Yes, the "simpler" representation is TAI, but we still live in a human world, and with humans entering and reading the data we must conform to their ways in some things.
The binary and text encodings both store in Gregorian so that codecs don't have the added responsibility of converting between representations (keeps the implementations simple).
TAI is like lye: useful, but not meant to be handled by humans. And since this format is designed for human usability, human time wins the day.
When I look at human readable logfiles I see a trend towards using epoch seconds for exactly these reasons. It's so easy to convert, it outweighs the anti-usability aspects inherent in formatted strings. I've converted all my own human readable file formats to epoch times for exactly these reasons and my customers seem to strongly prefer it.
with humans entering and reading the data we must conform to their ways in some things
I don't understand what is stopping you from converting from tai to human readable and vice versa at the form/input and display/output layers, respectively.
What's stopping me is that time conversion is complex and incredibly difficult to get right. In order for a format to be successful, it must be usable on all platforms big and small, and across all programming languages. The more complex the task of implementing the format, the less likely it is to be successful.
TAI conversions are not equally available or of the same quality on all platforms and technologies, whereas Gregorian time is available pretty much everywhere and is well tested, robust, and easy to use (relatively speaking. Time is never easy).
If it turns out in future that I've made a serious blunder with this policy, I could just add a new type "TAI time" and bump the Concise Encoding spec version (Concise Encoding documents all specify which version of the spec they adhere to). There are still about 30 or so RESERVED type slots in the binary spec so we're not about to run out. We then get the type, just a little late. The alternative is that I make the format needlessly complicated by pre-emptively adding TAI, and risk destroying its adoption. Being right doesn't mean success...
The binary and text encodings both store in Gregorian so that codecs don't have the added responsibility of converting between representations (keeps the implementations simple).
TAI is like lye: useful, but not meant to be handled by humans. And since this format is designed for human usability, human time wins the day.