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
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
|
# Oral-A recovery v2: BCI-v2-authorized standard-depth scaling
## Status and provenance boundary
This is a new confirmation branch opened only by the complete calibrated
oral-B-v2 pass. It does not reopen or relabel the failed learned-vectorizer
Oral-A-v1--v4 branches, and it does not alter the closed legacy recovery whose
prerequisite was the failed original oral-B R2 gate.
The experiment and every numerical threshold below are copied verbatim from
`ORAL_A_RECOVERY.md`, which was committed before any standard-depth endpoint
was generated. The only protocol change is the prerequisite: the new branch
requires both the complete D4 accept pass and
`oral_b_v2_calibrated_recovery_confirmation_v1`. No recovery-v2 test endpoint
may be generated until this document, its runner, analyzer, and smoke audit
are committed together in a clean tracked revision.
This separation matters for interpretation. The calibrated BCI-v2 experiment
establishes that role-vectorized somato-dendritic innovation can carry local
TD/outcome surprise without Kolen--Pollack reciprocity. The panel below then
tests the complementary engineering claim that the already established
dynamic-innovation/reciprocal substrate retains performance and alignment as
a canonical image network becomes deeper. Neither result alone is presented
as proving the other.
## Fixed panel and reuse boundary
Use CIFAR-10 ResNet-20/32/56 with width 16, BatchNorm, option-A shortcuts,
batch size 128, all 50,000 training examples, 200 epochs, SGD momentum 0.9,
weight decay `1e-4`, base learning rate 0.1 with 10-fold drops at epochs 100
and 150, no warmup, and model/data-loader seeds 10--14. Test is evaluated
exactly once at the endpoint; validation and intermediate test evaluations are
zero.
The matched methods are:
1. exact BP on the identical forward architecture;
2. tuned DFA at its already selected hidden rate 0.03 and output rate 0.1;
3. clean reciprocal Kolen--Pollack at hidden/output rate 0.1; and
4. dynamic innovation on the same reciprocal substrate, with traffic ratio 4,
traffic seed `5000 + model_seed`, 64 instruction-off calibration examples,
one closed-form neutral projection, and no later predictor update.
The already frozen D4 clean-KP and dynamic records at ResNet-20/seeds 10--14
are reused verbatim rather than rerun. Recovery v2 therefore adds exactly 50
records: BP and DFA at depth 20, and all four methods at depths 32 and 56.
Together with D4 this forms an exact 60-cell panel. All new records must share
one clean tracked source revision; all D4 records retain their separate frozen
source revision. No completed record may be deleted, replaced, or pooled with
a pilot.
Fixed HFA remains closed after its preregistered short gate failed. FF, PEPITA,
canonical EP, BurstCCN, and Dual Prop retain their method-native audits and are
reported as separate baselines rather than silently moved onto this matched
standard-ResNet grid.
## Frozen mechanism and cost invariants
Every dynamic record must retain all D4 invariants: finite 200-epoch
trajectory; mean early-third credit alignment measured only diagnostically;
final feedback/forward cosine at least 0.98 and epoch-151--200 mean at least
0.97; instruction-to-teaching RMS error at most `1e-4`; maximum post-projection
traffic ratio and absolute soma slope at most `1e-5`; zero instruction leakage
into neutral observations; exactly 64 predictor-update examples; exactly one
neutral observation per ordinary example; and zero task-loss queries.
At every depth, dynamic and clean KP must each use at most `1.34x` the matched
BP MAC estimate. Dynamic elementwise projection work is reported separately
and must be positive. Peak allocated memory may not exceed 8 GiB on the
authorized GTX 1080 devices. Wall time is reported but is never substituted
for MACs, queries, or memory. Every record must identify an NVIDIA GTX 1080
and visibility restricted to authorized timan107 physical GPU 5 or 7.
## Frozen oral-A gate
The five model/data seeds are the independent paired units. One-sided 95%
Student-t bounds use four degrees of freedom. Recovery v2 passes only if every
record and every audited value is finite and all of the following hold:
1. Mean BP accuracy is at least 90% at every depth and every dynamic endpoint
is at least 87%.
2. At each depth, mean dynamic accuracy is within 2 points of matched BP and
the one-sided 95% upper bound on the paired deficit is at most 3 points.
3. At each depth, mean dynamic accuracy is within 1.5 points of clean KP and
the upper bound on that paired deficit is at most 2.5 points.
4. At depth 56, dynamic exceeds tuned DFA by at least 2 points on average and
the one-sided lower bound on the paired advantage is at least 1 point.
5. Dynamic gains at least 0.5 points from depth 20 to 56 on average, its
one-sided lower bound is nonnegative, and at least four of five paired
seeds improve. Its mean depth gain must also be no more than 1 point below
BP's paired depth gain.
6. Mean dynamic early-third alignment is at least 0.85 at every depth; every
depth-56 seed is at least 0.80; and the depth-56 mean retains at least 90%
of the depth-20 mean.
7. Every mechanism, query, MAC, elementwise-work, memory, source, split, and
evaluation invariant above passes without exception.
The executable runner and immutable complete-grid analyzer are
`experiments/oral_a_dynamic_scaling_v2.py` and
`experiments/analyze_oral_a_dynamic_scaling_v2.py`. A complete pass raises the
formal milestone from 8 to 9 and establishes the standard-depth oral-A claim.
Any failure is retained, leaves the score at 8, and closes this branch without
depth, seed, optimizer, or threshold repair.
|
