WEBVTT

1
00:00:00.000 --> 00:00:08.633
This overview explains how the retained B2/B4 cloud runs
extend the evidence across separately recorded environments.

2
00:00:08.633 --> 00:00:17.733
Observed ranges across machines do not automatically form
a confidence interval or replace the original aggregation design.

3
00:00:18.033 --> 00:00:21.633
Suppose one formula is timed one hundred times.

4
00:00:21.633 --> 00:00:26.367
Those rows can describe timing variability for that formula and setup.

5
00:00:26.367 --> 00:00:31.467
They do not create one hundred independently selected formulas or establish coverage of an

6
00:00:31.467 --> 00:00:34.433
entire input family.

7
00:00:34.433 --> 00:00:39.900
Now run the same frozen job on three hosts and imagine ratios zero point eight, one, and

8
00:00:39.900 --> 00:00:41.500
one point two.

9
00:00:41.500 --> 00:00:44.867
The host transfer changes the observed result.

10
00:00:44.867 --> 00:00:49.433
These three points are illustrative host results, not automatically an uncertainty

11
00:00:49.433 --> 00:00:53.833
interval with a declared statistical interpretation.

12
00:00:53.833 --> 00:00:58.767
Within-host repetitions and across-host changes address different questions.

13
00:00:58.767 --> 00:01:03.467
A statistical analysis must specify what its units represent.

14
00:01:03.467 --> 00:01:10.333
Collapsing the two axes into one pile of rows can make evidence look broader than it is.

15
00:01:10.333 --> 00:01:10.336
Replication asks whether an observation persists
under separately recorded environments.

16
00:01:10.636 --> 00:01:18.023
Formula balancing gives each formula equal influence despite different numbers of repeated timing rows.

17
00:01:18.023 --> 00:01:22.589
Row weighting gives every timing row equal influence.

18
00:01:22.589 --> 00:01:35.384
Consider a teaching example: formula X has ratio zero point five measured three times, and formula Y has ratio one point five measured once.

19
00:01:35.684 --> 00:01:41.864
For an arithmetic mean in this teaching example, the four rows average to zero point seven five.

20
00:01:41.864 --> 00:01:45.629
Giving each of the two formulas equal weight gives one.

21
00:01:45.629 --> 00:01:50.682
Repeating X changes the row-weighted result but not the equal-formula result.

22
00:01:50.682 --> 00:01:56.859
These teaching arithmetic means are not a reconstruction of the historical estimator.

23
00:01:57.159 --> 00:02:04.268
Formula-cluster uncertainty keeps each formula and its repeated rows together when resampling.

24
00:02:04.268 --> 00:02:11.110
It describes formula variation within the recorded run, not a guarantee across machines.

25
00:02:11.110 --> 00:02:20.898
The companion links the exact retained aggregation and ratio protocol; use that protocol for the historical numbers.

26
00:02:20.898 --> 00:02:27.495
Replication asks whether an observation persists
under separately recorded environments.

27
00:02:27.495 --> 00:02:34.229
These pod summaries use row weighting, whereas
the local headline uses formula balancing.

28
00:02:34.229 --> 00:02:37.662
The aggregation procedures differ.

29
00:02:37.662 --> 00:02:41.529
The retained three-pod bare CM-over-CSE-flat

30
00:02:41.529 --> 00:02:49.262
time ratios span 0.9026 to 0.9126 overall.

31
00:02:49.262 --> 00:02:57.195
At support 16 they span 0.9746 to 0.9766.

32
00:02:57.195 --> 00:03:02.162
Each value belongs to its own recorded machine and run.

33
00:03:02.162 --> 00:03:10.329
Each of the three estimates retains its own machine and run
identity, under the compiled-kernel timing boundary.

34
00:03:10.329 --> 00:03:17.795
These are descriptive replication values, not endpoints
of a predeclared pooled confidence interval.

35
00:03:17.795 --> 00:03:21.829
They support the scoped direction on those runs.

36
00:03:21.829 --> 00:03:28.929
They neither erase wrapper overhead nor establish
that the current implementation will repeat them.

37
00:03:28.929 --> 00:03:33.986
These are historical retained runs, rather than a new benchmark.

38
00:03:33.986 --> 00:03:33.995
These are historical retained runs, rather than a new benchmark.

39
00:03:34.295 --> 00:03:39.729
A replication should keep the relevant job identity explicit: source, inputs,

40
00:03:39.729 --> 00:03:44.795
dependencies, wrapper, output contract, and timed boundary.

41
00:03:44.795 --> 00:03:50.129
A change in one of these can turn replication into a different experiment.

42
00:03:50.129 --> 00:03:53.962
Check that both hosts produce the required exact outputs before

43
00:03:53.962 --> 00:03:56.529
interpreting their timing ratios.

44
00:03:56.529 --> 00:04:00.995
A faster run that omitted output work or changed the assignment universe is not a

45
00:04:00.995 --> 00:04:05.162
successful replication of the original task.

46
00:04:05.162 --> 00:04:09.629
A matching replication strengthens evidence about the declared task and reveals

47
00:04:09.629 --> 00:04:12.729
sensitivity to the tested environments.

48
00:04:12.729 --> 00:04:17.562
It does not by itself establish a result for unrelated formula families or

49
00:04:17.562 --> 00:04:21.595
for a different requested artifact.

50
00:04:21.595 --> 00:04:21.611
Retain each run's scope and weighting method.

51
00:04:21.611 --> 00:04:25.495
Retain each run's scope and weighting method.

52
00:04:25.495 --> 00:04:31.687
Separate descriptive replication ranges from inferential uncertainty.