experiments: freeze oral A development gates

1 files changed, 148 insertions, 0 deletions

diff --git a/ORAL_A.md b/ORAL_A.md
new file mode 100644
index 0000000..e3427c4
--- /dev/null
+++ b/ORAL_A.md
@@ -0,0 +1,148 @@
+# Oral-A preregistration: standard CIFAR residual networks
+
+## Claim boundary and status
+
+This protocol asks whether the low-query learned-feedback backbone can train a
+standard deep convolutional family without reverse-mode credit assignment.  It
+does **not** rescue the failed Oral-B biological-signature gate.  With
+`nuisance_scale=0`, the predictor is inactive and SDIL reduces to learned
+direct-feedback calibration in the sense of Lansdell et al.; standard-network
+scaling is therefore an engineering/credit-assignment contribution, while the
+Harnett-specific innovation claim remains supported only in the separately
+audited mixed-traffic experiments.
+
+No CIFAR image/ResNet accuracy was observed before this protocol was frozen.
+Mechanics-only CPU checks are complete: standard ResNet parameter counts,
+option-A shortcuts, augmentation determinism, local Conv+BatchNorm eligibility,
+BatchNorm-coupled causal perturbations, running-stat isolation, and both apical
+vectorizers pass `experiments/cifar_image_smoke.py` and
+`experiments/conv_local_smoke.py`.
+
+## Fixed data, architecture, and accounting
+
+- CIFAR-10 uses the original Python batches under `/home/yurenh2/sdrn/data`.
+- Development uses a 45,000/5,000 training-only stratified split with seed 2027,
+  validation-index SHA-256
+  `8328b206a97c420e49e54e3eca4abe3274c4756b084355784ea3fb8059e4515b` and
+  exactly 500 validation examples per class.
+- Training augmentation is random 32x32 crop after four-pixel zero padding and
+  horizontal flip with probability 0.5. Validation/test are not augmented.
+- The architecture is the canonical CIFAR `6n+2` ResNet topology with 3x3
+  convolutions, BatchNorm, ReLU, global average pooling, and parameter-free
+  option-A shortcuts. ResNet-20/32/56 contain respectively 269,722, 464,154,
+  and 853,018 forward parameters.
+- BatchNorm's complete minibatch Jacobian is evaluated within its own layer.
+  No local update reads a downstream forward weight. Counterfactual `+/-`
+  trials are separate batches and do not update running statistics.
+- Because BatchNorm couples examples, causal calibration uses the batch-mean
+  loss derivative multiplied by batch size. This is an unbiased estimate of
+  the summed-loss hidden gradient. Logical batch queries, observed causal
+  scalars, per-example CE terms, forward-equivalent examples, MAC components,
+  peak memory, and wall time are all recorded separately.
+- SDIL uses one antithetic direction every four task updates (`K=1/every=4`),
+  selected before this project reached convolutional models by the passed C3
+  gate. `sigma=0.01`; no direction/cadence tuning is allowed here.
+
+## A1: establish a valid exact-BP reference
+
+The primary seed-0 ResNet-20 validation run uses 200 epochs, batch 128, SGD
+momentum 0.9, weight decay `1e-4`, learning rate 0.1 with 10x drops at epochs
+100 and 150, and no warmup. It must reach at least 90% final validation
+accuracy with finite loss.
+
+If and only if the primary fails, one predeclared architecture-recovery run is
+allowed: cosine learning rate from 0.1, five warmup epochs, and weight decay
+`5e-4`, with everything else fixed. The higher final validation endpoint is
+selected; if neither reaches 90%, the current implementation is not a valid
+standard-network testbed and all later Oral-A stages stop. This recovery tests
+the forward architecture/optimizer, not SDIL.
+
+## A2a: apical calibration screen with frozen forward weights
+
+On seed 0, ResNet-20, the first 10,000 post-split training examples, and 100
+feedback-only minibatches, cross:
+
+- vectorizer: `spatial_template`, `channel_gated`;
+- initial feedback scale: `0`, `1`;
+- apical regression rate: `0.001`, `0.01`, `0.1`.
+
+Forward weights, BN affine parameters, and the readout remain bitwise fixed.
