As the LED warms up its light output drops, so I don't see it as a "regulator" failing - more than likely it is poor heatsinking and/or poor/no thermal paste, etc..
The other option is that the cell's can't keep up with the current demand, and quickly sag to their steady state voltage given the current draw. It would be great to drive this head from high current cells or a stiff power supply to test this theory.
Has anyone seen one of these open to look at how the LED is attached?
Will
The graph shows a drop to 80%, with almost all of it occurring within 1-2 minutes, and the balance within 20 minutes or so. If it's temperature related, and assuming they can't raise the current in response to rising junction temperature because they don't want to exceed the 3 A LED rating, that would correspond to an increase in T sub J from 25 C to about 120 C (ref:
http://www.cree.com/products/pdf/xlampxm-l.pdf). Perhaps that is what is happening. I wouldn't think anyone would design a light that runs that hot in terms of T sub J - that's way above the boiling point of water - but I suppose they might have.
As for regulation ... Justin Case convincingly calculates total LED + driver power of 11 W, which is only 1.8 W per cell. That's nothing for a CR123A. Figure 2.4 V at 0.75 A. A good CR123A (i.e., Paranasonic, Duracell, or Energizer) will run over an hour before hitting 2.4 V (ref:
http://www.powerstream.com/cr123a-tests.htm), and even half-decent CR123As will run 40 minutes before hitting 2.4 V.
If Surefire doesn't make a regulator that holds the output current at about 3.35 V constant for an input varying from 18 to 12 V (much less 14.4 V, which is 2.4 V per cell), it's because they don't want to, not because they can't. And if they don't make a temperature-compensated regulator to hold the LUMENS absolutely constant, assuming our first paragraph accurately describes the limiting factor, it's because they don't want to lower the initial output from 900 to 720 lumens (a difference which makes no practical effect in the real world, but only in marketing), which would allow a line as flat as a table until cell depletion.
Bottom line: the heat sinking is probably very poor, because the T sub J is rising 95 C above ambient, and sure as hell the flashlight casing isn't rising more than one quarter that much, because you couldn't hold it.