experiment: freeze cross-architecture crossover matrix

1 files changed, 192 insertions, 0 deletions

diff --git a/CROSS_ARCHITECTURE_CROSSOVER.md b/CROSS_ARCHITECTURE_CROSSOVER.md
new file mode 100644
index 0000000..3d2ea1e
--- /dev/null
+++ b/CROSS_ARCHITECTURE_CROSSOVER.md
@@ -0,0 +1,192 @@
+# Cross-architecture local-learning crossover
+
+## Purpose and claim boundary
+
+This protocol tests two claims without selecting architectures or methods by
+their observed endpoints:
+
+1. SDIL lies on the accuracy--training-cost Pareto frontier among matched
+   local-learning methods.
+2. SDIL retains useful optimization as the size of a standard architecture
+   family increases.
+
+The primary development panel contains nine architecture--scale points and
+nine training methods, hence 81 cells.  A failed, nonfinite, chance-level, or
+timed-out cell is a result and remains in the panel.  A method cannot be
+removed from larger models because it was weak on a smaller model, and a size
+cannot be collapsed into an architecture-family average.  Smoke runs verify
+mechanics only and never replace a complete cell.
+
+This panel is a matched crossover, not a claim that all methods have equal
+wall time or implement the same inference process.  Accuracy, wall time,
+forward-equivalent work, peak memory, task-loss queries, relaxation steps, and
+method-specific observations are reported separately.  Native author results
+remain in `BASELINES.md` and are never silently mixed with matched cells.
+
+## Frozen cell registry
+
+The method axis is:
+
+| ID | Method |
+|---|---|
+| `bp` | exact reverse-mode backpropagation |
+| `fa` | ordinary layer-recursive feedback alignment |
+| `dfa` | direct feedback alignment |
+| `pepita` | PEPITA |
+| `ff` | Forward-Forward |
+| `ep` | equilibrium propagation |
+| `dualprop` | Dual Propagation |
+| `clean_kp` | reciprocal Kolen--Pollack learning without nuisance traffic |
+| `sdil` | dynamic somato-dendritic innovation learning |
+
+`fa` is not DFA and is not a learned feedback rule.  Its feedback tensors are
+independent fixed random draws.  On a residual DAG it follows the actual
+parent--child graph, including exact parameter-free shortcut and local
+normalization/activation Jacobians, but it never reads or copies a downstream
+forward convolution.  This is the existing hierarchical FA mechanism under
+the paper-facing ordinary-FA label.  `clean_kp` and `sdil` use the same
+reciprocal credit-transport substrate; SDIL alone receives the frozen
+four-RMS soma-predictable traffic and paired-neutral subtractive innovation
+operation.
+
+The architecture--scale axis is:
+
+| Family | ID | Frozen forward topology |
+|---|---|---|
+| plain CNN | `minicnn` | author miniCNN: C64-MP, C64-MP, linear-10 |
+| plain CNN | `vgglike` | author VGGlike: C128-MP, C256-MP, C512-MP, C512-MP, linear-10 |
+| plain CNN | `vgg16` | author CIFAR VGG16: 13 convolutions with 64/128/256/512 stages, five max-pools, dense 4096, dense 4096, linear-10 |
+| residual CNN | `resnet20` | standard CIFAR `6n+2` ResNet, base width 16 |
+| residual CNN | `resnet32` | standard CIFAR `6n+2` ResNet, base width 16 |
+| residual CNN | `resnet56` | standard CIFAR `6n+2` ResNet, base width 16 |
+| decoder Transformer | `transformer4` | four decoder blocks |
+| decoder Transformer | `transformer8` | eight decoder blocks |
+| decoder Transformer | `transformer12` | twelve decoder blocks |
+
+The plain-CNN definitions are copied from the frozen Dual Propagation author
+revision `7b2595b34421e1483a721dbfdeff8cdabda3a1ff`, not reconstructed from a
+name.  Its VGG16 has convolutional widths
+`64,64,128,128,256,256,256,512,512,512,512,512,512`, pooling after
+convolutions 2, 4, 7, 10, and 13, then two 4096-unit ReLU layers.  The
+residual-CNN definitions are the already audited `CIFARLocalResNet`
+implementations.
+
+The Transformer depth axis is frozen here, but its exact task, width, context,
+and algorithm adapters must be frozen in a separate tracked family protocol
+before any formal Transformer endpoint is launched.  In particular, an
+adapter must specify how PEPITA perturbs continuous token embeddings, how
+Forward-Forward constructs positive/negative autoregressive examples, and how
+EP and Dual Propagation expose relaxable block states.  A generic local
+cross-entropy surrogate may not be relabelled as one of those methods.
+
+Additional width or depth ladders may be added only as a new complete family
+revision containing every method.  They do not replace, merge, or selectively
+extend the 81 primary cells.
+
+## Matched family rules
+
+Plain CNN and residual CNN use CIFAR-10 with the same normalized images,
+augmentation, train/validation split, minibatch order, initial forward
+parameters, and evaluation batches within an architecture--seed pair.
