target

A deliberately small sentiment classifier with a deterministic benchmark, so the full ResearchForge journey runs offline in minutes and produces the same numbers on every machine. See docs/demo.md for the walkthrough.

• src/classifier.py — keyword classifier with two real weaknesses.
• src/config.py — the tunable settings experiments patch.
• benchmarks/evaluate.py — writes artifacts/results.json with f1 and a deterministic p95_latency_ms proxy; --quick is the screening subset.
• researchforge.example.yaml — a ready-to-review contract (copy to researchforge.yaml); latency budget: p95_latency_ms <= 200.

Known behaviors (why it demos well):

Change in src/config.py	f1	p95_latency_ms	Outcome
baseline	0.75	72	frozen reference
NORMALIZE = True	0.90	72	the winner
NGRAM_EXPANSION = True	0.82	312	better f1, violates the constraint
a broken import	—	—	recorded failure

Try it:

cp -r examples/simple-python /tmp/demo && cd /tmp/demo
git init -b main && git add . && git commit -m "baseline"
researchforge init --claude    # then drive it from Claude Code, or:
cp researchforge.example.yaml researchforge.yaml