# Normalized response-mirror baseline protocol

## Status and attribution

This protocol was frozen before observing any CIFAR-prefix response-mirror
alignment or accuracy endpoint. A mechanics-only pilot on synthetic Gaussian
images was used to choose the bounded mirror-rate grid; it did not load the
frozen CIFAR development split or evaluate validation/test accuracy.

The method is an inherited strong baseline, not SDIL novelty. It is a
normalized, bias-blocked local system-identification variant of Akrout et al.'s
[weight mirror](https://proceedings.neurips.cc/paper_files/paper/2019/file/f387624df552cea2f369918c5e1e12bc-Paper.pdf).
Gaussian parent probes pass through ordinary local forward synapses and a
separate update receives only the probe and child preactivation response. The
local correlation estimates the forward kernel; Q/R track that estimate by an
exponential delta rule. The task-error path remains the audited hierarchical
FA residual DAG.

Any BP-like scale obtained here belongs to weight estimation. It can raise the
bar SDIL must beat but cannot raise SDIL novelty or rescue failed Oral-A-v1.
Only a later, separately frozen demonstration that somato-dendritic innovation
is load-bearing on top of this path can improve the paper-level claim.

## WM-0: mechanics gate

Before endpoints, the smoke suite must show:

- one 16-probe local observation has mean feedback/forward cosine above 0.985
  and minimum cosine above 0.95 on a deterministic tiny ResNet;
- every feedback/forward norm ratio is in `[0.90, 1.10]`;
- after observations are fixed, changing all forward parameters changes the
  feedback update by exactly zero;
- all hierarchical-gradient and convolutional local-update audits remain
  green.

The gate passed at mean/min cosine `0.988701/0.966066`, norm ratios
`[0.953729,1.044368]`, and update independence error exactly zero.

## WM-1: frozen-forward CIFAR causal-capture gate

Use seed-0 ResNet-20, random feedback scale 1, the first 10,000 examples of the
frozen development split, batch 128 for the diagnostic graph, and the fixed
64-example training-prefix exact-gradient audit. Forward weights, readout,
BatchNorm state, and affine parameters remain bitwise fixed.

Record one fixed-HFA reference. For each `eta_M in {0.03, 0.1, 0.3}`, run 20
local mirror observations with convolutional probe batch 1, Gaussian noise
standard deviation 1, and mirror seed 3000. No task loss or label is consumed
by mirroring; the final alignment audit alone uses labels. Select maximum
early-third alignment, then all-layer alignment, then lower rate.

WM-1 passes only if all four records are finite and selected WM:

1. reaches early-third teaching alignment at least `0.40`;
2. reaches all-layer alignment at least `0.50`;
3. reaches mean feedback/forward cosine at least `0.85`;
4. keeps every feedback/forward norm ratio in `[0.5, 1.5]`;
5. records zero logical task-loss queries during mirror learning.

No mirror batch, step count, noise scale, rate, covariance estimator, or
per-layer schedule is added after observation.

## WM-2: bounded short accuracy gate

Only a WM-1 pass opens a matched 10k-example, 20-epoch ResNet-20 validation
screen. Copy A2b data, augmentation, batch 128, cosine schedule, no warmup,
momentum 0.9, weight decay `1e-4`, and output LR 0.1. Cross hidden LR
`{0.03,0.1}`. Use the selected mirror rate, 20 mirror warmup observations,
then one batch-1 mirror observation every 16 task updates. Select validation
accuracy, then lower total MACs/rate.

Advance only if both runs are finite, selected WM reaches at least 65%, stays
within 10 points of BP's matched 74.94%, retains early alignment at least 0.30,
uses zero task-loss queries for feedback learning, and costs no more than 1.15x
matched BP estimated MACs. Failure closes the baseline without tuning.

## WM-3: full validation baseline

Only a WM-2 pass opens one 200-epoch seed-0 validation run on all 45,000
development-training examples. Copy its selected hidden rate, mirror settings,
output LR 0.1, and the A1 step drops at epochs 100/150. No recovery branch is
allowed. A finite endpoint at least 88% with early alignment at least 0.30 and
cost at most 1.15x BP is considered a strong inherited baseline. It does not
authorize test access; it instead opens design of a hierarchical innovation
intervention on the same frozen recipe.

## Audited WM-1 outcome (2026-07-22)

All four records are finite and every frozen capture check passes. The selected
rate is `eta_M=0.1`:

| rule | early alignment | all-layer alignment | feedback/forward cosine | norm range |
|:--|--:|--:|--:|:--|
| fixed HFA | -0.000263 | 0.011279 | 0.001614 | [0.959, 1.055] |
| WM 0.03 | 0.075263 | 0.235197 | 0.625186 | [0.697, 0.787] |
| WM 0.10 | **0.446141** | **0.542159** | **0.915480** | **[0.889, 1.157]** |
| WM 0.30 | 0.112177 | 0.228114 | 0.853343 | [0.999, 1.697] |

Each learned record uses 20 local observations, zero logical task-loss queries,
and estimated work `1.6305e9` MACs. WM-1 status is `passed`; WM-2 opens with
the predeclared `eta_M=0.1`. No validation accuracy was used for this selection
and CIFAR-10 test remains untouched.

## Audited WM-2 outcome (2026-07-22)

Both trajectories are finite and retain strong feedback credit:

| hidden LR | validation accuracy | early alignment | all-layer alignment | MACs / BP |
|--:|--:|--:|--:|--:|
| 0.03 | 62.68% | 0.827404 | 0.825811 | 0.9968x |
| 0.10 | **64.04%** | **0.939256** | **0.916222** | **0.9968x** |

Selected LR 0.1 improves on fixed HFA by 20.52 points and uses zero task-loss
queries for feedback learning. It passes finite, alignment, query, and cost
checks. However, it misses the 65% threshold by 0.96 points and remains 10.90
points below matched BP, missing the within-10-point gate by 0.90 points.

WM-2 status is `failed`. WM-3 is not launched, no cadence/batch/rate tuning is
allowed, and test remains untouched. The high `0.939` hidden-credit cosine but
11-point accuracy gap also warns that directional alignment alone does not
guarantee matched finite-horizon optimization.