From e9f1342fc8e233a4841b7eb3c1363324e90ecda9 Mon Sep 17 00:00:00 2001 From: YurenHao0426 Date: Sat, 29 Aug 2026 18:43:05 -0500 Subject: figure: show SDIL transfer across digital learners --- visual-composer/digital-additive-transfer.md | 33 ++++++++++++++++++++++++++++ 1 file changed, 33 insertions(+) create mode 100644 visual-composer/digital-additive-transfer.md (limited to 'visual-composer/digital-additive-transfer.md') diff --git a/visual-composer/digital-additive-transfer.md b/visual-composer/digital-additive-transfer.md new file mode 100644 index 0000000..74867f8 --- /dev/null +++ b/visual-composer/digital-additive-transfer.md @@ -0,0 +1,33 @@ +# 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. -- cgit v1.2.3