protocol: freeze dynamic projection training gate

1 files changed, 90 insertions, 0 deletions

diff --git a/DYNAMIC_INNOVATION.md b/DYNAMIC_INNOVATION.md
new file mode 100644
index 0000000..19e39da
--- /dev/null
+++ b/DYNAMIC_INNOVATION.md
@@ -0,0 +1,90 @@
+# Dynamic neutral-projection development protocol
+
+## Claim boundary
+
+This is a new post-S0 development branch.  It does not reinterpret or rescue
+MT-1: MT-1 remains failed, MT-2/MT-3 remain sealed, and the reviewer score
+remains 5/10.  S0 ruled out extending a fixed one-sided predictor margin.  The
+new question is whether a fast *local and instruction-off* projection can keep
+the multiplicative neutral-residual mode inside its two-sided stability window
+as somatic statistics change.
+
+The slow predictor is still fitted exactly once on the original 64-example
+neutral calibration prefix and is frozen during task learning.  Before each
+task update, every cell receives a paired instruction-off observation of its
+current soma `h` and ordinary apical traffic `a0`.  From the neutral residual
+
+```text
+e0 = a0 - (P h + b),
+```
+
+it forms the current-batch local affine projection
+
+```text
+q = Cov(e0, h) / Var(h),
+e_stable = (e0 - mean(e0)) - q (h - mean(h)).
+```
+
+The plasticity signal is `r = s + e_stable`.  The coefficient fit sees zero
+task-instruction observations; it reads no label, loss, forward weight,
+feedback weight, or downstream state, and it does not modify the slow
+predictor.  For the diagonal affine traffic used by frozen MT-1, the controller
+nulls the current empirical residual-coupling coefficient.  With weight decay,
+the scalar homogeneous mode is therefore placed near `k=-lambda`, inside both
+Jury boundaries, instead of at a fixed negative coefficient whose product with
+an evolving covariance can cross the lower boundary.
+
+This controller requires one paired neutral apical microphase per task batch.
+Its elementwise work and observation count must be charged separately in any
+later endpoint comparison.  It uses zero task-loss perturbation queries and no
+reverse-mode differentiation.
+
+## D0: mechanics (completed before task-data evaluation)
+
+The implementation must prove on synthetic float64 tensors that:
+
+- an inaccurate frozen slow predictor leaves a nonzero affine neutral mode;
+- fast projection reduces the post-projection neutral/traffic RMS ratio and
+  residual-soma slope below `1e-14` for diagonal affine traffic;
+- projected innovation equals clean instruction below `1e-14`;
+- the projection consumes zero instruction observations and leaves every slow
+  predictor parameter bitwise unchanged;
+- forward and reciprocal correlations remain independently computed; and
+- the extra elementwise work is exposed explicitly.
+
+These checks passed at clean revision `eeb50b0`.
+
+## D1: frozen training-prefix gate
+
+There is exactly one candidate and no hyperparameter grid.  Keep the MT-1
+ResNet-20 architecture, CIFAR-10 45k training split, seed-0 initialization,
+four-to-one traffic intervention, augmentation, batch size 128, learning rate
+0.1, momentum 0.9, decay `1e-4`, and reciprocal KP path unchanged.  Use a
+zero-margin closed-form slow predictor, freeze it, and apply the fast neutral
+projection on every one of the first 352 shuffled task minibatches.  Restore
+the task-loader state after calibration.  Do not evaluate validation or test
+examples.
+
+The candidate passes only if all conditions below hold:
+
+- every recorded loss, signal diagnostic, parameter, optimizer state,
+  BatchNorm statistic, and projection diagnostic remains finite for 352 steps;
+- maximum task minibatch loss is at most 10 and final-32 mean loss is at most
+  2.5;
+- the used-signal/instruction RMS ratio stays within `1e-4` of one;
+- the maximum post-projection neutral/traffic RMS ratio is at most `1e-5`;
+- the maximum absolute post-projection residual-soma slope is at most `1e-5`;
+- every projection uses the current task batch's 72--128 local neutral
+  observations and exactly zero instruction observations;
+- maximum absolute forward and reciprocal-feedback weight is at most 10;
+- maximum absolute forward and reciprocal momentum is at most 50;
+- the 64-observation slow fit has zero stability margin, the slow predictor is
+  frozen during task training, initial traffic calibration is within `1e-5`
+  of ratio 4, and task-loader state restoration is exact; and
+- validation and test evaluation counts are both zero.
+
+A failure closes this paired-neutral projection branch without changing a
+threshold or adding a gain/clipping hyperparameter.  A pass supplies no task
+accuracy claim and cannot change the reviewer score.  It only authorizes a new,
+separately committed validation protocol with an explicit cost boundary.
+