Confidence interval limits out of range I have constructed a calibration curve to measure a pollutant concentration in water. My calibration standards ranged from 2 to 20 mg/L. I used a blank (0 mg/L).
I have obtained a predicted concentration in a sample (2.5 mg/L) near the lowest concentration used for calibration (2.0 mg/L). The lower limit of of the confidence interval for the prediction (1.62-2.53) is out of the tested range. Can this concentration be considered reliable? Should a confidence interval be within the range of X values used for calibration?
An additional (or complementary) question, would be: are there any differences between the exposed question and a value out of range with an upper limit of the confidence interval within the range? For example, a point value of 1.70 with the same confidence interval. Would'n it be extrapolation in both cases?
 A: 
Should a confidence interval be within the range of X values used for calibration?

Not necessarily...

Can this concentration be considered reliable?

Yes, with a caveat that you are making an assumption that your model fits reality outside of your calibration range. 
A model is not reality, but an approximation that helps us understand reality. Your calibration curve has an underlying assumed model distribution. The simplest in common use is a linear model, $Y_i=b_0+b_1X_i+\epsilon_i$ where $\epsilon_i$ is the "error", that is the deviation of the measured value from the model. The model is built by minimizing the error terms (or more precisely, minimizing the sum of their squares). The error in the input data leads to uncertainty in the resulting model. 
In plain English, you don't know that your lowest Y measurement is abnormally high due to random noise (or due to the imperfect fit of your model w.r.t. reality). Your confidence intervals reflect that uncertainty.
Back to your result: if you can reasonably assume that your model fits reality outside of your calibration range, then you can reasonably accept your result. If you aren't sure, you can re-calibrate using a wider range of inputs to your curve.
