Trace sources: score traces kept in MLflow

A TraceSource that reads an agent harness’s MLflow traces, scores them and writes the scores back.

Source: examples/sources. Read the guide for the concepts behind it.

studio-mlflow.yaml

# A trace source: evalsid reads the traces an agent harness already keeps in
# MLflow, scores them with the project's online policies, and writes the
# scores back as MLflow assessments (docs/guides/trace-sources.md).
#
#   kubectl apply -f studio-mlflow.yaml          # with the operator
#   evalsi source apply -f studio-mlflow.yaml --server $URL
apiVersion: evals.si/v1alpha1
kind: TraceSource
metadata:
  name: studio-mlflow
  labels: {evals.si/project: support}
spec:
  connector: mlflow
  endpoint: https://mlflow.example.com
  locations: ["1"]                       # MLflow experiment IDs
  # A Secret the chart mounts (sources.secrets), granted to the project for
  # this host in credentials.grants as file:mlflow-token/token.
  credentials: {file: mlflow-token/token}
  backfill: {since: 720h}                # read 30 days of history first
  poll: {interval: 30s}
  maxTraceDuration: 30m                  # a trace running longer is missed
  policies: [support-agent]
  writeBack: {enabled: true}