WEBVTT

1
00:00:00.000 --> 00:00:07.733
This overview explains how an advice policy can choose
the exact fallback and still spend time deciding not to rewrite.

2
00:00:07.733 --> 00:00:14.833
Zero accepted rewrites does not mean zero advice
cost when features and decision logic still execute.

3
00:00:15.133 --> 00:00:20.433
Start with a direct exact control costing ten illustrative units.

4
00:00:20.433 --> 00:00:24.833
It produces the required answer under the declared contract.

5
00:00:24.833 --> 00:00:29.533
This is the baseline request the gated route must match.

6
00:00:29.533 --> 00:00:35.933
The gate spends two units deciding to abstain, then runs the same ten-unit control.

7
00:00:35.933 --> 00:00:39.567
The answer remains exact and no rewrite was applied, but the

8
00:00:39.567 --> 00:00:42.633
complete request costs twelve.

9
00:00:42.633 --> 00:00:49.100
Abstention avoids an unsuitable action; it does not erase the gate's own cost.

10
00:00:49.100 --> 00:00:53.567
If all three requests take this fallback path, the direct total is thirty

11
00:00:53.567 --> 00:00:56.267
and the gated total thirty-six.

12
00:00:56.267 --> 00:01:01.600
Every answer can be identical while the gated route is slower on every request.

13
00:01:01.600 --> 00:01:07.533
These are teaching units, not reconstructed historical measurements.

14
00:01:07.567 --> 00:01:15.367
An abstention policy can preserve the exact fallback while still
spending time obtaining features and deciding to abstain.

15
00:01:15.367 --> 00:01:19.967
Safe advice and profitable advice are different tests.

16
00:01:19.967 --> 00:01:24.767
Suppose a toy request takes 10 cost units without advice.

17
00:01:24.767 --> 00:01:32.900
A gate spends 2 units to decide that no rewrite is worth
applying, then runs the unchanged 10-unit control.

18
00:01:32.900 --> 00:01:40.233
The final output is correct and the decision avoids
a risky transformation, but the request costs 12.

19
00:01:40.233 --> 00:01:48.000
An out-of-range input should also abstain, with its reason
recorded separately from insufficient predicted gain.

20
00:01:48.000 --> 00:01:56.133
The retained frozen D9 policy abstained on all
evaluation workloads and preserved exactness.

21
00:01:56.133 --> 00:02:02.100
Its charged gate remained slower, so no rewrite policy was promoted.

22
00:02:02.100 --> 00:02:07.267
Do not present zero accepted rewrites as a zero-cost path.

23
00:02:07.567 --> 00:02:12.533
Suppose a toy gate is validated only for a declared range of inputs.

24
00:02:12.533 --> 00:02:17.700
An outside-range input should not be treated as if the same evidence applies.

25
00:02:17.700 --> 00:02:24.900
This is a schematic eligibility rule, separate from any specific retained D9 guard.

26
00:02:24.900 --> 00:02:29.967
The gate abstains from rewriting and uses an exact eligible fallback.

27
00:02:29.967 --> 00:02:34.833
This can be the correct safe behavior for the request, even if it is slower than calling

28
00:02:34.833 --> 00:02:37.800
that fallback directly.

29
00:02:37.800 --> 00:02:43.367
Report how often the gate acts or abstains, whether every output stays exact, and how much

30
00:02:43.367 --> 00:02:45.533
the complete route costs.

31
00:02:45.533 --> 00:02:50.100
Safe abstention and profitable routing require different evidence.

32
00:02:50.100 --> 00:02:55.800
Keep the retained all-abstained result and its slowdown visible.

33
00:02:55.800 --> 00:02:55.815
Charge the decision to abstain.

34
00:02:55.815 --> 00:02:59.000
Charge the decision to abstain.

35
00:02:59.000 --> 00:03:05.263
Compare the full advised path with calling the fixed exact arm directly.