protocol: freeze reciprocal KP baseline funnel

1 files changed, 78 insertions, 0 deletions

diff --git a/KP_BASELINE.md b/KP_BASELINE.md
new file mode 100644
index 0000000..25696f7
--- /dev/null
+++ b/KP_BASELINE.md
@@ -0,0 +1,78 @@
+# Modified Kolen--Pollack baseline protocol
+
+## Attribution and boundary
+
+This is the reciprocal local-plasticity mechanism of Akrout et al., *Deep
+Learning without Weight Transport* (NeurIPS 2019), not an SDIL contribution.
+Their equations 16--18 give forward and feedback synapses equal adjustments
+from reciprocal local activity pairs and equal decay, so initially distinct
+weights converge without transmitting either weight. The public author
+repository is frozen for reference at revision
+`688f47addd2131684da1b7829b20365f585eee66`:
+
+<https://github.com/makrout/Deep-Learning-without-Weight-Transport>
+
+The public Python implementation is a fully connected pedagogical version and
+explicitly transposes a precomputed update. The implementation here instead
+recomputes each reciprocal convolutional correlation from its own cached
+parent activity and child teaching field. It follows the actual ResNet
+residual DAG, including local ReLU and BatchNorm Jacobians and parameter-free
+shortcut adjoints. The reciprocal update never reads a forward weight or a
+forward update tensor.
+
+KP is introduced because the frozen residual-response mirror full run tests a
+different, intermittent probe-response mechanism. KP is not a cadence or rate
+rescue of that branch. No KP task endpoint was generated before this protocol
+and its executable analyzer were committed.
+
+## KP-0: mechanics gate
+
+The convolutional smoke suite requires all of:
+
+- independently recomputed forward and reciprocal local directions agree to
+  absolute error below `1e-14` under the storage/sign convention;
+- changing every W and Q value after local activities are fixed changes the
+  reciprocal update by exactly zero;
+- exact `Q=W`, `R=-W_out^T` remains exact after two momentum-plus-decay steps;
+- the public KP task step leaves all parameters graph-free;
+- every pre-existing convolutional local-gradient and feedback check remains
+  green.
+
+The implemented errors are all exactly zero in float64.
+
+## KP-1: frozen 20-epoch useful-scale gate
+
+Run exactly one seed-0 ResNet-20 development record on the frozen 45k/5k
+CIFAR-10 split: batch 128, standard augmentation, hidden/output LR 0.1,
+epoch-100/150 step drops (therefore constant during this screen), no warmup,
+momentum 0.9, weight decay `1e-4`, BatchNorm, feedback scale 1, and a 32-example
+alignment audit. There is no LR, decay, initialization, or feedback-scale grid.
+
+The already frozen matched BP trajectory has 81.02% validation accuracy at
+epoch 20 and `1.09487808e14` linearly scaled training MACs. KP-1 passes only if:
+
+- the record, every epoch loss, and every tracking diagnostic are finite;
+- validation accuracy is at least 70% and within 15 points of BP epoch 20;
+- final early-third teaching alignment is at least 0.50;
+- final mean feedback/forward cosine is at least 0.80 and its epoch-11--20
+  mean is at least 0.70;
+- feedback learning uses zero task-loss queries;
+- total MACs, including a separate correlation at every reciprocal synapse,
+  are at most `1.40x` matched BP.
+
+Failure closes KP without an extra learning-rate/decay screen. Passing opens
+one full development run.
+
+## KP-2: conditional full baseline
+
+Copy KP-1 exactly for 200 epochs. KP-2 passes only if it is finite, reaches
+88% validation accuracy, retains early alignment 0.80, reaches final mean
+feedback/forward cosine 0.95, keeps the epoch-151--200 mean cosine at least
+0.95, uses zero task-loss queries, and costs at most `1.40x` the frozen full BP
+MACs. It never authorizes test access.
+
+KP remains an inherited baseline even if both gates pass, so it cannot raise
+the reviewer score. A KP-2 pass only opens a separately frozen
+raw-versus-norm-matched-raw-versus-innovation mixed-traffic experiment. The
+somato-dendritic innovation must be load-bearing there to affect the SDIL
+paper assessment.