protocol: freeze stability-margin training screen

1 files changed, 66 insertions, 0 deletions

diff --git a/STABLE_INNOVATION.md b/STABLE_INNOVATION.md
new file mode 100644
index 0000000..d499dc5
--- /dev/null
+++ b/STABLE_INNOVATION.md
@@ -0,0 +1,66 @@
+# Stability-calibrated innovation development protocol
+
+## Status and claim boundary
+
+This is a new post-MT-1 development branch, not a reinterpretation or rescue
+of the frozen mixed-traffic result. MT-1 remains failed, MT-2/MT-3 remain
+untouched, and no result here changes the reviewer score. The sole question is
+whether a local one-sided neutral-regression certificate can remove the
+multiplicative instability diagnosed in `THEORY.md` before any new validation
+endpoint is designed.
+
+The four-to-one traffic intervention, ResNet-20 architecture, seed-0 forward
+initialization, 45k training split, augmentation, batch size 128, LR 0.1,
+momentum 0.9, decay `1e-4`, and reciprocal KP path remain unchanged. The
+predictor is fitted once from the same 64-example instruction-off calibration
+prefix by per-cell affine least squares, then frozen throughout task training.
+No validation or test example is evaluated.
+
+For fitted slope `p`, use the stability upper margin
+
+```text
+p_stable = p + delta (1 + |p|).
+```
+
+Thus the measured neutral residual-soma slope is one-sided nonpositive rather
+than a point estimate with uncontrolled sign. This adds a small local
+anti-Hebbian component when the traffic fit is exact. It does not lower the
+traffic ratio or inspect a forward/feedback weight.
+
+## S0: frozen training-prefix stability grid
+
+The already observed diagnostic candidates `delta=0` and
+`delta=sqrt(float32 epsilon)` are ineligible: both retain finite task-active
+parameters for one epoch but produce enormous loss/parameter growth and
+nonfinite BatchNorm running variances. Before evaluating another margin, freeze
+the only remaining grid to
+
+```text
+delta in {0.001, 0.003, 0.01, 0.03}.
+```
+
+Run exactly the first 352 shuffled/augmented task minibatches for every value.
+All four records must share one clean source revision. The loader state is
+restored after the neutral fit. Each record must report zero validation and
+test evaluations.
+
+A candidate is eligible only if all of the following hold:
+
+- all recorded losses, signal diagnostics, parameters, optimizer states, and
+  BatchNorm running statistics remain finite for 352 steps;
+- maximum task minibatch loss is at most 10 and the final 32-step mean is at
+  most 2.5 (clean KP epoch-1 mean is 1.8315);
+- the maximum used-innovation/instruction RMS ratio is at most 2;
+- maximum absolute forward and reciprocal-feedback weight is at most 10;
+- maximum absolute forward and reciprocal momentum is at most 50;
+- the closed-form fit uses 64 observations, has maximum positive neutral
+  residual-soma slope at most `1e-7`, and every predictor update flag during
+  task training is false;
+- initial traffic calibration remains within `1e-5` of ratio 4, task loader
+  state restoration is exact, and no held-out evaluation occurs.
+
+Select the smallest eligible margin. If no margin is eligible, close this
+one-sided predictor branch without extending the grid. A pass only authorizes
+writing a separately frozen validation protocol with a new evaluation
+boundary; it supplies no accuracy claim and leaves the paper score at 5/10.
+