As far as rolling resistance, I understand that most of the Crr values posted tend to be from roller tests. Some people run their own tests in controlled outdoor tests, but that is less public or reliable.
Ultimately, I think an important aspect of rolling resistance has been neglected: energy lost due to vibration.
While roller tests can provide a good understanding of how a tire itself loses energy under the stresses of rolling motion, there is a larger component of energy loss that is the kinetic energy that is transferred to the bicycle and rider. So instead of the energy summation being that of tire deformation losses and aerodynamic losses, there's an additional cost of unnecessary accelerations. I believe this is what causes the "runaway" Crr values when high tire pressures are tested in real-world conditions.
Since the vertical compliance of most road tires is limited ( I personally get pinch-flats too often if I start a ride at less than 110psi), I would think that a road bike or TT bike with damped axle mounts would be more efficient in real world conditions. By this, I don't mean the gimmicky systems some manufacturers introduce to provide "vertical compliance," because those still transfer energy to the frame and rider and are not likely to significantly reduce a metric like Grms (https://femci.gsfc.nasa.gov/random/randomgrms.html).
Another related thought to real-world testing: There should be a real-time function that divides N-second-avg-power by speed. Convergence can be checked in real-time rather easily and say, turn "green". Real-world testing could then proceed with A-B-A-B... testing, where repeated runs could be checked to see if the baseline efficiency has changed. A change in constant wind or slope would cause a "pure" shift in efficiency, while gusts or road quality variance would cause a change in variance in the data (also affecting time or samples to convergence). The real-time tests can then be examined later with more confidence using Aerolab or similar, to extract Crr or CdA values. Using variance (in particular within regions where one does not expect variance) or tests for shifts in data could also be used to correct traditional methods.
Ultimately, I think an important aspect of rolling resistance has been neglected: energy lost due to vibration.
While roller tests can provide a good understanding of how a tire itself loses energy under the stresses of rolling motion, there is a larger component of energy loss that is the kinetic energy that is transferred to the bicycle and rider. So instead of the energy summation being that of tire deformation losses and aerodynamic losses, there's an additional cost of unnecessary accelerations. I believe this is what causes the "runaway" Crr values when high tire pressures are tested in real-world conditions.
Since the vertical compliance of most road tires is limited ( I personally get pinch-flats too often if I start a ride at less than 110psi), I would think that a road bike or TT bike with damped axle mounts would be more efficient in real world conditions. By this, I don't mean the gimmicky systems some manufacturers introduce to provide "vertical compliance," because those still transfer energy to the frame and rider and are not likely to significantly reduce a metric like Grms (https://femci.gsfc.nasa.gov/random/randomgrms.html).
Another related thought to real-world testing: There should be a real-time function that divides N-second-avg-power by speed. Convergence can be checked in real-time rather easily and say, turn "green". Real-world testing could then proceed with A-B-A-B... testing, where repeated runs could be checked to see if the baseline efficiency has changed. A change in constant wind or slope would cause a "pure" shift in efficiency, while gusts or road quality variance would cause a change in variance in the data (also affecting time or samples to convergence). The real-time tests can then be examined later with more confidence using Aerolab or similar, to extract Crr or CdA values. Using variance (in particular within regions where one does not expect variance) or tests for shifts in data could also be used to correct traditional methods.