> "frequency is measured from when the backup ends"

Many posters have made this statement; please state your sources as this
is demonstrably false.

 Policy Name:           DEMO-prod-std
 Include:   /
 Schedule:          full
   Frequency:       every 1 hour
   Daily Windows:
         Thursday   09:00:00  -->  Thursday   19:00:00
 
 Backup ID:         someserver_1147969653
 Policy:            DEMO-prod-std
 Sched Label:       full
 Backup Time:       05/18/2006 11:27:33
 Elapsed Time:      000:04:16
 
 Backup ID:         someserver_1147973257
 Policy:            DEMO-prod-std
 Sched Label:       full
 Backup Time:       05/18/2006 12:27:37
 Elapsed Time:      000:03:04

The reason schedules "slip" in NetBackup (pre-NetBackup 6) is the
scheduler wake-up interval of ten minutes (default) and scheduler
workload.  The time from the ideal frequency plus last start time (say,
24 hours and 18:01:33) until the next time the scheduler wakes up after
18:01:33 the next day is the earlies the job can be scheduled.  On
average, a schedule would slip five minutes per day (potentially missing
a window).  This (and the "I ran it as a test yesterday morning so it
didn't run last night" oversight) are why setting the frequency to the
window size became popular.  support.veritas.com/docs/209227

This was addressed in two ways that I know of:
1.  logic was added (4.5? 5.x?) to check the frequency for exact
multiples of a day, and start those backups at the beginning of the
window each time--it's "interval mode" which another poster spoke to
this earlier.  (I believe this is covered in a technote but I don't have
it.)
2.  NetBackup 6 sets an alarm in advance for when backups are due rather
than check for this every ten minutes, IIRC from training materials.