experiment: freeze ResNet crossover protocol

1 files changed, 121 insertions, 0 deletions

diff --git a/RESNET_CROSSOVER.md b/RESNET_CROSSOVER.md
new file mode 100644
index 0000000..28c8395
--- /dev/null
+++ b/RESNET_CROSSOVER.md
@@ -0,0 +1,121 @@
+# Residual-CNN crossover protocol
+
+## Scope and immutable topology
+
+This is the residual-family realization of
+`CROSS_ARCHITECTURE_CROSSOVER.md`.  The primary validation panel contains all
+nine registered methods at ResNet-20, ResNet-32, and ResNet-56, hence 27
+cells.  A failed, nonfinite, chance, out-of-memory, or timed-out cell remains
+in the panel.  No endpoint may remove a method or collapse a depth.
+
+Every cell uses `CIFARLocalResNet`: the standard CIFAR `6n+2` basic-block
+topology, base width 16, BatchNorm after every convolution, option-A identity
+shortcuts, residual multiplier 1, global average pooling, and a linear
+10-class readout.  Trainable forward tensor shapes and counts are identical
+within a depth.  The shared data path is the stratified seed-2027 45k/5k
+CIFAR-10 train/validation split, standard crop/flip augmentation, batch 128,
+float32, model/loader seed 0, momentum 0.9, and weight decay `1e-4`.  Test is
+loaded by the inherited data utility but never iterated, selected, or
+reported in P1/P2.
+
+## Method adapters and completed R0 mechanics
+
+- BP is exact reverse mode on the common forward graph.
+- Ordinary FA is `CIFARHierarchicalFAResNet`: independent fixed convolutional
+  feedback follows every child edge; option-A shortcut adjoints and local
+  ReLU/BatchNorm Jacobians are exact.
+- DFA is the existing independent output-to-population spatial vectorizer.
+- PEPITA performs the ordinary and error-modulated presentations, with a
+  fixed output-error-to-input He-uniform projection times 0.05.  A hidden
+  first-minus-second activity field directly multiplies the modulated
+  presynaptic state; convolutional correlations average spatial positions.
+- Forward-Forward overlays a candidate label on the first ten input
+  coordinates, normalizes each edge input, greedily trains one residual-DAG
+  population at a time with the softplus goodness objective, and scores all
+  candidates using layers after the first.  BatchNorm updates only at the
+  target layer.  The linear readout is the final goodness layer.
+- Equilibrium Propagation has one hard-sigmoid state per residual-DAG
+  population, synchronous leaky dynamics, 20 free and 4 nudged steps,
+  `dt=0.5`, beta magnitude 0.5 with random sign, squared output cost, symmetric
+  forward-edge transposes, and free-versus-nudged local correlations.
+- Dual Propagation uses the author DP-transpose setting: alpha 0, beta 0.1,
+  16 topological `fwK` inference passes, and the same forward edge for
+  symmetric feedback.  Each option-A shortcut contributes its exact
+  parameter-free transpose.
+- clean-KP is the existing independently recomputed reciprocal-correlation
+  rule.  It never reads a forward weight or forward update.
+- SDIL uses that same KP substrate plus four-RMS soma-predictable traffic, a
+  64-example instruction-free closed-form predictor fit, and the paired
+  neutral projection on every task minibatch.
+
+The deterministic float64 audit requires matched forward outputs or forward
+parameter counts and explicit local equations.  Current errors are:
+Dual-Propagation contrastive directions `1.07e-15`; PEPITA local directions
+`1.52e-14`; EP contrastive directions `3.36e-14`; and Forward-Forward local
+objective exactly zero error with zero non-target gradients and parameter
+changes.  Existing BP/FA/DFA/KP/SDIL mechanics remain covered by
+`experiments/conv_local_smoke.py`.  Autograd is used only by BP, by the
+single detached target layer in FF, and by independent post-training
+alignment diagnostics.
+
+## R1 bounded validation-only selection
+
+R1 uses ResNet-20, seed 0, ten constant-rate epochs, full 45k training data,
+5k validation data, validation after every epoch, no early stopping, no
+alignment probe, and no test evaluation.  Fixed canonical or previously
+validated settings are BP 0.1, FF 0.03, clean-KP 0.1, and SDIL 0.1.  The
+complete bounded grids are:
+
+| Method | Hidden/base learning rates | Output rate |
+|---|---|---|
+| ordinary FA | 0.01, 0.03, 0.1 | 0.1 |
+| DFA | 0.01, 0.03, 0.1 | 0.1 |
+| PEPITA | 0.0003, 0.001, 0.003 | same |
+| EP | 0.0003, 0.001, 0.003 | same |
+| Dual Propagation | 0.01, 0.025, 0.05 | same |
+
+Select maximum best validation accuracy, then a finite final loss, then the
+lower rate.  If every candidate is nonfinite, select the lowest rate and
+retain all failures.  The selected rate is copied without depth-specific
+tuning to ResNet-20/32/56.  R1 development records are not primary panel
+cells.
+
+## R2 complete 27-cell validation panel
+
+BP, ordinary FA, DFA, clean-KP, SDIL, and Dual Propagation receive 200 global
+epochs with 10x drops at epochs 100 and 150; their R1-selected base rates and
+the fixed output rates above are unchanged.  PEPITA receives 100 epochs with
+its selected rate and 10x drops at epochs 60 and 90.  FF receives 40 epochs
+per greedy layer at its fixed Adam rate.  EP receives 100 constant-rate
+epochs with its selected rate and fixed relaxation parameters.  Each method
+sees the same 45k ordinary examples per applicable global epoch; extra
+presentations, candidate-label passes, free/nudged relaxations, symmetric
+feedback, and local correlations are charged separately.
+
+Every cell records the complete command, source and protocol hashes, selector
+hash, split and CIFAR source hashes, final and best validation metrics,
+completed epochs, first nonfinite step, wall time, physical GPU, peak allocated
+and reserved memory, parameter count, forward MACs, ordinary examples,
+presentations, candidate evaluations, relaxation passes, causal task-loss
+queries, and method-specific invariants.  A per-cell 48-hour wall cap is fixed
+before launch.  A capped or failed cell is recorded and the shard continues.
+The complete analyzer rejects missing/duplicate cells, source or architecture
+drift, test evaluation, ordinary-example drift, and silent failure omission.
+
+## R3 untouched confirmation
+
+No R3 test endpoint opens until all 27 R2 records and the complete analyzer
+are committed.  A later commit must freeze the untouched seeds and evaluate
+every method/depth cell, never an SDIL-only winner subset.  Native author
+records remain separately labelled fidelity evidence and cannot substitute
+for a matched cell.
+
+## Prespecified analyses
+
+Report all methods at all depths in accuracy-versus-wall and
+accuracy-versus-hardware-independent-work panels, with failure markers.
+Report ResNet-20 to ResNet-56 paired change, forward parameter count, and an
+independent fixed alignment probe.  clean-KP equality in a clean task is a
+substrate sanity check, not an SDIL novelty result.  A residual-family scaling
+claim requires the complete panel and cannot be inferred from the already
+confirmed four-method panel alone.