experiment: freeze KP teaching-signal ablation

1 files changed, 83 insertions, 0 deletions

diff --git a/KP_TEACHING_SIGNAL_ABLATION.md b/KP_TEACHING_SIGNAL_ABLATION.md
new file mode 100644
index 0000000..7b1207e
--- /dev/null
+++ b/KP_TEACHING_SIGNAL_ABLATION.md
@@ -0,0 +1,83 @@
+# KP teaching-signal ablation
+
+## Question
+
+The dynamic SDIL image model and clean KP use the same reciprocal KP credit
+transport.  Their relevant local signals are
+
+```
+clean KP:       u = s
+raw KP:         u = a = s + t
+matched raw KP: u = ||r|| / ||a|| * a
+dynamic SDIL:   u = r = s + t - prediction(t).
+```
+
+Here `s` is the ordinary clean KP instruction and `t` is four-RMS
+soma-predictable apical traffic.  This experiment therefore changes neither
+the forward architecture nor the feedback-learning substrate.  It asks
+whether the subtractive innovation operation is necessary to recover useful
+credit when the same KP substrate receives mixed apical traffic.
+
+Raw and norm-matched KP execute the same frozen 64-example slow fit and the
+same instruction-off fast neutral projection on every ordinary minibatch.
+They report the projection but do not apply its subtractive direction.  Raw
+uses the mixed apical vector unchanged.  Matched raw preserves that vector's
+direction and changes only its per-example norm to the projected innovation
+norm.  Thus a matched-raw deficit cannot be attributed to a larger update
+norm, and neither deficit can be attributed to a cheaper observation budget
+for SDIL.
+
+This is a conditional robustness intervention, not a claim that dynamic SDIL
+outperforms clean KP when no nuisance traffic is present.
+
+## KTS-0: mechanics
+
+Before any endpoint launch, the CPU mechanics test must establish:
+
+- raw with the sham projection is bitwise equal to `s + t`;
+- matched raw has projected innovation's per-example norm to relative error
+  below `1e-12` and raw's direction to error below `1e-12`;
+- raw, matched, and innovation receive the identical instruction-off
+  projection report;
+- the projection observes zero task-instruction examples and changes no slow
+  predictor parameter;
+- all three rules form their local KP forward and reciprocal correlations from
+  the selected used vector without reverse-mode differentiation.
+
+## KTS-1: frozen validation endpoint
+
+KTS-1 inherits the passed
+`results/kp_dynamic_projection_full_gate.json` endpoint without rerunning or
+selecting it: clean KP is 91.26% and dynamic SDIL is 91.18% on the frozen
+45,000/5,000 CIFAR-10 split.  The only new endpoints are `raw` and `matched`.
+Use ResNet-20, width 16, model/data-loader/traffic seed 0/0/4000, batch 128,
+200 epochs, augmentation, batch normalization, SGD momentum 0.9, weight decay
+`1e-4`, learning rate 0.1 with 0.1 drops at epochs 100 and 150, traffic ratio
+4, one closed-form 64-example predictor fit, a frozen predictor thereafter,
+and the fast instruction-off projection on every task batch.  Evaluate the
+validation split exactly once at the endpoint and never evaluate test.
+
+Both jobs must share one clean tracked source revision containing this
+protocol, its runner, analyzer, mechanics test, and the sham-control
+implementation.  No endpoint may be deleted, replaced, or relaunched under
+the same name.  A nonfinite trajectory is retained as a terminal control
+collapse, not repaired by lowering the learning rate or traffic strength.
+
+The mechanism advantage passes only if every provenance, split, argument,
+traffic, observation, query, projection, norm-control, and cost invariant
+passes and dynamic SDIL exceeds raw by at least 5 percentage points and
+matched raw by at least 3 percentage points.  A nonfinite control counts as a
+performance/stability failure of that control, while the raw record and its
+failure location remain part of the evidence.  A finite control must have a
+valid endpoint accuracy.  Every job must use zero task-loss queries, one
+64-example slow fit, one instruction-off projection observation per ordinary
+example, no predictor updates during task learning, at most 1.34 times the
+frozen BP affine-MAC estimate, and separately reported positive elementwise
+work.
+
+Passing KTS-1 establishes that innovation subtraction, rather than KP
+transport alone or signal-norm rescaling, is necessary under the specified
+soma-predictable traffic.  It does not establish superiority to Dual Prop,
+BP, or clean KP in the clean setting.  Multi-seed and cross-architecture
+confirmation remain subsequent gates.
+