In the case where the logging interval is faster than the reporting interval, ie with a reporting interval of 15 minutes and a logging interval of 5 minutes, I'd expect the value at 2:00 to be the actual sample: 13.
The actual values logged could be instantaneous or the average of all the samples since the last logged value. For instance I have the modbus data source poll time set as low as I can without the polls starting to overlap, about 10s. Most variables I simply log the instantaneous value every 15 minutes, but for particularly noisy ones I log the average of the 6 samples/minute * 15 minutes = 90 samples. The result is there is some time delay in averaged samples, I think that it is acceptable, but I suppose it depends on the characteristic of the process.
For the second case, where the report interval does not coincide with the logging interval, I would linearly interpolate between each point. Thus to report on the value at 2PM a value after or at 2PM would have to be logged.
Slightly unrelated, but I've been meaning to look into how RRD (http://oss.oetiker.ch/rrdtool/) stores data as I believe it sets the bar for high performance logging and archival of time series data. There is a java port as well I believe.