Level 1: Bounded Document Refinement

This example is the smallest complete loop in the repository: generate a draft, critique it, revise it, and stop as soon as the work passes both deterministic checks and the critic threshold.

What the loop enforces

• Generator-critic pattern: separate drafting, critique, and revision agents.
• Word count bounds: each draft must land between 180 and 260 words.
• Style checks: the draft must include a concrete example, distinguish generation from verification, and avoid unsupported citation phrases.
• Bounded iteration: the controller defaults to five iterations, fifteen model calls, and a stagnation threshold of three.

Flow

The model does not get to declare success on its own. The critic score matters, but deterministic validators still veto completion if the draft is too long, lacks a concrete example, or drifts into unsupported “according to…” style claims.

Example configuration

This is the real run configuration from src/adk_loop_lab/examples/level_1_refinement/example.py:

run = LoopRun(
    example_id="level_1_refinement",
    goal=(
        "Write a concise explanation of why deterministic verification is "
        "necessary in agentic loops."
    ),
    budgets=BudgetConfig(
        max_iterations=max_iterations or 5,
        max_model_calls=15,
        stagnation_threshold=3,
    ),
)

The deterministic checks are equally direct:

results.append(range_validator(word_count, 180, 260, "word_count"))

results.append(
    not_contains_validator(
        normalized,
        ["according to", "studies show", "research indicates", "experts agree"],
        "unsupported_citations",
    )
)

Why it matters

This example is intentionally narrow. It shows that even a small loop benefits from explicit state, bounded retries, and a verification phase that can reject superficially good language when it misses concrete constraints.

Continue to Level 2 for evidence tracking and gap-driven iteration.