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
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
|
# Dynamic neutral-projection development protocol
## Claim boundary
This is a new post-S0 development branch. It does not reinterpret or rescue
MT-1: MT-1 remains failed, MT-2/MT-3 remain sealed, and the reviewer score
remains 5/10. S0 ruled out extending a fixed one-sided predictor margin. The
new question is whether a fast *local and instruction-off* projection can keep
the multiplicative neutral-residual mode inside its two-sided stability window
as somatic statistics change.
The slow predictor is still fitted exactly once on the original 64-example
neutral calibration prefix and is frozen during task learning. Before each
task update, every cell receives a paired instruction-off observation of its
current soma `h` and ordinary apical traffic `a0`. From the neutral residual
```text
e0 = a0 - (P h + b),
```
it forms the current-batch local affine projection
```text
q = Cov(e0, h) / Var(h),
e_stable = (e0 - mean(e0)) - q (h - mean(h)).
```
The plasticity signal is `r = s + e_stable`. The coefficient fit sees zero
task-instruction observations; it reads no label, loss, forward weight,
feedback weight, or downstream state, and it does not modify the slow
predictor. For the diagonal affine traffic used by frozen MT-1, the controller
nulls the current empirical residual-coupling coefficient. With weight decay,
the scalar homogeneous mode is therefore placed near `k=-lambda`, inside both
Jury boundaries, instead of at a fixed negative coefficient whose product with
an evolving covariance can cross the lower boundary.
This controller requires one paired neutral apical microphase per task batch.
Its elementwise work and observation count must be charged separately in any
later endpoint comparison. It uses zero task-loss perturbation queries and no
reverse-mode differentiation.
## D0: mechanics (completed before task-data evaluation)
The implementation must prove on synthetic float64 tensors that:
- an inaccurate frozen slow predictor leaves a nonzero affine neutral mode;
- fast projection reduces the post-projection neutral/traffic RMS ratio and
residual-soma slope below `1e-14` for diagonal affine traffic;
- projected innovation equals clean instruction below `1e-14`;
- the projection consumes zero instruction observations and leaves every slow
predictor parameter bitwise unchanged;
- forward and reciprocal correlations remain independently computed; and
- the extra elementwise work is exposed explicitly.
These checks passed at clean revision `eeb50b0`.
## D1: frozen training-prefix gate
There is exactly one candidate and no hyperparameter grid. Keep the MT-1
ResNet-20 architecture, CIFAR-10 45k training split, seed-0 initialization,
four-to-one traffic intervention, augmentation, batch size 128, learning rate
0.1, momentum 0.9, decay `1e-4`, and reciprocal KP path unchanged. Use a
zero-margin closed-form slow predictor, freeze it, and apply the fast neutral
projection on every one of the first 352 shuffled task minibatches. Restore
the task-loader state after calibration. Do not evaluate validation or test
examples.
The candidate passes only if all conditions below hold:
- every recorded loss, signal diagnostic, parameter, optimizer state,
BatchNorm statistic, and projection diagnostic remains finite for 352 steps;
- maximum task minibatch loss is at most 10 and final-32 mean loss is at most
2.5;
- the used-signal/instruction RMS ratio stays within `1e-4` of one;
- the maximum post-projection neutral/traffic RMS ratio is at most `1e-5`;
- the maximum absolute post-projection residual-soma slope is at most `1e-5`;
- every projection uses the current task batch's 72--128 local neutral
observations and exactly zero instruction observations;
- maximum absolute forward and reciprocal-feedback weight is at most 10;
- maximum absolute forward and reciprocal momentum is at most 50;
- the 64-observation slow fit has zero stability margin, the slow predictor is
frozen during task training, initial traffic calibration is within `1e-5`
of ratio 4, and task-loader state restoration is exact; and
- validation and test evaluation counts are both zero.
A failure closes this paired-neutral projection branch without changing a
threshold or adding a gain/clipping hyperparameter. A pass supplies no task
accuracy claim and cannot change the reviewer score. It only authorizes a new,
separately committed validation protocol with an explicit cost boundary.
## Audited D1 outcome
D1 passes all frozen checks at clean revision `8b28522`. All 352 updates and
states remain finite; maximum minibatch loss is `2.9281` and the final-32 mean
is `1.6293`. The uncontrolled neutral residual reaches `0.005833` of traffic
RMS and the fast correction slope reaches `0.02968`, so the controller is not
a numerical no-op. Projection reduces the worst neutral/traffic RMS ratio to
`3.03e-8` and the worst absolute residual-soma slope to `1.52e-8`. Maximum
forward/feedback weights are `1.181/0.833` and maximum momenta are
`0.600/0.364`. No held-out example was evaluated and the score remains 5/10.
## D2: frozen short validation endpoint
D1 opens exactly one 20-epoch seed-0 validation run. Use the same ResNet-20,
45k/5k CIFAR-10 split, augmentation, batch size 128, constant learning rate
0.1 (the inherited KP short schedule), momentum 0.9, decay `1e-4`, four-to-one
traffic, reciprocal KP path, zero-margin 64-example closed-form slow fit, and
fast neutral projection on every task minibatch. Evaluate the validation set
once after training, probe alignment on 32 training examples only after the
endpoint is fixed, and never evaluate test.
D2 passes only if:
- the final record, all 20 epoch records, feedback trajectories, signal
trajectories, projection certificates, and diagnostics are finite;
- validation accuracy is at least 80%, within three points of the frozen clean
KP short endpoint (`82.66%`), at least 70 points above each failed MT-1 raw
and norm-matched endpoint, and final loss is finite;
- early-third used-signal alignment is at least 0.70, final feedback/forward
cosine is at least 0.80, and the epoch-11--20 mean feedback cosine is at
least 0.70;
- every epoch keeps used/instruction RMS within `1e-4` of one, maximum
post-projection neutral/traffic RMS at most `1e-5`, maximum absolute
post-projection soma slope at most `1e-5`, and zero task-instruction
observations in the fast fit;
- exactly one closed-form 64-example slow fit occurs, task predictor cadence is
zero, neutral-projection observation count equals ordinary training examples,
initial traffic ratio error is at most `1e-5`, and task-loss queries are zero;
- total MACs are at most `1.34x` the frozen 20-epoch BP record, all additional
elementwise work is reported, and peak allocated memory is at most 2.5 GiB;
and
- there is exactly one validation evaluation and zero test evaluations.
D2 failure closes the branch. A pass remains a one-seed short validation
result and leaves the paper score at 5/10, but opens the already specified D3
full validation endpoint below.
## D3: conditionally frozen full validation endpoint
If and only if D2 passes, run one seed-0 200-epoch validation record with the
same method and data boundary. The only schedule change is the inherited KP-2
step schedule: LR 0.1, drops by 0.1 at epochs 100 and 150, no warmup. Evaluate
validation exactly once at the end and test never. D3 passes only if every
D2 mechanics/cost/query certificate remains satisfied (using the corresponding
200-epoch BP cost), validation accuracy is at least 89%, lies within 1.5 points
of BP (`91.62%`) and clean KP (`91.26%`), final early alignment is at least
0.90, final feedback cosine is at least 0.98, the final-50-epoch mean feedback
cosine is at least 0.97, and all trajectories remain finite.
A D3 pass would establish a frozen full standard-ResNet endpoint and make a
5-to-6 reviewer-score update eligible, subject to an adversarial audit. It
would then open a separately frozen independent multi-seed test confirmation;
no confirmation seed or test endpoint may be touched before that protocol is
committed. A D3 failure leaves the score at 5 and closes this branch.
|
