experiment: freeze Rain EP bias confirmation

1 files changed, 54 insertions, 0 deletions

diff --git a/RAIN_EP_BIAS_CONFIRMATION.md b/RAIN_EP_BIAS_CONFIRMATION.md
new file mode 100644
index 0000000..9aa22c1
--- /dev/null
+++ b/RAIN_EP_BIAS_CONFIRMATION.md
@@ -0,0 +1,54 @@
+# Frozen Rain EP layer-state confirmation
+
+This protocol was frozen after the single-seed S1 development screen in
+`results/ep_bias/s1_summary.json`. S1 selected the only promoted corruption
+ratio, `0.01`; stronger `0.1` and `4.0` conditions remain recorded failures.
+
+## Fixed protocol
+
+- Author implementation: `rain-neuromorphics/energy-based-learning`, revision
+  `6b253fd8a5d267535f58ab79992256ef10031ceb`.
+- Model and learner: author `ConvHopfieldEnergy28` 32--64--10 with positive EP,
+  nudging `0.25`, 12 training relaxation iterations, 30 inference iterations,
+  author local parameter rules and SGD settings.
+- Data: FashionMNIST training set only. Data seed `6100` fixes disjoint 10,000
+  training and 2,000 validation examples. The test set is not evaluated.
+- Model/order seeds: `1989, 1990, 1991, 1992, 1993`; batch size 128; three
+  epochs; no epoch selection.
+- Conditions: clean, raw structured bias, same-RMS zero-mean noise, constant
+  predictor, affine innovation predictor, and oracle subtraction.
+- Bias: per-neuron affine function of the local first-phase state, normalized
+  by the experimenter to `0.01` times the initial clean layer-state-difference
+  RMS. This normalization is not visible to either predictor.
+- Predictor: normalized local LMS at rate `0.2`. Constant and innovation see
+  the same 128 instruction-off observations in the existing first EP phase of
+  the first training minibatch. Both are then frozen. There is no extra
+  equilibrium phase and no backpropagation.
+- Hardware: all six conditions for a seed run sequentially on one physical
+  GPU. Different seeds may run on GPUs 5 and 7 in parallel.
+
+The injected bias is deliberately inside the affine predictor class. Passing
+therefore confirms correction, causality and transfer to an EP implementation;
+it is not independent evidence that a real device exposes the same feature.
+
+## Frozen gate
+
+Across all five paired seeds:
+
+1. clean, noise, constant, innovation and oracle complete three finite epochs;
+2. mean clean validation accuracy is at least 70%;
+3. raw loses at least 10 accuracy points relative to clean in every seed;
+4. innovation beats raw and constant in every seed, with 95% lower bounds of
+   at least 10 and 5 accuracy points respectively;
+5. the 95% upper bounds on clean-minus-innovation, absolute
+   innovation-minus-oracle, and absolute noise-minus-clean are each below
+   three accuracy points;
+6. innovation has lower final residual/clean state-difference RMS than constant
+   in every seed;
+7. constant and innovation each use exactly 128 neutral observations;
+8. innovation/clean mean wall-time ratio is at most `1.15`;
+9. all records share one author revision, one SDIL revision, the fixed data
+   split and `autodiff_used_for_learning=false`.
+
+Failure closes this exact confirmation. It does not authorize a new ratio,
+predictor rate, seed replacement or extra calibration on the same holdout.