WEBVTT

1
00:00:00.000 --> 00:00:09.367
This overview explains how C23 tests the best-artifact task
on fresh generator families and separately recorded hosts.

2
00:00:09.367 --> 00:00:18.194
Fresh input families and a new host test different forms
of transfer, and neither implies a universal performance ranking.

3
00:00:18.194 --> 00:00:18.200
Fresh input families and a new host test different forms
of transfer, and neither implies a universal performance ranking.

4
00:00:18.500 --> 00:00:24.500
Imagine a bounded task with three valid artifacts, scored three, two, and one under a

5
00:00:24.500 --> 00:00:27.633
declared smaller-is-better toy ranking.

6
00:00:27.633 --> 00:00:31.267
The true best candidate in this bounded set is C.

7
00:00:31.267 --> 00:00:35.167
Exactness and rank are both visible.

8
00:00:35.167 --> 00:00:39.167
If the reference omits C, it names B as the winner.

9
00:00:39.167 --> 00:00:43.933
A method returning C may then appear to disagree with the reference even though it found

10
00:00:43.933 --> 00:00:46.367
the better permitted artifact.

11
00:00:46.367 --> 00:00:51.833
The problem lies in the reference's incomplete task coverage.

12
00:00:51.833 --> 00:00:56.333
Rebuild the reference to cover every candidate it claims to score.

13
00:00:56.333 --> 00:01:01.067
Keep the original defective freeze in the correction trail and bind the corrected result

14
00:01:01.067 --> 00:01:03.833
to the actual enumeration bounds.

15
00:01:03.833 --> 00:01:10.832
The example does not claim that a bounded enumeration covers an unlimited search space.

16
00:01:10.832 --> 00:01:10.833
The example does not claim that a bounded enumeration covers an unlimited search space.

17
00:01:10.833 --> 00:01:20.667
C23 keeps the seven methods and best-artifact task while
changing to previously unused Yosys generator-family functions.

18
00:01:20.667 --> 00:01:24.633
The first incomplete task freeze remains recorded;

19
00:01:24.633 --> 00:01:30.267
the corrected bounded reference must
cover every candidate it claims to score.

20
00:01:30.267 --> 00:01:36.900
On the retained Windows table, source-packed
ANF narrowly leads the screened paths.

21
00:01:36.900 --> 00:01:44.433
The later unchanged Linux table changes the ordering:
compiled screened CM leads in aggregate.

22
00:01:44.433 --> 00:01:49.700
Both host results are complete, and their narrow aggregate winners differ.

23
00:01:49.700 --> 00:01:56.933
Both host lanes retain the same canonical artifact
contract, even though their timing rankings differ.

24
00:01:56.933 --> 00:02:02.367
Source-family freshness and machine transfer are different checks.

25
00:02:02.367 --> 00:02:06.233
The retained report includes both Windows and Linux results,

26
00:02:06.233 --> 00:02:12.067
yet narrow headroom and individual
regressions keep production promotion false.

27
00:02:12.067 --> 00:02:19.333
A correct selected artifact and a new host do
not authorize training or a default change.

28
00:02:19.633 --> 00:02:23.867
A new source family and a new machine are independent changes.

29
00:02:23.867 --> 00:02:27.233
Place comparisons in a two-by-two grid so it is clear which

30
00:02:27.233 --> 00:02:29.700
combinations were actually tested.

31
00:02:29.700 --> 00:02:34.267
One filled cell does not establish the result in the others.

32
00:02:34.267 --> 00:02:38.267
Two implementations can prepare and screen data differently while owing

33
00:02:38.267 --> 00:02:40.800
the same canonical artifact.

34
00:02:40.800 --> 00:02:44.767
Compare both reconstruction and the complete ranking fields.

35
00:02:44.767 --> 00:02:50.533
Different internal timing orderings do not redefine the caller's contract.

36
00:02:50.533 --> 00:02:54.200
A narrow aggregate lead can change across hosts while every

37
00:02:54.200 --> 00:02:57.133
returned artifact remains exact.

38
00:02:57.133 --> 00:03:01.133
Preserve both host records and individual regressions.

39
00:03:01.133 --> 00:03:05.200
Neither a new host nor a correct artifact alone authorizes training

40
00:03:05.200 --> 00:03:09.167
changes or production promotion.

41
00:03:09.167 --> 00:03:09.168
Preserve each host's exactness and timing results.

42
00:03:09.168 --> 00:03:13.667
Preserve each host's exactness and timing results.

43
00:03:13.667 --> 00:03:19.986
Keep individual regressions and the
unchanged non-promotion decision visible.