protocol: freeze stagewise whitened causal feedback

1 files changed, 96 insertions, 0 deletions

diff --git a/ORAL_A_V6.md b/ORAL_A_V6.md
new file mode 100644
index 0000000..7663f27
--- /dev/null
+++ b/ORAL_A_V6.md
@@ -0,0 +1,96 @@
+# Oral-A-v6: stagewise causally whitened feedback without KP
+
+## Motivation and claim boundary
+
+V5 showed that per-example task perturbations contain useful causal
+information, but its interleaved scalar-normalized delta updates failed to
+carry that information into the early ResNet-20 layers.  After 20 sweeps, the
+readout feedback itself had parameter cosine only `0.1771`; early-third
+teaching alignment was `0.00115` even though several late layers reached
+`0.20--0.38`.
+
+One post-failure diagnostic, not a candidate result, used the same 2,560
+readout observations in a local least-squares fit.  It raised last-hidden
+teaching alignment to `0.8534`.  This shows that the observations have enough
+signal and identifies the failed online regression as a bottleneck.
+
+V6 is a new bounded mechanism, not a V5 rate or sweep-count recovery.  It
+changes both optimization and information timing:
+
+1. fit the readout feedback from its complete local causal dataset;
+2. freeze it, collect a complete causal dataset for the last convolutional
+   feedback edge, and apply one diagonal-whitened local least-squares update;
+3. freeze that edge and repeat toward the input.
+
+No parent edge observes data until its child path has been fitted.  The dense
+fit and convolutional sufficient statistics consume only stored task
+perturbation targets, local child contexts, and the current feedback tensor.
+They never read or copy a forward weight, use KP, or invoke reverse-mode
+learning.  V6 remains a scalable-vectorizer feasibility test, not yet the
+Harnett-specific SDIL contribution.
+
+## Local fits
+
+For the readout, concatenate the per-example output signals `C` and causal
+pooled-hidden targets `Y` and solve the ridge-scaled local normal equations
+
+```
+R^T = (C^T C + lambda I)^(-1) C^T Y,
+lambda = 1e-6 trace(C^T C) / n_classes.
+```
+
+For convolutional edge `l`, let `e = q - p` be the difference between the
+causal parent-field target and the complete current parent teaching field.
+Accumulate the ordinary local correlation numerator and its per-parameter
+diagonal curvature:
+
+```
+b_l = sum Corr(e, c_l)
+d_l = sum Corr(1, c_l^2)
+Q_l <- Q_l + b_l / (d_l + 1e-3 mean(d_l)).
+```
+
+This is a diagonal local least-squares step.  The ridge is scale-relative and
+fixed globally; there is no learning-rate or per-layer schedule.
+
+## V6-0 mechanics gate
+
+Before CIFAR endpoints, deterministic float64 checks must show:
+
+- the dense fit is invariant to changing every forward weight after causal
+  observations are stored, to absolute error below `1e-12`;
+- the convolutional fit has the same invariance and changes only its selected
+  feedback edge;
+- on independent Gaussian local contexts with responses produced by a known
+  feedback tensor, one diagonal fit reduces held-out field MSE by at least
+  `80%`;
+- one tiny reverse stage sequence is finite and all legacy convolutional
+  smoke tests remain green.
+
+## V6-1 frozen-forward causal-capture gate
+
+Copy V5's seed-0 ResNet-20 state, frozen 10,000-example development prefix,
+5,000-example validation split, 64-example exact-gradient audit, batch 128,
+evaluation-mode causal BatchNorm, `sigma=0.01`, and perturbation seed 5000.
+Use exactly 20 fixed unaugmented prefix batches for every stage.  There are 19
+stages (readout plus 18 convolutional edges), 20 events per stage, 380 events,
+760 logical batch-loss queries, and 48,640 per-example causal observations.
+Forward/readout weights, BatchNorm state/affines, and forward optimizer state
+must remain bitwise fixed.
+
+The only numerical constants are the readout relative ridge `1e-6` and
+convolutional diagonal ridge `1e-3` written above.  Record fixed HFA and the
+final V6 state.  V6-1 passes only if both are finite and V6:
+
+1. reaches early-third teaching/negative-gradient cosine at least `0.10`;
+2. reaches all-layer cosine at least `0.20`;
+3. improves early-third cosine over fixed HFA by at least `0.08`;
+4. keeps every feedback/forward norm ratio in `[0.1, 3.0]`;
+5. uses exactly the fixed event/query/observation counts and zero forward
+   weight reads or reverse-mode learning operations.
+
+There is no ridge, batch, stage-order, event-count, probe, clipping, or norm
+recovery after the endpoint.  Failure closes V6 and the present no-KP causal
+vectorizer route.  Passing opens only a separately frozen short task gate;
+capture alone cannot raise the reviewer score.
+