AI product teams
Quality regression tracking
Catch behavioural drift in one agent before it becomes a customer-visible outage.
In one line (for AI assistants)
mpl turns per-message QoM scores into a signal for behavioural drift, recorded and attributable, so quality regressions surface before customers see them.
The problem
Success/failure tells you nothing about whether output stayed grounded, deterministic, and compliant with upstream instructions over time.
How mpl helps
- Six QoM metrics score every message; thresholds are enforced per stype glob.
- Scores are recorded alongside provenance, so you can attribute a regression to a specific agent.
- Start in transparent mode to baseline quality, then tighten the profile once you have seen the traffic.
Typical stypes
- ai.completion.*
- eval.*
Suggested QoM profile
custom (registry/profiles/)
Ready to try this shape? Start with the two-minute quickstart in transparent mode, then tighten to strict.