results: evaluate centered EP under hardware bias

1 files changed, 24 insertions, 0 deletions

diff --git a/RAIN_EP_RELEASED_PROFILE.md b/RAIN_EP_RELEASED_PROFILE.md
index c5bed5e..796f36e 100644
--- a/RAIN_EP_RELEASED_PROFILE.md
+++ b/RAIN_EP_RELEASED_PROFILE.md
@@ -81,3 +81,27 @@ split, one epoch compares centered clean, fixed raw/intercept/SDIL, released-
 profile raw/intercept/SDIL, and a released-profile oracle. Downstream holdout
 accuracy is the only selection endpoint. Residual diagnostics are retained
 only to catch implementation errors.
+
+R1 used 10,000 FashionMNIST training examples and a disjoint 2,000-example
+holdout drawn from the official training set, one epoch, one fixed seed, and
+the author's comparative ConvHopfieldEnergy32 network. Holdout accuracy was:
+
+| hardware condition | correction | accuracy |
+|---|---|---:|
+| no hardware bias | none | 46.00% |
+| constant per-parameter bias | none | 10.00% |
+| constant per-parameter bias | intercept-only | 45.45% |
+| constant per-parameter bias | affine SDIL | 45.45% |
+| released state-dependent profile | none | 32.70% |
+| released state-dependent profile | intercept-only, cadence 10 | 43.85% |
+| released state-dependent profile | affine SDIL, cadence 10 | 42.35% |
+| released state-dependent profile | exact oracle subtraction | 46.00% |
+
+This is positive evidence that a local neutral measurement can rescue a
+centered-EP learner from a damaging constant hardware offset: both local
+correctors recover 45.45% from 10.00%, close to the 46.00% clean endpoint.
+It is negative evidence for the stronger state-dependent claim at the frozen
+R1 setting: affine SDIL is 1.50 percentage points below the matched local
+intercept and 3.65 points below the oracle. The state-dependent result must
+therefore be treated as unconfirmed until a paired multi-seed accuracy study
+shows otherwise. No residual-bias diagnostic can override this endpoint.