Your first experiment¶

from whatifd.adapters.stub import StubTraceSource, StubTraceSpec
from whatifd.adapters.protocols import RawTrace
from whatifd.cache.summary import CachePolicySnapshot, CacheSummary
from whatifd.pipeline import run_pipeline
from whatifd.serialization import encode_report_v01
from whatifd.types.manifest import EnvironmentFingerprint, RunManifest
from whatifd.types.policy import DecisionPolicy, TrustFloor
from whatifd.types.statistical import (
    BootstrapMethodDisclosure,
    EffectSizeDisclosure,
    JudgeMethodDisclosure,
    MethodologyDisclosure,
    MultiplicityDisclosure,
)

specs = [
    StubTraceSpec(trace_id=f"f-{i:02d}", user_message=f"failure {i}",
                  original_response=f"resp {i}", cohort="failure")
    for i in range(20)
] + [
    StubTraceSpec(trace_id=f"b-{i:02d}", user_message=f"baseline {i}",
                  original_response=f"resp {i}", cohort="baseline")
    for i in range(20)
]
source = StubTraceSource(specs=specs)

# In a real run, this comes from your runner + scorer.
# Here we make it deterministic so the example is reproducible.
def delta_fn(rt: RawTrace) -> float:
    return 0.4 if rt.cohort == "failure" else 0.05

report = run_pipeline(
    source=source,
    delta_fn=delta_fn,
    floor=TrustFloor(),
    policy=DecisionPolicy(),
    manifest=RunManifest(
        run_id="demo-run",
        whatifd_version="0.1.0",
        env=EnvironmentFingerprint(...),
    ),
    methodology=MethodologyDisclosure(...),
    cache_summary=CacheSummary(
        mode="off",
        profile=CachePolicySnapshot(...),
        hits=0, misses=0, writes=0,
        stale_hits=0, corrupted_entries=0,
        schema_version="v1", key_version="v1",
        models_distribution={},
    ),
)
print(report.verdict_state)        # "ship" / "dont_ship" / "inconclusive"
print(encode_report_v01(report))   # JSON

from whatifd.contract import TraceInput, ReplayConfig, ToolCache, ReplayOutput
from anthropic import Anthropic

client = Anthropic()

def run(
    trace_input: TraceInput,
    config: ReplayConfig,
    tool_cache: ToolCache,
) -> ReplayOutput:
    response = client.messages.create(
        model="claude-haiku-4-5",
        max_tokens=1024,
        system=config.system_prompt or "You are a helpful assistant.",
        messages=[{"role": "user", "content": trace_input.user_message}],
    )
    return ReplayOutput(text=response.content[0].text)