I don't fake my numbers, and especially not on such a corner market as the old HD7000/6000 mining Heavy. However, the result are intriguing, what version do you use? If that's the B-X, so it could be false negative shares (good shares badly detected as wrong and thus not sent to pool) which would say my hybrid sync still need fixes, even on HD7000.
If that's on the b5/b6, so... i've little explaination so far, except that my new implementation itself (not counting the hybrid parts) is glitchy and already produces a lot of false negative. In such case, it would be a very major bug. I'll do some emergency tests, lucky i've a lot of old cards to test on.

A less extreme explanation would be that you used (explicitly or not) a very high Diff, making your fast card on ease but not your slow cards. In such case, the effective hashrate of the slow cards would start low and converge to the theorical hashrate after a very long time. Like when you mine with CPU on Nicehash.

for HD6950, a user on Github advises to use worksize=32 and multi_hash=864, one one thread.
Take a look at the HD6000 user topic here:
https://github.com/jceminer/cn_gpu_miner/issues/26

also try with worksize=16 and 8

edit: hashing Ryo with low Diff (5000) on my dual HD7950 (totaling 845h/s) and 840h/s effective hashrate so far with the B-x, with 1% bad shares, as expected. I strongly advise you redo your test with either a fixed low Diff or your big GPU paused.