CRJ700 Captain reports early turn to the LOC course with the autopilot armed for an ILS. The problem occurred to this reporter previously but with no predictability.

Narrative

Incorrect capture of localizer by FGC. ATC gave us vectors ILS. When we were given clearance for the approach; the final intercept heading was 250 degrees. This approach was as stable as it gets. Wings level; a 30 degree intercept angle; and exactly on the glideslope; with autopilot engaged. At 12 DME; we were still north of the localizer; with the localizer needle still on the peg. A perfect vector by ATC. That's when the localizer was captured by the FGC -- with the needle still on the peg. The FGC commanded a right turn. I got it stopped at 275 degrees; using the heading bug; turned back to the 250 degree intercept heading; and reprogrammed the approach. All of this occurred with the needle staying on the peg. ATC made a comment about the early turn; but also saw that we were correcting. A proper capture occurred at 7 DME; and the rest of the approach was uneventful. It's unknown why this happened. It should be noted that even though this is a problem that some pilots are aware of; it is still not possible to predict which failure mode will occur. In this case; the airplane turned to the right toward a parallel heading; but it could have turned to the left; flown through the localizer; and performed at least one 'S' turn before stabilizing. The technique of delaying the arming of the approach function until the localizer comes alive must be used with caution to avoid an overshoot and can lead to two problems. One is that any overshoot of the localizer is likely to get a comment from ATC. The other problem is the penetration of the No Transgression Zone when traffic is on a parallel ILS. Fortunately there was no traffic on the parallel approach in this case; since the other ILS had been shut down three weeks prior to this event.