Evals and code evidence answer different questions.
Use runtime evaluations to understand behavior on a task. Use a Metal Mantra report to inspect signals in the source. Read both with their scope and limitations.
What LangSmith and Langfuse contribute
LangSmith supports offline evaluations on datasets and online evaluations of production outputs. Its evaluators include code and LLM-as-judge methods. See the official evaluation documentation.
Langfuse supports evaluation of live traces and experiments on predefined datasets, with manual or automated evaluators. See the Langfuse evaluation overview.
These results help a team compare changes and investigate behavior. Their meaning depends on the test cases, evaluator, model and application version.
What Metal Mantra contributes
Metal Mantra publishes automated static signals from selected source files, with file references, scan dates and coverage. It can identify references to tools such as LangSmith and Langfuse; an import or declared dependency does not prove that tracing is enabled or that evaluation succeeded.
A scan is never a security certification. A clean selected file set cannot establish that an agent is safe at runtime. Evaluation results and code scores are not interchangeable.
Browse code reports referencing LangSmith or Langfuse. Search results depend on what the scanner read, not every integration a project may use.
Attach an evaluation with enough context
Sign in to the account managing your current public report and open its Evaluation results section. Attach a public results link from LangSmith, Langfuse or another tool, or a redacted report hosted elsewhere. Include the dataset and sample count, evaluator and scoring method, results and limitations, agent commit or release, and evaluation date.
Links must use HTTPS without credentials, query parameters or fragments. If an experiment is private, publish a redacted summary you are entitled to share. Do not publish secrets or personal data.
One submission is shown per repository. The managing account can replace or remove it. It is labelled self-reported: Metal Mantra has not verified maintainer identity or reproduced the evaluation, and it never changes the automated score.
How to read the evidence together
Check that the evaluated version matches the version you intend to run. Compare the evaluation dataset with your own workload, inspect failures as well as averages, and check whether the source scan covered the relevant files.
A high eval result can coexist with a serious code finding. An old eval may no longer describe current code. Missing results mean unknown, not failure. Neither tool provides a blanket guarantee that an agent should receive credentials or sensitive data.
Metal Mantra is independent and is not affiliated with LangChain or Langfuse. Read the user guide or scan a public repository.