AD725 outputs wrong colors

I designed device based on AD725ARZ. Device is built basing on the reference circuit diagram on page 12.

TV is connected to **composite output of the circuit**. It does not have S-video input, thus no way to test the S-video image (at least for now).

There're two encoders connected in parallel to the CSYNC and RGB lines:

CXA1645-based circuit, and it display correct colors:

AD725ARZ-based circuit, and it displays incorrect colors:

The obvious idea comes to mind is that I just confused R and B input wires, but it is NOT the case. I checked the wiring several times, and also looked with scope onto the input pins of the AD725: the pin 8 has most activity, which must be the blue color (as picture consists 90% of blue).

The standard used is NTSC. Picture is stable, which means that sync is OK. Board is wired using air-wires, suffers from some degree of dot crawling but that's a problem of wiring and cabling.

Now another piece of information: the board based on AD725 has logic to switch to PAL in case it detects vertical sync of 50 Hz (you can see the related stuff in the reference circuit in AD725 datasheet). When machine switches to PAL mode, colors change! Here is the pic:

And now we see green color instead of blue one! But that's not all... even in NTSC 60 Hz mode, if I force AD725 to work in PAL mode, it starts showing same green screen instead of red...

Any ideas what could be the cause?

What can I do to troubleshoot the problem further?

**Update 1:** board images as requested by Wendall. It is a prototype. CXA1645 properly working board is made the same way.

**Update 2:** It seems there's a DC offset at all the color inputs of the chip, and sometimes colors saturate. I can not understand the source of it, all three colors are decoupled using 0.1 uF.

But CXA1645 does also have DC offset of about 3 V at its inputs, and works properly. I guess it performs some kind of leveling inside removing common DC component out of R/G/B.

**Update 3:** here's circuit diagram.

**Update 4:** I have removed all the logic out of equation, connecting pin 3 directly to 14.31818 generator, 16 to TTL CSYNC signal, 15 to +5V, and 1 to +5V (together with another arm of YTRAP circuit). Result is the same - red instead of blue. Seems chip is defective.