+Each run uses K1 antithetic BatchNorm-coupled calibration and a fixed 32-example
+training-prefix alignment probe. A candidate is eligible only if all values are
+finite and early-third teaching/negative-gradient cosine is positive. Within
+each vectorizer family, select maximum early-third cosine, then maximum
+all-layer mean cosine, then lower `eta_A`, then lower `a_scale`. Both families
+must yield an eligible candidate or A2 stops.
+
+The spatial template has one class vector per spatial unit. The channel-gated
+candidate shares `A c + tanh(h) * G c` over positions; at ResNet-56 it has
+40,640 rather than 5,324,800 vectorizer parameters. This capacity/cost choice
+is frozen before accuracy is observed.
+
+## A2b: bounded short accuracy screen
+
+Use seed 0, ResNet-20, 10,000 training examples, 20 epochs, final validation
+only, cosine decay, no warmup, momentum 0.9, and weight decay `1e-4`:
+
+- one exact-BP run at hidden/output LR 0.1;
+- DFA at hidden LR `{0.01, 0.03, 0.1}`, output LR 0.1;
+- each A2a-selected SDIL vectorizer at the same three hidden rates, output LR
+  0.1, its selected A hyperparameters, and 100 A-only warmup minibatches.
+
+Select DFA by maximum final validation accuracy, then lower MAC cost/rate.
+Select SDIL by maximum accuracy, except that a channel-gated candidate within
+0.5 points of the global maximum wins; remaining ties use lower total MACs,
+then lower hidden rate. Advancement requires BP at least 50%, SDIL at least
+35%, SDIL no more than 10 points below selected DFA, and all trajectories
+finite. Failure stops Oral-A rather than expanding the grid.
+
+## A3: full seed-0 ResNet-20 development gate
+
+Train the A1-selected BP schedule, A2b-selected DFA, and A2b-selected SDIL for
+200 epochs on all 45,000 training examples. DFA/SDIL retain their selected
+hidden base rate, use output base rate 0.1 and the same epoch-100/150 10x drops;
+SDIL retains its 100-step A-only warmup and K1/every4 joint calibration. No
+nuisance traffic or predictor learning is present.
+
+All of the following must hold to open the independent test panel:
+
+1. BP final validation accuracy is at least 90%.
+2. SDIL is within 5 points of BP and at least 2 points above tuned DFA.
+3. SDIL early-third teaching alignment is at least 0.05.
+4. Every loss/metric is finite and SDIL's estimated total training MACs do not
+   exceed exact BP's estimate.
+
+No recovery branch follows an A3 failure.
+
+## A4: independent ResNet-20/32/56 confirmation
+
+Only after A3 passes, run exact BP, selected DFA, and selected SDIL at depths
+20, 32, and 56 for model/data-loader seeds 10--14. Hyperparameters and epoch
+budgets are copied without depth-specific tuning. The CIFAR-10 test set is
+evaluated exactly once at the end of every run; no intermediate test metric is
+logged or used for selection.
+
+The Oral-A gate requires:
+
+1. all 45 trajectories are finite;
+2. at every depth, mean SDIL test accuracy is within 2 points of exact BP;
+3. at depth 56, SDIL exceeds DFA by at least 2 points;
+4. paired SDIL depth-20 to depth-56 accuracy changes by no less than -2 points;
+5. depth-56 early-third alignment is at least 0.05 and at least 30% of the
+   depth-20 value;
+6. SDIL remains nondominated among matched local methods in accuracy versus
+   estimated MACs, logical causal queries, and peak memory. Wall time is shown
+   but cannot substitute for these hardware-independent coordinates.
+
+Native BurstCCN and Dual Prop runs remain separate executable-fidelity
+comparisons because their architectures, epochs, and objectives differ. Their
+published multi-seed values are not pooled with this matched panel.
+
+## Stop rules
+
+- No test seed 10--14 may be touched before A3 passes.
+- No failed threshold may be weakened after observing an endpoint.
+- No extra depth, direction count, perturbation frequency, vectorizer, or
+  optimizer branch is added after its corresponding screen closes.
+- A strong standard-network result cannot be described as evidence that the
+  natural Harnett causal-role/error-derivative signature emerged.