[Suggestions for realizator]
In the linked function, the vector ekk looks like a vector of the absolute projection errors (I do not remember which projections are supposed to be done, I suppose one is left points on right image given the current parameters relating the two cameras, but for sure ekk concatenates the errors of two projections).
The assert makes the function fail when this error is large for at least one point. I do not understand the math behind this function at first glance, but I guess that ekk should logically be close to zero, except if there are bad inputs in the first place.
I think it would be impractical to try to find errors based on this metrics, you better check the input data another way.
In particular, I would guess that if you drew the detected chessboard with cv2.drawChessboardCorners, you would see either bad detection or detected points that are not homologous in the two images (rotated, skipped corner or whatever). You might see errors in other pairs too, but if the resulting error is small enough it might just screw up the computed parameters silently.
In stereomaton v2, if ever I manage to find time to work on this project, I planned to use ChArUco board to define the points used in the stereo calibration instead. This board has oriented, numbered and encoded tags so that each corner has a unique identifier whatever the orientation of the pattern relative to the camera (works even if a part of the board is cut). It also has large corners where subpixel position refinement can be computed. A prior check of the points can be done based on their ID (is this ID detected on both images of this pair) which, in my opinion, will remove most of the errors that several users reported in the forum.
See https://docs.opencv.org/trunk/df/d4a/tu ... ction.html