docs: add audited dynamic stability caption

1 files changed, 20 insertions, 0 deletions

diff --git a/results/figs/figureS_dynamic_stability_caption.md b/results/figs/figureS_dynamic_stability_caption.md
new file mode 100644
index 0000000..08613e8
--- /dev/null
+++ b/results/figs/figureS_dynamic_stability_caption.md
@@ -0,0 +1,20 @@
+**Figure S — Dynamic neutral projection repairs the multiplicative
+mixed-traffic failure.** **a,** Training-only batch losses for the frozen D1
+dynamic controller and the earlier closed-form/frozen-predictor controller
+lesion. The lesion curve is a labeled pre-grid diagnostic rather than a
+preregistered endpoint; it visualizes the failure that motivated D1. Both
+records use the same ResNet-20 family, four-to-one soma-coupled traffic,
+momentum, learning rate, and 352-step boundary. **b,** Before each D1 update,
+the neutral residual remaining after the slow predictor grows with the task
+state. A fast affine projection based only on paired instruction-off local
+soma/apical observations holds the post-projection residual below
+`3.03e-8` of traffic RMS. The fast coefficient fit observes no task
+instruction and leaves the slow predictor frozen. **c,** The separately
+frozen 20-epoch D2 endpoint reaches `83.58%` validation accuracy, versus
+`82.66%` for clean reciprocal KP and `10%` for each raw, norm-matched, and
+original frozen-innovation MT-1 condition. The MT-1 bars do not receive the
+new controller and are retained as historical failed controls. D2 uses one
+seed, one final validation evaluation, zero test evaluations, zero task-loss
+queries, and `1.326x` the matched BP affine-MAC estimate; the paired neutral
+microphase and 4.070e12 elementwise operations are reported separately.
+