Nolma Lens

Lens turns user behavior into actionable recommendations. It answers: are users actually happy with what your agents produce?

How it works

User interacts with AI output
        |
Your code calls nolma.signal()
        |
Lens tracks acceptance rates,
edit distances, retry rates
        |
After 100+ signals: Claude generates
evidence-backed recommendations
        |
Dashboard shows: "Switch to gpt-4o-mini,
save $43/mo, low quality risk"

What Lens measures

Metric	Description
Acceptance rate	% of outputs used without changes
Edit rate	% of outputs the user modified
Edit distance	How many characters users change
Retry rate	% of times user regenerated
Abandon rate	% of times user left without using
Downgrade score	0-100 readiness to use cheaper model

Downgrade readiness score

Calculated weekly per agent:

Component	Weight
Acceptance rate	40%
Stability (low retry rate)	25%
Focus (low abandon rate)	20%
Consistency (low edit rate)	15%

Score >= 75 → Shadow mode recommended
Score >= 90 → Safe to switch models

Getting started

Add 2 lines to your agent code:

# After user acts on the output:
await nolma.signal_async(session_id, "accepted")

That’s all. Lens starts working after 100 signals per agent.

​Nolma Lens

​How it works

​What Lens measures

​Downgrade readiness score

​Getting started

Nolma Lens

How it works

What Lens measures

Downgrade readiness score

Getting started