rea-workbench

You have a black-box business rule and a pile of input→output traces. This workbench recovers the rule, proposes a rebuild, validates evidence claims, runs behavior evals, and emits a pass/fail HTML report — in one command.

Not another empty CI framework. A product you can try in under two minutes. The gate library it consumes is rea-rebuild-gate.

The problem

Teams reverse-engineer fees, percent buttons, ramps, and other opaque rules from production behavior. Agents then rewrite. A demo that “runs” is not enough: you need recovered claims with confidence, a rebuild that matches the traces, and a human-readable audit before you ship.

Try it (< 2 minutes)

Stdlib Python 3.10+. No install, no API keys, no network.

git clone https://github.com/BrickerP/rea-workbench.git
cd rea-workbench
PYTHONPATH=src:vendor python3 -m rea_workbench demo

Open docs/demo-report.html in a browser.

RESULT: PASS — HTML report → docs/demo-report.html

What the demo owns end-to-end:

case	traces	recovered rule	evals
fee_schedule	24	3 tiers (rate+floor) + under-$50 surcharge	24/24
percent_rule	12	+/− vs ×÷ percent value formulas	12/12
speed_ramp	8	SPEED / ACCELERATION / MAX_SPEED clamp	8/8

What one run does

1. Load I/O traces (fixtures generated locally; your own JSON works too).
2. Recover a candidate rule from traces only (black-box source never imported into the rebuild).
3. Propose rebuild.py + evidence.json + evals.json.
4. Gate via vendored rea-rebuild-gate: confidence floor, claim→eval coverage, behavior evals.
5. Emit docs/demo-report.html with PASS/FAIL, claims, and per-eval rows.

Your own traces

PYTHONPATH=src:vendor python3 -m rea_workbench run fee_schedule \
  --traces path/to/traces.json \
  --out /tmp/fee-art \
  --report /tmp/fee-report.html

Trace shape:

[
  {"id": "t1", "input": {"amount": 100, "tier": "gold"}, "output": 2.0}
]

Supported recoverers today: fee_schedule, percent_rule, speed_ramp.

Layout

src/rea_workbench/   # product: recover → propose → gate → report
vendor/              # rea-rebuild-gate (dependency, not the product)
fixtures/traces/     # sample I/O traces
docs/                # demo HTML + screenshots from our run
reports/             # latest artifacts

Development

PYTHONPATH=src:vendor python3 -m unittest discover -s tests -v
PYTHONPATH=src:vendor python3 -m rea_workbench demo

Author

Built by Yupeng Lu — AI agent & backend engineer. I take agents from demo to production: MCP tools, guardrails, evals, billing, real-time systems.

Hire me: https://brickerp.github.io/hire/

MIT License.