path: root/MIXED_TRAFFIC.md

# 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.
```

The scale was chosen from a pre-endpoint signal calculation, not an accuracy
pilot. If instruction and traffic are uncorrelated and the RMS ratio is
`rho=4`, the instruction contributes only `1/(1+rho^2)=1/17=5.88%` of raw
signal power. Ideal normalized LMS leaves coefficient fraction
`(1-0.1)^20=0.12158` after warmup, so residual traffic has only
`4*0.12158=0.4863` times the instruction RMS and innovation's instruction
fraction rises to `1/(1+0.4863^2)=80.88%`. Ongoing neutral updates reduce it
further. These are falsifiable scale predictions under the stated
orthogonality idealization, not guaranteed accuracy outcomes.

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-0 passes. CPU float64 errors are zero or at most `8.88e-16`; the exact
predictor error is `1.39e-17`. A task-data-free synthetic ResNet-20 batch-128
audit of the most expensive matched path on a GTX 1080 remains finite and peaks
at 0.857 GB allocated / 0.906 GB reserved after reset, with float32 layer-ratio
error `4.77e-7`. The machine-readable record is
`results/kp_innovation_mechanics.json`. No CIFAR endpoint was evaluated.

## 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-1 fails. All three clean-revision records become nonfinite during epoch 1
and end at 10.00% validation accuracy with NaN loss. Thus innovation does not
separate from either raw control at the frozen four-to-one intervention. The
initial layer-ratio error is only `4.77e-7`, neutral warmup leaves a traffic
residual ratio of `0.139942`, all conditions use zero task-loss queries, and
each costs `1.3271x` the epoch-20 BP MAC reference. Calibration, warmup, query,
and cost invariants therefore pass, but they do not rescue the failed finite,
accuracy, alignment, tracking, or norm checks. MT-2 and MT-3 remain untouched;
the branch closes without changing traffic strength or predictor schedule.

A post-failure training-only audit performs no validation or test evaluation
and stops at the first training-active nonfinite tensor. The first failures are
the stem forward weight and its momentum at task step 23 for raw, 69 for
norm-matched raw, and 74 for innovation. The predictor slope and reciprocal
feedback weights remain finite at those points. Norm matching and innovation
therefore delay the same multiplicative local instability rather than remove
it. `THEORY.md` derives the linearized residual-coupling term `D W C` and shows
why residual RMS alone cannot certify stability.

## 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.