protocol: freeze KP mixed-traffic innovation gates

1 files changed, 145 insertions, 0 deletions

diff --git a/MIXED_TRAFFIC.md b/MIXED_TRAFFIC.md
new file mode 100644
index 0000000..0bc0ee2
--- /dev/null
+++ b/MIXED_TRAFFIC.md
@@ -0,0 +1,145 @@
+# KP mixed-traffic innovation protocol
+
+## Question and attribution boundary
+
+This protocol asks whether the Harnett-specific somato-dendritic innovation is
+load-bearing on a standard ResNet once a strong non-backpropagating credit path
+is available. The reciprocal Kolen--Pollack (KP) pathway is inherited from
+Akrout et al. and receives no SDIL novelty credit. The candidate contribution
+is only the neutral-period subtraction used in the mixed apical compartment.
+
+At every hidden population, first compute the hierarchical KP instruction
+`s_l`. Add fixed soma-predictable traffic
+
+```text
+t_l = g_l B_l h_l,
+a_l = s_l + t_l,
+a_hat_l = P_l h_l + b_l,
+r_l = a_l - a_hat_l.
+```
+
+`B_l` is a per-unit log-normal coefficient drawn once from traffic seed 4000.
+The positive gain `g_l` is fixed once at initialization using the first 64
+unaugmented development-training examples so that every layer has
+`RMS(t_l)/RMS(s_l)=4`. Labels enter only this one training-prefix scale
+calibration through `s_l`; no validation endpoint is used. The target ratio is
+not retuned across signals, epochs, or seeds.
+
+The diagonal affine predictor is trained only with instruction switched off.
+All three conditions receive exactly 20 neutral warmup minibatches at
+`eta_P=0.1` and one neutral predictor update every 16 task minibatches. Raw and
+norm-matched controls pay the same predictor-learning cost even though raw
+does not consume its prediction. The three used signals are
+
+```text
+raw:        u_l = a_l,
+matched:    u_l = ||r_l||_example / ||a_l||_example * a_l,
+innovation: u_l = r_l.
+```
+
+Norms are taken over every non-batch coordinate separately for each example
+and layer. Consequently the matched control has the innovation's update-signal
+norm but the raw signal's direction at the same network state. Every condition
+uses its own used signal for both the forward local correlation and the
+independently recomputed reciprocal KP correlation; matched optimizer and
+decay dynamics are otherwise unchanged.
+
+This is a controlled predictable-traffic test, not the Oral-B natural-traffic
+claim. The broad endogenous top-down traffic and desired-velocity branches
+remain failed and are not reopened by this protocol.
+
+## MT-0: mechanics and cost gate
+
+Before any task endpoint, the float64/CPU smoke suite must establish all of:
+
+- setting traffic gain to zero makes raw, matched, and innovation exactly the
+  clean KP instruction when `P=b=0`;
+- setting `P_l=g_l B_l, b_l=0` removes the predictable traffic exactly;
+- layerwise gain calibration realizes ratio 4 to relative error below `1e-12`;
+- matched raw has the innovation norm to relative error below `1e-12` and
+  preserves raw cosine direction to absolute error below `1e-12`;
+- the predictor update depends only on local soma and instruction-off apical
+  traffic, not labels, output errors, or forward/feedback weights;
+- independently recomputed forward and reciprocal correlations remain equal
+  under every used signal, and no parameter acquires an autograd graph;
+- affine MACs, predictor/traffic elementwise operations, loss queries, memory,
+  and wall time are reported separately. All three conditions execute the
+  same calibration and predictor schedule.
+
+Failure closes this branch. Passing mechanics does not change the reviewer
+score.
+
+## MT-1: frozen 20-epoch load-bearing screen
+
+This screen is conditional on a KP-2 pass. Run exactly raw, matched, and
+innovation at seed 0 on the frozen 45,000/5,000 CIFAR-10 split. Copy KP-1:
+ResNet-20 width 16, batch 128, standard augmentation, BatchNorm, hidden/output
+LR 0.1, epoch-100/150 step drops (hence constant in this screen), no optimizer
+warmup, momentum 0.9, and weight decay `1e-4`. Use the traffic and predictor
+settings above and a 32-example final alignment audit. There is no traffic,
+predictor, learning-rate, decay, or signal-rule grid.
+
+MT-1 passes only if all of the following hold:
+
+- all three records, epoch losses, tracking trajectories, and diagnostics are
+  finite, clean, and validation-only;
+- innovation accuracy is at least 75% and no more than 5 points below the
+  frozen clean KP-1 accuracy of 82.66%;
+- innovation exceeds raw by at least 5 accuracy points and norm-matched raw by
+  at least 3 points;
+- innovation's used-signal early-third alignment is at least 0.70, and on the
+  same final innovation network its innovation alignment exceeds raw-apical
+  alignment by at least 0.15;
+- innovation's final feedback/forward cosine is at least 0.80 and its
+  epoch-11--20 mean is at least 0.70;
+- every initial layer ratio is within `1e-5` of 4, the post-warmup predictor
+  traffic-residual RMS ratio is at most 0.25, and matched norm relative error
+  is at most `1e-6`;
+- feedback learning uses zero task-loss queries and each condition's affine
+  MAC estimate is at most `1.40x` the frozen epoch-20 BP reference.
+
+The accuracy gaps and same-state alignment rotation are conjunctive: clipping
+an unstable raw update is not enough. MT-1 failure closes the branch without a
+weaker traffic ratio. A pass opens MT-2 but leaves the reviewer score at 5/10.
+
+## MT-2: frozen full validation gate
+
+Copy MT-1 exactly for 200 epochs. The three jobs remain seed 0 and do not touch
+the test set. In addition to finite/clean provenance, exact ratio calibration,
+zero queries, matched norms, and the `1.40x` BP MAC ceiling, require:
+
+- innovation validation accuracy at least 88% and within 3 points of KP-2;
+- innovation gains of at least 5 points over raw and 3 over matched raw;
+- innovation used-signal early alignment at least 0.80 and a same-network
+  advantage of at least 0.15 over raw-apical alignment;
+- final and epoch-151--200 mean feedback/forward cosine both at least 0.95;
+- post-training traffic-residual RMS ratio at most 0.05.
+
+MT-2 failure closes the confirmation panel. A pass is strong development
+evidence but one seed is insufficient for acceptance, so the formal score
+remains 5/10.
+
+## MT-3: untouched five-seed test confirmation
+
+Only an MT-2 pass opens seeds 10--14. For each seed, pair an independently
+drawn traffic field with seed `5000 + model_seed`. Freeze the chosen method and
+run clean KP, raw, matched, and innovation after refitting on all 50,000
+CIFAR-10 training examples. Evaluate the test set exactly once per run; no
+validation endpoint, seed deletion, or replacement is permitted.
+
+The confirmation passes only if all 20 records are finite and clean, mean
+innovation test accuracy is at least 88%, and:
+
+- the mean paired innovation deficit to clean KP is at most 3 points;
+- mean paired innovation gains are at least 5 points over raw and 3 over
+  matched raw;
+- the one-sided 95% lower confidence bound of each paired innovation gain is
+  above zero;
+- at least four of five paired gains are positive against each control;
+- all mechanism, query, and cost invariants from MT-2 remain satisfied.
+
+A complete pass raises the strict reviewer estimate from 5/10 to 6/10 (weak
+accept): it would establish a Harnett-specific, norm-controlled innovation
+effect on a standard ResNet across untouched network and traffic draws. It
+would not establish natural biological traffic or broad architecture scaling;
+those remain Oral-B and Oral-A work respectively.