WEBVTT

1
00:00:00.000 --> 00:00:08.033
This overview explains which scoped performance findings
survived the corrected comparison with sharing-aware CSE-flat.

2
00:00:08.033 --> 00:00:15.667
Keeping kernel and wrapper results separate prevents a narrow
retained advantage from becoming an unsupported universal claim.

3
00:00:15.967 --> 00:00:19.367
A comparison row is more than two elapsed times.

4
00:00:19.367 --> 00:00:25.033
It identifies the input, the implementation paths, the required output, and the boundary

5
00:00:25.033 --> 00:00:26.600
that was timed.

6
00:00:26.600 --> 00:00:32.733
Keep those fields attached to the corrected B2 and B4 result when interpreting it.

7
00:00:32.733 --> 00:00:36.633
A full vector at support sixteen contains sixty-five thousand

8
00:00:36.633 --> 00:00:40.067
five hundred thirty-six Boolean outputs.

9
00:00:40.067 --> 00:00:44.067
That tells us the assignment universe, not the complexity of every

10
00:00:44.067 --> 00:00:46.567
expression over those inputs.

11
00:00:46.567 --> 00:00:52.000
Two width-sixteen formulas can have very different structure.

12
00:00:52.000 --> 00:00:58.467
The measured comparison remains tied to its declared family, wrapper, and output task.

13
00:00:58.467 --> 00:01:04.433
It should not be turned into a universal width-sixteen statement or a point-query result.

14
00:01:04.433 --> 00:01:11.161
The corrected row is useful precisely when a new workload matches its scope.

15
00:01:11.461 --> 00:01:17.896
Formula balancing gives each formula equal influence despite different numbers of repeated timing rows.

16
00:01:17.896 --> 00:01:21.568
Row weighting gives every timing row equal influence.

17
00:01:21.568 --> 00:01:32.540
Consider a teaching example: formula X has ratio zero point five measured three times, and formula Y has ratio one point five measured once.

18
00:01:32.840 --> 00:01:38.660
For an arithmetic mean in this teaching example, the four rows average to zero point seven five.

19
00:01:38.660 --> 00:01:42.425
Giving each of the two formulas equal weight gives one.

20
00:01:42.425 --> 00:01:47.595
Repeating X changes the row-weighted result but not the equal-formula result.

21
00:01:47.595 --> 00:01:53.981
These teaching arithmetic means are not a reconstruction of the historical estimator.

22
00:01:54.281 --> 00:02:00.414
Formula-cluster uncertainty keeps each formula and its repeated rows together when resampling.

23
00:02:00.414 --> 00:02:06.141
It describes formula variation within the recorded run, not a guarantee across machines.

24
00:02:06.141 --> 00:02:14.629
The companion links the exact retained aggregation and ratio protocol; use that protocol for the historical numbers.

25
00:02:14.629 --> 00:02:14.635
The corrected row is useful precisely when a new workload matches its scope.

26
00:02:14.635 --> 00:02:23.401
The accepted B2/B4 V3 result compares compiled
CM and sharing-aware CSE-flat after compilation.

27
00:02:23.401 --> 00:02:29.801
The formula-balanced CM-over-CSE-flat time ratio is 0.8906,

28
00:02:29.801 --> 00:02:38.535
with a formula-cluster 95% interval from 0.8741 to 0.9073.

29
00:02:38.535 --> 00:02:43.868
At live support 16, that ratio narrows to 0.9612,

30
00:02:43.868 --> 00:02:50.035
with interval 0.9290 to 0.9942.

31
00:02:50.035 --> 00:02:53.435
These are two labelled cohort estimates.

32
00:02:53.435 --> 00:02:57.635
Neither is a speedup that applies to every formula.

33
00:02:57.635 --> 00:03:02.568
In the same correction history, the public wrapper is slower than CSE-flat;

34
00:03:02.568 --> 00:03:07.368
the bare evaluator result does not pay for all request stages.

35
00:03:07.368 --> 00:03:14.635
These are historical retained estimates for their recorded
workload, host, comparator and timing boundary.

36
00:03:14.635 --> 00:03:18.501
They are not a benchmark of today's modified code.

37
00:03:18.501 --> 00:03:25.268
The cluster intervals describe within-run
formula variation, not an all-machine guarantee.

38
00:03:25.268 --> 00:03:32.865
Current implementation changes require a separately
authorized matched replay before new performance claims.

39
00:03:32.865 --> 00:03:32.868
Current implementation changes require a separately
authorized matched replay before new performance claims.

40
00:03:33.168 --> 00:03:37.668
Consider fictional costs that demonstrate the logic of a correction.

41
00:03:37.668 --> 00:03:44.635
A six-unit candidate compared with a ten-unit control has a time ratio of zero point six.

42
00:03:44.635 --> 00:03:49.768
These teaching numbers are not the B2 or B4 measurements.

43
00:03:49.768 --> 00:03:55.401
If a corrected wrapper reduces the control to five units while preserving the same output,

44
00:03:55.401 --> 00:03:59.135
the candidate's new ratio is one point two.

45
00:03:59.135 --> 00:04:02.435
The candidate did not become logically wrong.

46
00:04:02.435 --> 00:04:08.168
The old performance comparison stopped representing the stronger baseline.

47
00:04:08.168 --> 00:04:12.768
Preserve the historical record and clearly identify the correction.

48
00:04:12.768 --> 00:04:17.101
Claims that depended on the weaker comparison must be revised.

49
00:04:17.101 --> 00:04:22.168
Other claims require their own evidence; they neither disappear automatically nor gain

50
00:04:22.168 --> 00:04:25.501
support from the correction.

51
00:04:25.501 --> 00:04:25.516
Preserve the retained result with its workload, host, and timing boundary.

52
00:04:25.516 --> 00:04:31.101
Preserve the retained result with its workload, host, and timing boundary.

53
00:04:31.101 --> 00:04:36.468
New implementation claims require a new matched comparison.