protocol: freeze convolutional HFA baseline funnel

1 files changed, 64 insertions, 0 deletions

diff --git a/HFA_BASELINE.md b/HFA_BASELINE.md
new file mode 100644
index 0000000..1dc5f87
--- /dev/null
+++ b/HFA_BASELINE.md
@@ -0,0 +1,64 @@
+# Convolutional HFA baseline protocol
+
+## Status and claim boundary
+
+This protocol was frozen before observing any convolutional hierarchical-FA
+accuracy endpoint. It adds a matched local-learning baseline after the
+hierarchical oracle localized the ResNet bottleneck to spatial child-error
+fields. HFA is not an SDIL novelty claim: recursive random feedback is prior
+art, and a strong HFA result cannot rescue the failed frozen Oral-A A3 gate or
+open its sealed A4 test panel.
+
+The implementation follows the actual option-A residual DAG. Every learned
+forward convolution has an independent fixed 3x3 feedback tensor; shortcut
+adjoints are parameter-free, and ReLU and BatchNorm Jacobians are local. The
+feedback code never reads a downstream forward convolution. In an audit-only
+symmetric test, setting the feedback tensors to the forward tensors and the
+readout feedback to the negative readout transpose reproduces hidden negative
+gradients to below `2e-12` relative error and the exact-BP update to below
+`2e-7` absolute error. Actual experiments never make this copy.
+
+## HFA-S1: bounded short screen
+
+Use the frozen Oral-A development split and the A2b setting: seed-0 ResNet-20,
+the first 10,000 post-split training examples, 5,000 validation examples,
+20 epochs, batch 128, ordinary CIFAR augmentation, cosine decay without
+warmup, momentum 0.9, weight decay `1e-4`, and final validation evaluation
+only. Output learning rate is 0.1. Cross hidden learning rate
+`{0.01, 0.03, 0.1}` with feedback scale 1 and the single predeclared feedback
+seed derived from model seed 0. A 32-example training-prefix alignment probe is
+diagnostic only.
+
+Select maximum final validation accuracy, then lower estimated MACs, then
+lower hidden learning rate. The short screen opens one full validation run only
+if the selected trajectory is finite and reaches at least 50% validation
+accuracy. For interpretation, the already observed matched short endpoints are
+DFA `37.16%`, failed-v1 SDIL `41.98%`, and BP `74.94%`; these are comparisons,
+not selectable thresholds.
+
+## HFA-S2: full validation baseline
+
+If HFA-S1 passes, train the selected HFA setting for 200 epochs on all 45,000
+development-training examples. Copy the A1 schedule exactly: batch 128,
+momentum 0.9, weight decay `1e-4`, output base rate 0.1, and hidden base rate
+selected by HFA-S1, with 10x drops at epochs 100 and 150 and no warmup. Evaluate
+the same 5,000-example validation split only at the end. No recovery grid is
+allowed.
+
+This is a baseline endpoint, not an advancement test for SDIL. A finite result
+at or above 80% is considered a strong matched local baseline that any learned
+hierarchical SDIL variant must compare against. A weaker result remains
+reportable and cannot trigger additional HFA tuning. Neither outcome permits
+test-set evaluation.
+
+## Accounting and stop rules
+
+- Record fixed feedback parameters, feedback-convolution MACs, local
+  correlation MACs, peak memory, wall time, and zero causal queries.
+- All three HFA-S1 jobs must use the same clean git commit and must not touch
+  CIFAR-10 test data.
+- No feedback scale, seed, optimizer, architecture, or extra learning rate is
+  added after the S1 endpoints are read.
+- HFA-S2, if opened, copies the selected S1 hidden rate mechanically and does
+  not select on intermediate validation accuracy.
+