# Dynamic neutral-projection development protocol

## Claim boundary

This is a new post-S0 development branch.  It does not reinterpret or rescue
MT-1: MT-1 remains failed, MT-2/MT-3 remain sealed, and the reviewer score
remains 5/10.  S0 ruled out extending a fixed one-sided predictor margin.  The
new question is whether a fast *local and instruction-off* projection can keep
the multiplicative neutral-residual mode inside its two-sided stability window
as somatic statistics change.

The slow predictor is still fitted exactly once on the original 64-example
neutral calibration prefix and is frozen during task learning.  Before each
task update, every cell receives a paired instruction-off observation of its
current soma `h` and ordinary apical traffic `a0`.  From the neutral residual

```text
e0 = a0 - (P h + b),
```

it forms the current-batch local affine projection

```text
q = Cov(e0, h) / Var(h),
e_stable = (e0 - mean(e0)) - q (h - mean(h)).
```

The plasticity signal is `r = s + e_stable`.  The coefficient fit sees zero
task-instruction observations; it reads no label, loss, forward weight,
feedback weight, or downstream state, and it does not modify the slow
predictor.  For the diagonal affine traffic used by frozen MT-1, the controller
nulls the current empirical residual-coupling coefficient.  With weight decay,
the scalar homogeneous mode is therefore placed near `k=-lambda`, inside both
Jury boundaries, instead of at a fixed negative coefficient whose product with
an evolving covariance can cross the lower boundary.

This controller requires one paired neutral apical microphase per task batch.
Its elementwise work and observation count must be charged separately in any
later endpoint comparison.  It uses zero task-loss perturbation queries and no
reverse-mode differentiation.

## D0: mechanics (completed before task-data evaluation)

The implementation must prove on synthetic float64 tensors that:

- an inaccurate frozen slow predictor leaves a nonzero affine neutral mode;
- fast projection reduces the post-projection neutral/traffic RMS ratio and
  residual-soma slope below `1e-14` for diagonal affine traffic;
- projected innovation equals clean instruction below `1e-14`;
- the projection consumes zero instruction observations and leaves every slow
  predictor parameter bitwise unchanged;
- forward and reciprocal correlations remain independently computed; and
- the extra elementwise work is exposed explicitly.

These checks passed at clean revision `eeb50b0`.

## D1: frozen training-prefix gate

There is exactly one candidate and no hyperparameter grid.  Keep the MT-1
ResNet-20 architecture, CIFAR-10 45k training split, seed-0 initialization,
four-to-one traffic intervention, augmentation, batch size 128, learning rate
0.1, momentum 0.9, decay `1e-4`, and reciprocal KP path unchanged.  Use a
zero-margin closed-form slow predictor, freeze it, and apply the fast neutral
projection on every one of the first 352 shuffled task minibatches.  Restore
the task-loader state after calibration.  Do not evaluate validation or test
examples.

The candidate passes only if all conditions below hold:

- every recorded loss, signal diagnostic, parameter, optimizer state,
  BatchNorm statistic, and projection diagnostic remains finite for 352 steps;
- maximum task minibatch loss is at most 10 and final-32 mean loss is at most
  2.5;
- the used-signal/instruction RMS ratio stays within `1e-4` of one;
- the maximum post-projection neutral/traffic RMS ratio is at most `1e-5`;
- the maximum absolute post-projection residual-soma slope is at most `1e-5`;
- every projection uses the current task batch's 72--128 local neutral
  observations and exactly zero instruction observations;
- maximum absolute forward and reciprocal-feedback weight is at most 10;
- maximum absolute forward and reciprocal momentum is at most 50;
- the 64-observation slow fit has zero stability margin, the slow predictor is
  frozen during task training, initial traffic calibration is within `1e-5`
  of ratio 4, and task-loader state restoration is exact; and
- validation and test evaluation counts are both zero.

A failure closes this paired-neutral projection branch without changing a
threshold or adding a gain/clipping hyperparameter.  A pass supplies no task
accuracy claim and cannot change the reviewer score.  It only authorizes a new,
separately committed validation protocol with an explicit cost boundary.

