If there is any sort of analog or digital filtering or noise supression involved in the learning then one would expect to get unmodulated for any signal well above the frequency learning was designed for. What I didn't expect was for the pulses to be three times longer than in what we believe is the correct signal.

I don't really know how to proceed.

Anyone who has both a TSU3000 and any good IR learning device could investigate the nature of the bugs in the firmware and/or editing program and probably deduce a work around. I'd be glad to help with the analysis, but I don't have a TSU3000. I assume you don't have any good IR learning device.

If one knew or reverse engineered the format of the PCF file, one could investigate the bugs in the editing program.