Embedding AINL in Research Loops

Use AINL as a deterministic orchestration substrate inside evolutionary or self-improving agent loops.

In-memory Python flow

from compiler_v2 import AICodeCompiler
from runtime.engine import RuntimeEngine
from runtime.adapters.base import AdapterRegistry

ainl_source = """
S app api /api
L1: R core.ADD 2 3 ->sum J sum
"""

compiler = AICodeCompiler(strict_mode=True)
ir = compiler.compile(ainl_source, emit_graph=True)
if ir.get("errors"):
    raise RuntimeError(ir["errors"])

registry = AdapterRegistry(allowed=["core"])
engine = RuntimeEngine(ir=ir, adapters=registry, trace=True)
result = engine.run_label(engine.default_entry_label(), frame={})
trace = engine.get_trace()

Loop-friendly primitives

• ainl inspect workflow.ainl for canonical IR snapshots.
• ainl run workflow.ainl --trace-jsonl run.trace.jsonl for observable execution tape.
• MCP ainl_validate diagnostics now include llm_repair_hint.
• MCP ainl_fitness_report and ainl_ir_diff provide scoring and mutation signals.

Fitness score contract

ainl_fitness_report returns a bounded metrics.fitness_score in [0, 1] plus metrics.fitness_components for transparency.

Current weighted formula:

• reliability: 0.6
• latency: 0.2 (lower latency => higher component)
• steps: 0.1 (fewer runtime steps => higher component)
• adapter calls: 0.1 (fewer calls => higher component)

Use fitness_components.weights in tool output as the source of truth for downstream ranking logic.

Detailed MCP payload contract: docs/operations/MCP_RESEARCH_CONTRACT.md.