traceguard

Point-in-time correct LLM instrumentation — the time-integrity layer for LLM pipelines.

When you run LLMs over historical data — backtesting a signal, replaying a pipeline, re-scoring an archive — TraceGuard makes it structurally hard to accidentally use a model or prompt that did not yet exist at the point in time you are simulating.

It is not a dashboard or a gateway; it is the lower layer that guarantees the timeline underneath one. It interoperates with observability stacks (Langfuse / Phoenix via the optional traceguard[otel] exporter) rather than competing with them. See docs/POSITIONING.md.

• traceguard.registry.models — model registry with released_at / available_to_us_at; select_model(..., strict=...) with mandatory explicit mode (no default), so anachronistic choices fail loudly.
• traceguard.registry.prompts — git-tracked YAML prompt templates; load_prompt pins the content hash into every trace.
• traceguard.sdk.tracer — @tracer.trace decorator and tracer.span() context manager recording input hash, model/prompt versions, output, and perf into SQLAlchemy (SQLite by default).
• traceguard.sdk.normalizer — the single canonical normalize_input / input_hash (sorted keys, fixed float precision, normalized whitespace).
• traceguard.sdk.wrappers.anthropic — wrap_anthropic auto-instruments an Anthropic SDK client (extra: traceguard[anthropic]).
• traceguard.validators.lookahead — pure-function invariant validators (validate_feature_as_of, validate_model_timing, validate_reference_timing) that raise InvariantViolation; call them in pytest/CI.

Install

pip install traceguard

Requires Python 3.11+. Optional extras: pip install "traceguard[anthropic]" (Anthropic client wrapper) and pip install "traceguard[otel]" (OpenTelemetry / OpenInference export to Langfuse, Phoenix, or any OTLP backend).

Example

from datetime import datetime, timezone
from traceguard.registry.models import register_model, select_model
from traceguard.store.models import make_engine

engine = make_engine("sqlite:///traceguard.db")

register_model("demo-llm-2024", model_family="internal-ml",
               capability_class="general-llm",
               released_at=datetime(2024, 1, 10, tzinfo=timezone.utc),
               available_to_us_at=datetime(2024, 2, 1, tzinfo=timezone.utc),
               engine=engine)

# Backtesting as of mid-2025: models that arrived later are invisible.
model_id = select_model("general-llm",
                        available_at=datetime(2025, 6, 30, tzinfo=timezone.utc),
                        strict=True, engine=engine)

A complete runnable tour (synthetic data, no API keys) lives in examples/quickstart.

Contract

The binding interface contract — table schemas, SDK signatures, the four look-ahead invariants, SemVer rules — is in docs/SPEC.md.

Phase 0 scope: tracer, model/prompt registries, normalizer, invariants 1–3, Anthropic wrapper. Not yet: drift checks, replay sets (invariant 4), CLI, Postgres/TimescaleDB, OpenAI/Voyage wrappers — see TRACEGUARD_ROADMAP.md.

Development

cd packages/traceguard
uv sync
uv run pytest        # 70 tests

License

Apache-2.0.