AnyoneLearnAI
Start learning
Chapter DLLM tracing labPage 4 of 8

LLM tracing lab

Validate outputs and schemas

LLM tracing lab is production work only when one frozen failure can be reproduced, one measurable gate can stop a release, and one operator can safely reverse it.

~14 minValidation

Before you start

Why this matters

Read this incident aloud: a user reports a wrong answer and the trace must distinguish empty retrieval from a generation defect. In two minutes, write the earliest deterministic check that should fail, the telemetry signal you would inspect, and the action that must not happen automatically. Compare your answer with this chapter's boundary: span attributes contain IDs, counts, model versions, and redacted arguments—not prompts, API keys, or PII.

1Learn the idea

Read

Test behavior and evaluate quality

Testing this system requires more than checking that JSON parses. Build a small matrix with one normal case, one boundary case, the known failure, and one adversarial case. Freeze randomness and time. Stub the provider with recorded semantic outcomes rather than brittle prose snapshots. Then assert the decision, prohibited behavior, emitted metric, and absence of sensitive fields. The key behavioral assertion for this topic is assert [s.name for s in trace.spans] == ["request","embed_query","retrieve","generate"].

Evaluate at two resolutions. First, case-level reasons must tell a developer exactly which expectation failed. Second, aggregate complete trace ratio and per-span p95 duration must meet at least 0.99 context propagation and no sensitive attributes in exported spans on the frozen set and on important slices. A global pass can hide a severe intent, carrier, release, or customer workflow. Report numerator and denominator beside every rate; 0 failures over two cases is not strong evidence.

Before accepting a metric, attack it. If a shorter refusal raises a keyword score while usefulness collapses, the metric is being gamed. If lower cost excludes retries, the denominator is wrong. If latency ignores timeouts, the sample is censored. Compare automated scores with a small blinded human review and record disagreements as future fixtures.

For validation, execute the full matrix and print a machine-readable report plus a short developer summary. Exit nonzero on a violated critical gate; do not round a failing value into a pass.

Read

Focused implementation artifact

import pytest

CASES = [
    pytest.param({"request_id":"req_7b9f2","release":"rag-v12","hashed_user_id":"u_91c","sampled":true,"query":"[REDACTED]"}, id="known-boundary"),
]

@pytest.mark.parametrize("case", CASES)
def test_behavioral_gate(case):
    with tracer.start_as_current_span("retrieve") as span: chunks = store.search(query, k=4)
    assert [s.name for s in trace.spans] == ["request","embed_query","retrieve","generate"]

def test_release_gate(report):
    assert report.sample_count >= 10
    assert report.critical_failures == 0

Read

Build the evaluation report

Expand the parameterized test into four named fixtures: benign control, threshold boundary, known regression, and adversarial misuse. For every row, assert the decision and at least one negative fact, such as no leak, no unauthorized action, no duplicate write, or no candidate promotion. Provider prose can vary; policy outcomes and required evidence cannot.

Aggregate the case results into complete trace ratio and per-span p95 duration and preserve numerator, denominator, critical failures, and slice keys. Apply at least 0.99 context propagation and no sensitive attributes in exported spans exactly. Compare baseline and candidate on the same fixtures, then review disagreements where the automated score and a blinded human label diverge. Those disagreements are useful data, not noise to discard.

The regression named the retrieve span reports zero chunks but an uncorrelated model log makes generation look guilty must fail before the fix and pass after it. Verify llm_trace_complete_ratio is emitted for both outcomes and that the failing report points to fixture IDs and trace IDs without embedding sensitive content. A failed gate leads operators to mark retrieval as the fault domain and link the trace ID in the incident without copying prompt content.

Checking tutor…

Continue learning · glossary & guides