results: test SDIL on measured physical surrogate

2 files changed, 14 insertions, 0 deletions

diff --git a/results/figs/physical_bias_p1_surrogate.png b/results/figs/physical_bias_p1_surrogate.png
new file mode 100644
index 0000000..6888aa9
--- /dev/null
+++ b/results/figs/physical_bias_p1_surrogate.png
diff --git a/results/figs/physical_bias_p1_surrogate_caption.md b/results/figs/physical_bias_p1_surrogate_caption.md
new file mode 100644
index 0000000..7174c0b
--- /dev/null
+++ b/results/figs/physical_bias_p1_surrogate_caption.md
@@ -0,0 +1,14 @@
+**Development result on a measured-state-dependent two-edge surrogate.** A--B,
+combined task error after matching at least 6 s of nominal training time across
+switching periods. C--D, corresponding gate-space cycle span. The local affine
+bias fields are inferred from the released Dillavou et al. drift traces; they
+are not independent hardware measurements. Frozen constant calibration and
+SDIL receive the same 660/2040 neutral observations in the two task pairs,
+respectively, before learning. SDIL removes essentially the entire modeled
+raw-to-oracle gap and beats frozen constant calibration at every period. A
+constant estimator can also reach oracle performance when allowed another
+240--6000 online neutral observations per run. Most importantly, the ideal
+leading-order overclamping analogue reaches zero error without neutral
+observations and beats SDIL throughout. Thus this result verifies the local
+mechanism and observation tradeoff but does not pass the physical strong-
+baseline gate or demonstrate correction on hardware.