WEBVTT

1
00:00:00.000 --> 00:00:06.433
This overview explains how a corrected comparison
changes the claims supported by historical evidence.

2
00:00:06.433 --> 00:00:13.097
A later improvement on a different task cannot
retroactively validate an earlier mismatched comparison.

3
00:00:13.097 --> 00:00:13.100
A later improvement on a different task cannot
retroactively validate an earlier mismatched comparison.

4
00:00:13.400 --> 00:00:18.600
Imagine a candidate taking six units against a weak ten-unit control.

5
00:00:18.600 --> 00:00:21.867
The observed comparison favors the candidate.

6
00:00:21.867 --> 00:00:26.800
These fictional values illustrate how a correction changes a claim; they do not replace

7
00:00:26.800 --> 00:00:30.100
retained research measurements.

8
00:00:30.100 --> 00:00:35.467
A stronger control completes the same required task in five units.

9
00:00:35.467 --> 00:00:41.700
The candidate still produces the same exact answer, but its earlier advantage disappears.

10
00:00:41.700 --> 00:00:48.300
The relevant performance claim must change because its comparison premise changed.

11
00:00:48.300 --> 00:00:52.567
Keep the mathematical identity, the implementation description, and the

12
00:00:52.567 --> 00:00:55.667
speed claim as separate cards.

13
00:00:55.667 --> 00:00:59.933
A baseline correction can invalidate the speed comparison without

14
00:00:59.933 --> 00:01:03.533
disproving a separately checked identity.

15
00:01:03.533 --> 00:01:09.663
It also does not establish that the implementation is otherwise flawless.

16
00:01:09.663 --> 00:01:09.667
It also does not establish that the implementation is otherwise flawless.

17
00:01:09.667 --> 00:01:14.200
Keep the old evidence and mark the wording it no longer supports.

18
00:01:14.200 --> 00:01:19.533
A correction creates a traceable successor;
it does not make history disappear.

19
00:01:19.533 --> 00:01:24.633
The audit replaced a weak generic
comparator story with sharing-aware CSE-flat,

20
00:01:24.633 --> 00:01:30.233
checked exact outputs and execution
order, and separated kernel from wrapper.

21
00:01:30.233 --> 00:01:36.000
The corrected B2/B4 compiled-kernel advantage survived in its stated scope.

22
00:01:36.000 --> 00:01:40.667
EPFL remained near parity against CSE-flat.

23
00:01:40.667 --> 00:01:43.933
The wrapper comparison remained slower.

24
00:01:43.933 --> 00:01:52.800
Three conclusions remain separate: a scoped B2/B4 kernel
advantage, EPFL near parity, and a slower public wrapper.

25
00:01:52.800 --> 00:02:00.033
The withdrawn wrapper crossover and dramatic
mismatched-task SymPy claim must stay withdrawn.

26
00:02:00.033 --> 00:02:09.033
Later counting or packed-mask improvements answer their own questions;
they do not retroactively validate an unsupported historical headline.

27
00:02:09.033 --> 00:02:14.800
A claim registry needs both supersession
edges and the retained source identities.

28
00:02:15.100 --> 00:02:19.767
Suppose the corrected result concerned a fresh full-vector request.

29
00:02:19.767 --> 00:02:23.567
Its conclusion belongs to that contract, including the preparation and

30
00:02:23.567 --> 00:02:27.033
output work the caller required.

31
00:02:27.033 --> 00:02:32.700
A later scalar evaluation experiment can be valuable, but it asks for one answer rather

32
00:02:32.700 --> 00:02:34.733
than all outputs.

33
00:02:34.733 --> 00:02:41.300
A win on that new task cannot retroactively restore the old full-vector claim.

34
00:02:41.300 --> 00:02:45.900
Preserve the old record, attach the correction, and state the new task

35
00:02:45.900 --> 00:02:48.167
as a separate experiment.

36
00:02:48.167 --> 00:02:51.333
This gives the reader a trail they can follow without forcing them

37
00:02:51.333 --> 00:02:55.429
to infer which conclusions survived.

38
00:02:55.429 --> 00:02:55.433
to infer which conclusions survived.

39
00:02:55.433 --> 00:02:59.567
Preserve the old evidence and its supersession record.

40
00:02:59.567 --> 00:03:05.986
Support a new claim with a new matched comparison
rather than reviving a withdrawn headline.