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
|
# Frozen nonlinear CLLN confirmation
## Purpose
Confirm the hardware-realistic result on new component draws after the
nonlinear model, local predictor, baselines, task set, and endpoints have been
fixed. This is an in-silico component confirmation, not a fabricated-chip
experiment.
## Frozen protocol
- Simulator: `sdil/physical_grid.py` and
`experiments/physical_grid_bias_p5.py`.
- Topology: released 4-by-4 periodic CLLN with 32 learnable edges.
- Tasks: all 40 released ring-classification tasks: five input diameters and
eight label rotations.
- New component seeds: `20260833,20260834,20260835`.
- Published-scale component model: differential gain standard deviation
`0.01`, twin input mismatch `0.001 V`, multiplier offset `2.3 V/s`.
- Methods: clean, raw imperfection, 16-observation constant calibration,
16-observation degree-2 SDIL predictor, clean overclamping, imperfect
overclamping, and imperfect overclamping plus the same SDIL predictor.
- Standard learning: 600 epochs and `0.001 s` learning exposure per active
update.
- Overclamping: at most 1,000 epochs, released nudging, `0.0025 s/V`, and the
existing three-perfect-checkpoint early stop.
- No BP, autograd, device constant, or oracle-neutral method is used.
- Pairing: all six methods receive the same task, initial gates, and component
draw.
## Endpoints and statistics
The primary endpoint is final classification error. Stable zero-error fraction,
wall time, clamp displacement, and held-out predictor RMSE are secondary. The
40 tasks are the bootstrap unit; the three component draws are averaged within
task before percentile 95% intervals are computed.
The result supports confirmation only if:
1. every method completes all 120 task/device cells with finite task metrics;
2. standard SDIL improves raw error in every task-averaged pair and closes at
least 95% of the raw-to-clean mean error gap;
3. standard SDIL has no more than one percentage point excess mean error over
clean;
4. overclamping plus SDIL closes at least 95% of the overclamp-to-clean-
overclamp mean error gap;
5. constant calibration and uncorrected overclamping remain in the result
regardless of outcome.
The output is
`results/physical_bias/p11_nonlinear_hardware_confirmation.json`. Development
seeds `20260829` through `20260832` remain separate.
|
