AMOS: If reversed sign after CMB-correction, correlate errors

From structural model to CMB-corrected structural model, sometimes the key parameter will change sign and/or become insignificant. We can follow modification indices or connect the relevant errors, to balance the loadings in complex models.

In title picture, while all "abc ghi" were positive value in structural model, after adding CLF b and c become negative. Then I added a corelate curve between error 2 and 4, then b c become positive but h become negative. I added another corelate between error 3 and 4, now they all become positive.

I think this is because AMOS allocate correlation matrix wrong in complex models, so we can give more direction about the true model that we are testing. Welcome any critics and thoughts.

This is the final curves that makes b c positive, without making others negative: