1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
|
# Same-path contrastive-bias confirmation
## Status and purpose
This is a post-screen confirmation frozen after the B1 results were known. It
is not the untouched B2 in `CONTRASTIVE_BIAS.md`. B1 showed the intended raw
failure and innovation recovery at every activity-bias ratio, but its complete
gate failed because ordinary clean DP ended at 69.66% instead of 70% and a
common-bias run differed from it by 2.28 accuracy points despite exactly zero
teaching-difference error.
This confirmation removes that control mismatch. All four conditions use the
same explicit-difference training path. All four conditions for one seed run
sequentially on the same physical GPU.
## Frozen panel
The implementation, upstream revision, model, data split, batch order, and
local bias equations are unchanged from B1. The five new model/minibatch seeds
are `1989, 1990, 1991, 1992, 1993`. Every run uses author Dual Prop miniCNN on
CIFAR-10, `alpha=0`, `beta=0.1`, `fwK`, 16 inference passes, batch size 100,
and the complete 130-epoch author schedule: peak learning rate 0.025, warm-up
learning rate 0.001, 10 warm-up epochs, decay epoch 120, and final learning
rate `2e-6`. Test is not evaluated.
For every seed, the four ratio-4 conditions are:
1. `same_path_clean`: common activity bias with the raw rule. Its compartment
difference is exactly zero, so this is clean DP through the same code path.
2. `raw`: differential activity bias with the raw rule.
3. `innovation`: differential activity bias with the local neutral affine
subtraction.
4. `oracle`: differential activity bias with exact generated-bias subtraction.
The ratio is calibrated once at initialization and frozen. A raw run that
becomes nonfinite is retained. Clean, innovation, and oracle must complete all
130 epochs. Test remains untouched until the full validation panel and audit
exist.
## Confirmation gate
The panel passes only if all source, registry, hardware, split, and test
isolation checks pass and:
- same-path clean, innovation, and oracle are finite for 130 epochs in all five
seeds, and mean same-path-clean final validation accuracy is at least 80%;
- raw is nonfinite or at least 20 accuracy points below same-path clean in
every seed;
- innovation is above raw in every seed and the one-sided 95% lower bound on
the paired gain is above 20 points;
- the one-sided 95% upper bound on the paired clean-minus-innovation deficit is
below 2 points;
- the two-sided 95% bound on the absolute mean innovation-minus-oracle gap is
below 2 points;
- innovation's post-subtraction bias RMS is at most `1e-3` of raw bias in every
run, and every predictor reports zero task-instruction observations;
- every seed's four conditions record the same physical GPU UUID.
Failure is retained and narrows the claim. Passing permits a separately
frozen, once-only test evaluation; it does not retroactively make B1 pass.
|
