WEBVTT

1
00:00:00.000 --> 00:00:02.533
This overview explains how dispatch,

2
00:00:02.533 --> 00:00:10.300
source transfer, cost-policy fitting, and compiled
policies interact in C17 through C20.

3
00:00:10.300 --> 00:00:18.659
An aggregate improvement can coexist with a per-case
failure, and a retrospective replay is not fresh confirmation.

4
00:00:18.659 --> 00:00:18.667
An aggregate improvement can coexist with a per-case
failure, and a retrospective replay is not fresh confirmation.

5
00:00:18.967 --> 00:00:23.300
A router that selects an exact method is itself a program.

6
00:00:23.300 --> 00:00:28.967
In this fictional example, feature extraction costs two units, dispatch costs one, and the

7
00:00:28.967 --> 00:00:31.800
chosen exact arm costs ten.

8
00:00:31.800 --> 00:00:36.100
The caller pays thirteen units for the complete route.

9
00:00:36.100 --> 00:00:41.067
If every leaf of the decision tree chooses that same arm, evaluating the features cannot

10
00:00:41.067 --> 00:00:42.900
change the decision.

11
00:00:42.900 --> 00:00:47.333
Replacing the tree with a constant selection removes unnecessary feature work while

12
00:00:47.333 --> 00:00:50.500
preserving the chosen method.

13
00:00:50.500 --> 00:00:53.667
Follow two inputs down different paths.

14
00:00:53.667 --> 00:00:57.600
They reach different leaves but select the same control.

15
00:00:57.600 --> 00:01:00.500
This makes the redundancy concrete.

16
00:01:00.500 --> 00:01:06.033
Folding the constant decision is valid only after confirming that every relevant path has

17
00:01:06.033 --> 00:01:11.433
the same result and required side effects are preserved.

18
00:01:11.433 --> 00:01:11.439
C17 adds a task dispatcher around exhaustive and screened exact arms.

19
00:01:11.439 --> 00:01:18.367
C17 adds a task dispatcher around exhaustive and screened exact arms.

20
00:01:18.367 --> 00:01:22.200
C18 changes the source cohort.

21
00:01:22.200 --> 00:01:29.167
C19 separates development, validation
and confirmation for fitting a cost policy.

22
00:01:29.167 --> 00:01:36.533
C20 compiles the frozen policy when its learned
tree reduces to a constant screened choice.

23
00:01:36.533 --> 00:01:45.067
In a conceptual task, a dispatcher examines features,
chooses screening and verifies the selected artifact.

24
00:01:45.067 --> 00:01:51.167
If it always chooses screening, the features
and branch can become pure overhead.

25
00:01:51.167 --> 00:01:59.233
Constant-folding that frozen choice removes the decision
path while preserving the selected exact arm.

26
00:01:59.233 --> 00:02:07.367
A separate fresh test is still needed to establish
performance outside the cohort that motivated the change.

27
00:02:07.367 --> 00:02:10.933
The retained sequence includes aggregate improvements,

28
00:02:10.933 --> 00:02:18.900
a tiny-case tail failure, independent-source
evidence and a retrospective same-machine replay.

29
00:02:18.900 --> 00:02:25.600
A retrospective replay is not a fresh
confirmation simply because the code is new.

30
00:02:25.600 --> 00:02:33.287
None permits dropping exact reconstruction
or changing a per-case gate after observing failures.

31
00:02:33.287 --> 00:02:33.300
None permits dropping exact reconstruction
or changing a per-case gate after observing failures.

32
00:02:33.567 --> 00:02:36.767
Development evidence can guide revisions.

33
00:02:36.767 --> 00:02:42.167
Before confirmation, freeze the rule, source identities, and acceptance gates so that the

34
00:02:42.167 --> 00:02:46.300
test has a stable object to evaluate.

35
00:02:46.300 --> 00:02:51.667
A fresh eligible cohort tests the frozen rule under the declared criteria.

36
00:02:51.667 --> 00:02:56.867
Aggregate performance and bad-case limits can answer different questions, so retain both

37
00:02:56.867 --> 00:02:59.867
when they are part of the gate.

38
00:02:59.867 --> 00:03:04.333
Replaying the same old cohort can verify reproducibility.

39
00:03:04.333 --> 00:03:07.800
It does not create fresh source-family evidence.

40
00:03:07.800 --> 00:03:13.533
A no-op or constant dispatcher can be safe and exact while remaining unpromoted under the

41
00:03:13.533 --> 00:03:18.000
actual frozen performance criteria.

42
00:03:18.000 --> 00:03:18.015
Keep reconstruction checks and frozen per-case gates intact.

43
00:03:18.015 --> 00:03:23.033
Keep reconstruction checks and frozen per-case gates intact.

44
00:03:23.033 --> 00:03:29.167
Compare a constant policy with invoking the same fixed arm directly.