## Audited D1 outcome

D1 passes all frozen checks at clean revision `8b28522`.  All 352 updates and
states remain finite; maximum minibatch loss is `2.9281` and the final-32 mean
is `1.6293`.  The uncontrolled neutral residual reaches `0.005833` of traffic
RMS and the fast correction slope reaches `0.02968`, so the controller is not
a numerical no-op.  Projection reduces the worst neutral/traffic RMS ratio to
`3.03e-8` and the worst absolute residual-soma slope to `1.52e-8`.  Maximum
forward/feedback weights are `1.181/0.833` and maximum momenta are
`0.600/0.364`.  No held-out example was evaluated and the score remains 5/10.

## D2: frozen short validation endpoint

D1 opens exactly one 20-epoch seed-0 validation run.  Use the same ResNet-20,
45k/5k CIFAR-10 split, augmentation, batch size 128, constant learning rate
0.1 (the inherited KP short schedule), momentum 0.9, decay `1e-4`, four-to-one
traffic, reciprocal KP path, zero-margin 64-example closed-form slow fit, and
fast neutral projection on every task minibatch.  Evaluate the validation set
once after training, probe alignment on 32 training examples only after the
endpoint is fixed, and never evaluate test.

D2 passes only if:

- the final record, all 20 epoch records, feedback trajectories, signal
  trajectories, projection certificates, and diagnostics are finite;
- validation accuracy is at least 80%, within three points of the frozen clean
  KP short endpoint (`82.66%`), at least 70 points above each failed MT-1 raw
  and norm-matched endpoint, and final loss is finite;
- early-third used-signal alignment is at least 0.70, final feedback/forward
  cosine is at least 0.80, and the epoch-11--20 mean feedback cosine is at
  least 0.70;
- every epoch keeps used/instruction RMS within `1e-4` of one, maximum
  post-projection neutral/traffic RMS at most `1e-5`, maximum absolute
  post-projection soma slope at most `1e-5`, and zero task-instruction
  observations in the fast fit;
- exactly one closed-form 64-example slow fit occurs, task predictor cadence is
  zero, neutral-projection observation count equals ordinary training examples,
  initial traffic ratio error is at most `1e-5`, and task-loss queries are zero;
- total MACs are at most `1.34x` the frozen 20-epoch BP record, all additional
  elementwise work is reported, and peak allocated memory is at most 2.5 GiB;
  and
- there is exactly one validation evaluation and zero test evaluations.

D2 failure closes the branch.  A pass remains a one-seed short validation
result and leaves the paper score at 5/10, but opens the already specified D3
full validation endpoint below.

## Audited D2 outcome

D2 passes every frozen check at clean revision `f0fb6c7`.  The sole endpoint
reaches `83.58%` validation accuracy versus `82.66%` for clean KP and `10%` for
both failed MT-1 controls.  Early-third alignment is `0.8838`, final and
epoch-11--20 feedback cosines are `0.9041` and `0.8606`, and all trajectories
are finite.  The maximum unprojected neutral/traffic ratio grows to `0.00956`,
while the dynamic remainder stays below `3.04e-8` with maximum soma slope
`1.44e-8`.  The run uses zero task-loss queries, `1.3261x` BP MACs, 2.165 GB
peak allocated memory, exactly one validation evaluation, and zero test
evaluations.  The score remains 5/10 because this is a one-seed short endpoint.

## D3: conditionally frozen full validation endpoint

If and only if D2 passes, run one seed-0 200-epoch validation record with the
same method and data boundary.  The only schedule change is the inherited KP-2
step schedule: LR 0.1, drops by 0.1 at epochs 100 and 150, no warmup.  Evaluate
validation exactly once at the end and test never.  D3 passes only if every
D2 mechanics/cost/query certificate remains satisfied (using the corresponding
200-epoch BP cost), validation accuracy is at least 89%, lies within 1.5 points
of BP (`91.62%`) and clean KP (`91.26%`), final early alignment is at least
0.90, final feedback cosine is at least 0.98, the final-50-epoch mean feedback
cosine is at least 0.97, and all trajectories remain finite.

A D3 pass would establish a frozen full standard-ResNet endpoint and make a
5-to-6 reviewer-score update eligible, subject to an adversarial audit.  The
audit raises the score to 6 only if *every* stated D3 condition passes; code or
presentation changes cannot substitute for a failed endpoint.  It
would then open a separately frozen independent multi-seed test confirmation;
no confirmation seed or test endpoint may be touched before that protocol is
committed.  A D3 failure leaves the score at 5 and closes this branch.

