converge-ai 0.1.0

Does your AI-generated code actually do what you asked?

Converge

Does your AI-generated code actually do what you asked?

Converge validates AI-generated code against the original intent using three pillars:

Pillar	Question
Intent Fidelity	Does the code do what was asked?
Optimal Design	Is this the most natural solution?
Code Consistency	Is the code internally coherent?

Install

pip install converge
export ANTHROPIC_API_KEY=sk-...

Usage

from converge import ConvergeValidator

validator = ConvergeValidator()
result = validator.validate(code=my_code, intent=my_intent)

print(result)
# Converge — Iteration 1
# Verdict         : ✅ CONVERGED
# Intent Fidelity :  92.0  (threshold 80)
# Optimal Design  :  85.0  (threshold 80)
# Code Consistency:  88.0  (threshold 75)

if result.verdict != "CONVERGED":
    improved_code = agent_iterate(result.feedback)

Iterative loop

from converge import ConvergeValidator, ConvergeLoop

loop = ConvergeLoop(ConvergeValidator(), max_iterations=5)
for result in loop.run(initial_code, intent, generate_fn):
    print(result)
    if result.verdict == "CONVERGED":
        break

Verdicts

Verdict	Meaning
CONVERGED	All 3 pillars above threshold — ship it
PARTIAL	2/3 pillars OK — iterate on feedback
NON_CONVERGED	Needs significant rework

Without API key

Converge works offline using a fast syntactic analyser (no LLM calls). Set ANTHROPIC_API_KEY for full LLM-powered validation.

License

MIT