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
|
# Visual contract: additive digital transfer
- **Artifact:** four-panel quantitative result figure.
- **Target venue / format:** ICLR two-column paper, full-width figure.
- **Core claim:** the same local instruction-off subtraction improves four
digital local-learning backbones under matched teaching-channel
imperfection.
- **Reviewer question:** is SDIL a correction that transfers across learning
rules, or does it work only in one custom network?
- **Evidence layer:** main digital transfer result.
- **Source data:**
`results/contrastive_bias/c1_gate.json`,
`results/ep_bias/c1_gate.json`,
`results/coupled_ladder/p2_confirm_side4.json`, and
`results/coupled_ladder/p3_overclamp_side4.json`.
- **Statistics / uncertainty:** Dual Propagation and EP resample five seeds;
CLLN panels average three device draws within each of 40 tasks and resample
tasks. Error bars are percentile 95% bootstrap intervals.
- **Figure prototype:** coordinated 2-by-2 bar-chart small multiples.
- **Panel map:** author-code Dual Propagation, author-code EP, digital coupled
learning, and digital overclamped coupled learning.
- **Exact label inventory:** clean, same-RMS noise, raw, static calibration,
SDIL, oracle, task accuracy, clean gap recovered.
- **Caption role:** define raw-to-clean gap recovery within each panel and
state that datasets and absolute clean levels differ across panels.
- **Manuscript placement:** Part 1, after the algorithm definition.
- **Output formats:** editable SVG, vector PDF, PNG preview, analysis JSON, and
source CSV.
- **Traceability:** every bar is regenerated by
`experiments/plot_digital_additive_transfer.py` from the four JSON sources.
- **Constraint:** the EP five-seed result beats raw in every seed and recovers
96% of the mean clean gap, while its stricter preregistered gate failed. The
manuscript and caption must retain that distinction.
|
