diff options
Diffstat (limited to 'visual-composer/coupled-ladder-scaling.md')
| -rw-r--r-- | visual-composer/coupled-ladder-scaling.md | 37 |
1 files changed, 37 insertions, 0 deletions
diff --git a/visual-composer/coupled-ladder-scaling.md b/visual-composer/coupled-ladder-scaling.md new file mode 100644 index 0000000..c339551 --- /dev/null +++ b/visual-composer/coupled-ladder-scaling.md @@ -0,0 +1,37 @@ +# Visual contract: digital CLLN scaling pilot + +- **Artifact:** three-panel quantitative result figure. +- **Target venue / format:** ICLR two-column paper, full-width figure. +- **Core claim:** local innovation subtraction preserves digital coupled + learning under fixed component imperfection as the lattice grows. +- **Reviewer question:** does SDIL retain a task-level advantage at large + system size, and is the result stronger than static calibration or matched + zero-mean noise? +- **Evidence layer:** main scaling result, currently labelled exploratory. +- **Source data:** + `results/coupled_ladder/p1_imperfection_pilot.json` and + `results/coupled_ladder/p1_bias_baseline_pilot.json`. +- **Statistics / uncertainty:** component draws are averaged within task; + tasks are the bootstrap unit. The pilot has five task clusters and one + component draw per task and size. Error bars are percentile 95% intervals. +- **Figure prototype:** coordinated small-multiple lines. +- **Panel map:** + - (a) final classification error against learnable edge count; + - (b) fraction of runs that reach zero error and remain there through the + training horizon; + - (c) mean classification error over epochs at 2,048 edges. +- **Exact label inventory:** Clean, same-RMS noise, raw imperfection, static + calibration, SDIL, learnable edges, classification error, stable zero-error + runs, training epoch. +- **Caption role:** state the paired protocol, distinguish fixed bias from + matched noise, and label the evidence as a five-task pilot. +- **Manuscript placement:** Part 2, immediately after the digital CLLN and + component-imperfection definitions. +- **Output formats:** editable SVG, vector PDF, PNG preview, analysis JSON, + and source CSV. +- **Traceability:** every plotted aggregate is regenerated by + `experiments/analyze_coupled_ladder_scaling.py` from the two JSON files. +- **Constraint:** the figure reports an error-growth slope comparison only + when the static-calibration slope is positive. Raw imperfection is allowed + to show a high non-monotonic floor rather than a forced power law. + |