So the current heuristics are no longer valid (e.g. you were in a car wreck so you're hunched from back pain and have a black eye), and you're going to use the current heuristics as a gate to updating the heuristics?

Imagine if the "forgot password" link on a website gave you a form that said "enter your current password and then you can set a new password". This is the scenario you're describing.

Your solution to not being able to login is to login and change settings?

Think of it like Paxos. As long as a majority of unique data points agree, allow access. That way if your back is out which results in your posture being unrecognized, the other factors can still form a quorum and grant you access. There could be 20 such factors weighted by "difficulty to forge", so that a fingerprint counts 15x as much as posture for example. Then if you successfully authenticate with, say, 80% on all data points then the system can allow you to train a new factor.

Exactly! It becomes a machine learning problem. There was a post about this a while back ..

I was attempting to post some humor on HN, but it seems it's not quite welcome here. :)