Curve quality checking?

[ElectroLund]

I've been tasked to replace the human eyeball with some software automation  😆

We've got a calibration routine that produces a response curve which is exponential.  There are times when the calibration fails and we get a linear or otherwise "bad looking" curve.  

Here's a good response curve (in red) below.  Overlapping in blue is CVI's polynomial curve fitting from the advanced analysis lib.  Very powerful.

Honestly, I didn't comprehend all the data science, so I just made a quick tool to experiment with the output of functions like PolyFitEx, ExpFitEx, LinearFitEx.  Poly gave me the closest fit at 8th order and QR algorithm.

Here's the issue though.  It's one thing to fit the cal curve I have captured from this device.  However, that won't help me determine criteria for pass/fail.

I'm tempted to just take an average of "good" looking curves and then offset this for upper and lower thresholds and then check curves for being in this bound.

Thoughts?