figure: consolidate hardware-realistic CLLN evidence

1 files changed, 40 insertions, 0 deletions

diff --git a/visual-composer/physical-hardware-evidence.md b/visual-composer/physical-hardware-evidence.md
new file mode 100644
index 0000000..b271127
--- /dev/null
+++ b/visual-composer/physical-hardware-evidence.md
@@ -0,0 +1,40 @@
+# Visual contract: hardware-realistic CLLN evidence
+
+- **Artifact:** three-panel quantitative result figure.
+- **Target venue / format:** ICLR two-column paper, full-width figure.
+- **Core claim:** local innovation subtraction restores nonlinear CLLN
+  learning under published component imperfections and remains effective with
+  nonideal local sampling.
+- **Reviewer question:** does the method survive component mismatch, sampler
+  mismatch, noise, stale refresh, and their combination at the task endpoint?
+- **Evidence layer:** hardware-realistic simulation; no fabricated-chip claim.
+- **Source data:**
+  `results/physical_bias/p5_full_grid_bias_crossover.json`,
+  `results/physical_bias/p9_grid_correlated_autozero.json`, and
+  `results/physical_bias/p8_spice_autozero_primitive.json`.
+- **Statistics / uncertainty:** four device draws are averaged within each of
+  40 tasks; tasks are resampled for percentile 95% bootstrap intervals.
+- **Figure prototype:** horizontal method comparison, mismatch response curves,
+  and horizontal nonideality comparison.
+- **Panel map:**
+  - (a) nonlinear CLLN final error across clean, raw, calibration,
+    overclamping, SDIL, and their combination;
+  - (b) final error against measured local sampling-error RMSE for pedestal
+    and gain mismatch sweeps;
+  - (c) final error under common pedestal, sampling noise, stale neutral
+    refresh, and combined nonidealities.
+- **Exact label inventory:** nonlinear CLLN, component errors, sample-path
+  mismatch, residual sampling error, nonideal local sampling, final
+  classification error.
+- **Caption role:** identify the published component-error scale, the local
+  correlated-double-sampling operation, task/device counts, and the boundary
+  between simulation and fabricated hardware.
+- **Manuscript placement:** Part 3, after the digital scaling result.
+- **Output formats:** editable SVG, vector PDF, PNG preview, analysis JSON, and
+  source CSV.
+- **Traceability:** every plotted aggregate is regenerated by
+  `experiments/plot_physical_hardware_evidence.py`; the SPICE primitive is
+  carried in the analysis manifest and reserved for supplementary evidence.
+- **Constraint:** circuit-solver failures remain classification failures.
+  Sampler RMSE excludes the missing diagnostic value on failed solves, while
+  the task endpoint includes every trial.