I've been playing with accumulating the LSB from an unconnected ADC pin ... This seems to work OK but still not good enough
I had fun running the experiments but the results were not good. A four inch insulated wire gave the best results. Longer or shorter wires and the values turned to garbage (randomly speaking, of course). And [u]anything[/u] near the wire had an effect. The method is not very reliable.
Another option I'm thinking about is to use a semi-random byte generated as I just mentioned to load the RC oscillator adjust register, then count the time it takes to do an EEPROM write and use one byte from the result. Once again do this four times to get 32 bits.
I had looked at that as well. It appears that EEPROM write times are driven entirely by the processor clock. In other words, an EEPROM write appears to always take X processor clock ticks. But, I didn't pursue it long enough to determine that for certain.
Maybe the temp sensor can be thrown into the mix here as well
In general, the internal temperature sensors seem to be too stable. Especially if the processor is flowing any current (processor is above ambient temperature).
For the ATtiny85s I have, the internal temperature sensor only produces even values. The granularity is too course.
I guess this is similar to CBs watchdog arrangement
That didn't go too well either. For some processors, the results were good. For SMD processors, the results were very very bad.
I even tried this technique...
...without much success. I believe the sample-and-hold capacitor in the AVR processor has too little capacitance for the technique to work. I can't remember the details but I just was not getting good results.