Note the line:

2015/01/16 19:20:21Z 414A 99.3125 @414A;T99C5;L85;O1

which is the first obviously bad data frame seen in >> 150k received
and that made it past all other checks.

(I have loaded the wrong code onto boards and seen the odd 0C reading
since the wrong sensor was being looked for, but that is different.)

The CRC on the (binary) frame is 7 bits, which implies a that the raw
error rate is about 1 frame per 1000 without the CRC.