## Audited D3 outcome

D3 passes all 19 frozen checks at source revision `d945c42`.  The sole
endpoint reaches `91.18%` validation accuracy, only 0.44 points below BP
(`91.62%`) and 0.08 points below clean KP (`91.26%`).  Early-third alignment
is `0.999353`, final feedback/forward cosine is `0.999663`, and the epoch
151--200 mean is also `0.999663`.  The maximum unprojected neutral/traffic
ratio is `0.008474`; the projected ratio and soma slope remain below
`3.04e-8` and `2.13e-8`, with zero task-instruction observations in the fast
fit.  The run uses zero task-loss queries, `1.3261x` BP MACs plus separately
reported elementwise work, 2.165 GB peak allocated memory, one validation
evaluation, and zero test evaluations.  Under the predeclared score rule this
raises the strict reviewer score from 5 to 6 and opens, but does not complete,
D4.

## D4: conditionally frozen paired test confirmation

The D4 protocol and executable gate are committed before any D3 endpoint is
observed, but D4 remains inaccessible unless the audited D3 gate passes.  On a
pass, pair clean KP and dynamic innovation at model/data-loader seeds 10--14.
Each independently drawn dynamic traffic field uses seed `5000 + model_seed`.
Train on all 50,000 CIFAR-10 training examples for 200 epochs with the exact D3
architecture, optimizer, augmentation, and schedule.  Evaluate CIFAR-10 test
exactly once at the endpoint of each of the ten runs; perform no validation or
intermediate test evaluation.  All runs must share one clean source revision,
and no failed seed may be deleted or replaced.

D4 passes only if all ten trajectories and audited values are finite, mean
dynamic test accuracy is at least 89%, every dynamic seed reaches at least 87%,
the mean paired deficit to clean KP is at most 1.5 points, the one-sided 95%
upper confidence bound on that deficit is at most 2.5 points, and at least four
of five dynamic seeds lie within 2 points of paired clean KP.  Mean early-third
credit alignment must be at least 0.90 and every dynamic seed must individually
retain early alignment at least 0.85, final feedback cosine at least 0.98,
final-50-epoch mean feedback cosine at least 0.97, all D3 projection/leakage
tolerances, zero task-loss queries, at most `1.34x` the corresponding 50k BP MAC
estimate, separately reported positive elementwise work, and at most 2.5 GiB
peak allocated memory.

A complete D4 pass raises the strict reviewer score from 6 to 7 and establishes
the accept bar.  It does not establish the oral claim: standard-network depth
scaling and the broader biological signature panel remain separate gates.  A
D4 failure keeps the D3-supported score at 6 and closes this confirmation
without threshold repair.

The executable audit accepts exactly the ten named frozen records and rejects
missing or extra JSON cells. Once its deterministic gate exists, later audits
may only verify exact equality with that record; they cannot overwrite it.
The audit binds both the exact D3 prerequisite-gate digest and the detached
`0008f2c` source revision used to launch all ten records; merely sharing an
arbitrary clean revision is insufficient.
Every record must also report an NVIDIA GTX 1080 with physical visibility
restricted to the authorized timan107 GPU 5 or 7; device identity is an audit
condition rather than an inferred property of the launch script.

### D4 realized result

D4 passes on 2026-07-23 without seed replacement, threshold changes, or
intermediate endpoint access.  Dynamic innovation reaches `91.584%` mean test
accuracy across seeds 10--14 versus clean KP's `91.388%`; the paired
clean-minus-dynamic differences are
`[+0.06, -0.03, -0.78, -0.01, -0.22]` percentage points, with mean `-0.196`
and one-sided 95% upper bound `+0.131` points.  Every dynamic seed reaches at
least `91.51%`, and mean early-third alignment is `0.999687`.

All ten trajectories and audited metrics are finite.  The five dynamic records
also pass the frozen feedback, projection, instruction-leakage, query, MAC,
elementwise-work, peak-memory, hardware, source, data-split, and evaluation
checks with no invariant failure.  The immutable gate is
`results/kp_dynamic_projection_confirmation_gate.json`, and the ten raw
records are in `results/kp_dynamic_projection_confirmation/`.  This realizes
the predeclared 6-to-7 score change and opens oral-B R1; it does not establish
positive utility from added standard-network depth.