I don't disagree with anything you wrote. This is a complex problem, so of course the solution (or "solution space" in a universe of possible behaviors) is not "just" a matter of anything. But in order to discuss it in a forum such as this, one has to be able to propose a necessarily simplified description of an approach that seems promising. That's the intent of my subsequently posted prioritized list and I think it's a pretty sensible one, in response to @sleepydoc's prior question about what to do in conflicts between Vision and map.The trouble is it’s not quite as clear-cut as that. Based on the few hints we’ve had so far it’s pretty clear that the Vision system uses map data as an input to the NN .. so if the Vision system isnt sure (to use a too-crude example) if the road ahead has 2 or 3 lanes, it can use map information to resolve the ambiguity. Of course its much more complex than that, but it means that when map data and vision data differ the car may reach a point where its confidence level is too low to continue. Sure, as the vision system improves Tesla can weight that more heavily than map data, but you are still going to face times when the two are so dissonant that the car cannot progress by itself (which seems pretty sensible for an L2 system).
There is also safety to be considered. If the vision system sees a 50 mph road sign with 60% confidence, but the map system has a 40 mph limit, which should the car choose? If the car saw a 25 mph sign with the same 60% confidence, which should it choose? So it’s not just a matter of correctness, its a matter of what is the safest course of action.
I don't automatically subscribe to the notion that the AV should respond to everything the way a human would, but it's a reasonable principle when in doubt - first because other road users expect it and this is likely to be true for the foreseeable future, and second because the road engineering, to the extent it is properly executed, is based on providing human drivers with mostly familiar visual cues.
I'd be completely ready to say that the AV should take advantage of its specific and sometimes superhuman capabilities, to make up for its lack of human experience and intuition. And in theory, this could include the ability to utilize an accurate and precise (HD) map. However, even in the best of circumstances, if Tesla devoted significant resources maintenance of such maps for the entire ODD, they still cannot keep up with minute-by-minute developments. And we know that the current Tesla maps are very far from this; they seem to have many gaps, errors or insufficiencies - in fact they seem to be generating more than their fair share of map-vs-vision problems. Because of this reality, I conclude that FSDs reliance on maps should be about the same as a human driver's reliance on maps and/or memory of the route: a guide and expectation, but in no way a substitute for real-time sensory perception of the environment.
I certainly agree also that unexpected differences should be more heavily weighted in favor of "safer" behavior. If the car doesn't see a stop sign or a one-way restriction that the map says is there, that's not an excuse to ignore it. On the other hand, driving slower or stopping is not necessarily safe behavior in high speed traffic. That's why I suggest prioritizing proximate observed traffic behavior above the map info, to help resolve such dilemmas.
Finally, I really enjoy this kind of discussion, where the participants goal is to explore possible solutions and approaches. None of us has a fully-formed, unassailable solution, but including too many disclaimers in the proposals don't insulate against that; they just clog up the discussion. For some people, the point of the back-and-forth is to Be Right, draw sides and look for soft spots to tear down the wall. I don't perceive that at all in your posts, so I responded and I'll just reiterate that I still feel good about my prior comments; they're not meant to be the basis of simple codable Self-Driving Laws, but a proposed framework, implemented in the inherently softened logic of NNs, for resolving the well-known data conflicts under discussion.