Verdict Python SDK

PyPI distribution: cognifity-verdict. Python import: verdict.

The Verdict Python SDK. Auto-instruments your LLM calls via wrapt and captures them into a vendor-neutral Trace schema (attribute names follow the OpenTelemetry GenAI semantic conventions, but no OTel spans are emitted). Traces are written to SQLite by default (or any Storage adapter). Content capture (prompts/completions) is off by default — opt in with capture_content=True; when enabled, captured content is run through built-in pattern redaction recursively across supported JSON-compatible message and tool structures before Trace assignment and again at storage. Card candidates use Luhn validation; IPv6 candidates use standard-library address validation so clock values such as 12:34:56 remain intact, and email candidates use a linear @-anchored scanner so malformed or very long input cannot trigger regex backtracking. Unsupported objects fail closed. Traversal is bounded by node and character budgets, and cycles or repeated container references fail closed at every occurrence so sanitized output never retains caller-owned aliases. Redacted mapping-key collisions keep every value under deterministic suffixed keys rather than overwriting one entry. This is best-effort matching, not a compliance guarantee; keep content capture off when its documented coverage is insufficient.

For a customer proof of concept, follow the versioned 0.1.0a4 POC release profile. It pins the package set, provider entry points, persistence mode, and privacy boundary used for release verification.

Supported streams finalize after full consumption, iteration error, explicit close() / aclose(), context exit, or async cancellation. Garbage collection of a never-iterated unclosed stream is not a persistence guarantee. A supported instrumented provider call made inside a manual span records the innermost span's ID in Trace.parent_span_id. This is the sole automatic direction because multiple provider traces may share one manual span; automatic capture never chooses one reverse SpanRecord.trace_id. Manual-only work can bind to an existing stored trace with trace_context(trace_id) or set_context(trace_id=...). An unknown explicit trace ID is recorded as an unlinked span with a link-status attribute rather than as an orphan; spans with no provider call or explicit context remain standalone.

Provider SDK unset sentinels and other non-primitive numeric metadata are normalized to unavailable (None) before a Trace reaches storage. A synchronous telemetry persistence failure never replaces the provider call's result or exception; Verdict emits one warning per provider, storage type, and exception type instead of flooding application logs.

sample_rate controls the fraction of supported calls retained, and buffered_writes=True moves persistence to a background batched writer. Stored manual spans do not wait for provider acknowledgement or receive repair writes; each ended span is persisted once independently of provider success. flush() is a FIFO point-in-time barrier and accepts an optional timeout. close() rejects new reads/writes, drains every accepted FIFO write, stops and joins the worker, then closes the inner adapter; post-close flush() is an idempotent no-op. The 0.1.0a4 POC profile uses buffered_writes=False. Buffered mode requires an explicit shutdown() imported from verdict.client before process exit. Completed drift analyses use atomic DriftRun snapshots, including explicit zero-signal runs. Storage readers select a run marker and its exact signals from one snapshot; deleting a matched attributed signal window removes the completed run as a unit. prune_before() removes expired standalone and orphan span rows while preserving an old span referenced by a retained Trace. SQLite and PostgreSQL execute multi-table trace deletion and pruning atomically and serialize concurrent trace writers while they decide which shared parent spans must survive. Stored costs are best-effort estimates from Verdict's dated static base-price table; unknown models remain unpriced, and the values are not billing truth.

import verdict
from anthropic import Anthropic

verdict.init(service_name="my-app", storage="sqlite:///./verdict.db")
client = Anthropic()
# Use Anthropic normally — supported SDK calls are captured.

For 0.1.0a4, supported POC entry points are Anthropic messages.create(...) (including stream=True), OpenAI chat.completions.create(...) and its stream helper, and Google models.generate_content(...) / generate_content_stream(...). Anthropic messages.stream(...) and the OpenAI Responses API are not captured in this release.

See the repository README for the full picture, the architecture decisions, and the examples.

Apache 2.0.