+Transformer methods use the same tokenized data bytes, sampled token
+positions, forward initialization, and evaluation batches within a
+depth--seed pair.  The following rules hold in every family:
+
+- the forward topology and number of trainable forward parameters are matched;
+- BP uses exact reverse mode, while non-BP methods may use autograd only to
+  evaluate explicitly local Jacobian-vector products whose boundaries are
+  detached and audited;
+- FA/DFA random feedback never changes; KP feedback changes only through its
+  reciprocal local correlation and never reads a forward weight or update;
+- SDIL uses the same KP update plus the already frozen neutral observation,
+  predictor, traffic, and innovation rules;
+- every method receives the same number of ordinary training examples and the
+  same endpoint-selection split; extra queries, forwards, perturbations, or
+  inference iterations are charged rather than hidden;
+- method-specific inference hyperparameters follow the cited paper or author
+  implementation when available.  A bounded validation-only selection grid
+  must be committed before its endpoints and applied at all three sizes;
+- no test example is used for tuning, early stopping, divergence recovery, or
+  choosing a method variant.
+
+Forward-Forward, EP, and Dual Propagation necessarily have method-specific
+state or example construction.  Their matched cells preserve the frozen
+forward parameterization and task, while the native-author ledger separately
+shows executable fidelity.  Any unavoidable semantic deviation is named in
+the family protocol and table rather than described as exact reproduction.
+
+## Stages and completeness
+
+### X0: executable-rule audit
+
+Before a formal endpoint, each family adapter must pass deterministic CPU or
+one-minibatch tests that establish:
+
+- identical forward outputs and parameter counts across method wrappers at
+  initialization;
+- exact BP agreement in an audit-only symmetric-feedback limit for FA/KP;
+- independence of actual FA/DFA feedback from forward weights;
+- absence of reverse-graph leakage across detached local boundaries;
+- paper-equation agreement for PEPITA, Forward-Forward, EP, and Dual
+  Propagation on a small deterministic network;
+- clean-KP/SDIL identity when traffic is disabled and innovation prediction is
+  zero;
+- SDIL raw, matched-raw, and innovation signal identities from
+  `KP_TEACHING_SIGNAL_ABLATION.md`;
+- finite cost counters whose analytic values agree with enumerated operations.
+
+One-epoch or short-model GPU smokes may check compilation, memory, and runtime.
+They cannot select learning rates, inference steps, architecture cells, or
+reported endpoints.
+
+### X1: complete validation matrix
+
+Every family freezes its optimizer, schedule, validation-only hyperparameter
+grid, cap on inference/relaxation work, seed, split, hardware policy, and
+failure policy before X1.  X1 is complete only when all 81 registered cells
+have an immutable record from that frozen protocol.  The analyzer must reject
+duplicates, missing cells, untracked or dirty source, architecture drift,
+dataset drift, and method-dependent ordinary-example counts.
+
+The primary metric is validation top-1 accuracy for CIFAR-10 and validation
+negative log-likelihood/perplexity for autoregressive language modeling.
+Each record also includes final and best-validation values, the selection
+rule, wall time, peak allocated and reserved memory, total forward-equivalent
+examples, affine MACs or FLOPs, task-loss queries, local observation counts,
+relaxation passes, first nonfinite step, source revision, environment freeze,
+dataset hashes, and physical GPU identity.
+
+There is no numerical pass threshold for retaining a method in X1.  Timeout
+caps are architecture- and method-independent functions of the frozen BP
+runtime plus a predeclared multiplier; a timeout is plotted at its achieved
+metric and charged elapsed cost.  Nonfinite cells are plotted at chance (or
+their last finite language loss) with a failure marker, not omitted.
+
+### X2: complete confirmation matrix
+
+X2 uses untouched seeds and the X1-selected settings without a new grid.  The
+number and identities of confirmation seeds are committed before any X2
+endpoint.  Every one of the 81 cells receives the same seed set; a partial
+SDIL-only or winner-only confirmation cannot support the crossover claim.
+Test is evaluated once per completed confirmation record.  X1 and X2 remain
+separate in all tables.
+
+## Paper-facing analyses
+
+The crossover produces two prespecified views:
+
+1. An accuracy/loss versus measured-cost plot containing every method,
+   architecture size, timeout, and divergence marker.  A method is on the
+   empirical Pareto frontier only if no observed point is at least as accurate
+   and no more costly, with one strict inequality.  Wall time and
+   hardware-independent work receive separate panels.
+2. A within-family scaling plot showing the paired change from smallest to
+   largest size for every method.  Raw endpoints, parameter counts, and
+   teaching/negative-gradient alignment by depth are shown; family averages
+   cannot hide a failed largest model.
+
+The strongest acceptable conclusion is conditional: under the frozen tasks,
+budgets, and adapters, SDIL is Pareto-competitive and scales across the tested
+families.  The panel does not prove asymptotic scaling, biological
+implementation, clean-setting superiority over BP, or universal superiority
+over EP/Dual Propagation.