# Residual response-mirror baseline protocol

## Status and method boundary

Estimate-then-average WM passed its frozen capture gate but missed the short
accuracy gate by under one point. Its final feedback/forward cosine plateaued
near 0.81 in the deepest convolutional group despite frequent local probes.
The diagnosed issue is stationary estimator noise: a fresh finite-sample W
estimate is noisy even when Q already equals W.

Residual response mirroring (RRM) is a substantive update-rule change, not a
new cadence/batch/rate branch of failed WM. For each local Gaussian probe it
compares the observed forward child response with Q's predicted response and
updates from their difference. At Q=W, every individual stochastic update is
zero. The observation and update remain separated; the update never reads a
forward parameter. RRM is still inherited local predictive weight estimation,
not SDIL or Harnett-specific novelty.

A mechanics-only synthetic-image pilot was used to bound the mirror-rate grid.
No CIFAR development-prefix alignment or validation/test endpoint was observed
before this protocol and its executable selector were committed.

## RRM-0: mechanics gate

The convolutional smoke suite must verify:

- exact-symmetry response-residual fraction below `1e-14` for every probe;
- exact-symmetry parameter-update RMS below `1e-14`;
- changing all forward parameters after observations are fixed changes the
  feedback update by exactly zero;
- all prior hierarchical and local-gradient checks remain green.

This passes at response-residual fraction `3.08e-17`, update RMS `2.75e-18`,
and forward-parameter independence error exactly zero.

## RRM-1: frozen-forward capture gate

Copy WM-1 exactly: seed-0 ResNet-20, random feedback scale 1, frozen 10k
development prefix, 64-example exact-gradient audit, convolutional mirror batch
1, Gaussian noise standard deviation 1, mirror seed 3000, 20 observations, and
no forward/readout/BatchNorm update. Record fixed HFA and cross
`eta_M in {0.03,0.1,0.3}`. Select early alignment, then all-layer alignment,
then lower rate.

All four records must be finite, and selected RRM must reach early alignment
0.65, all-layer alignment 0.70, mean feedback/forward cosine 0.93, norm ratios
in `[0.5,1.5]`, and zero task-loss queries. No extra observation, rate, batch,
noise, cadence, or layer-specific setting follows a failure.

## RRM-2: short accuracy gate

Only an RRM-1 pass opens two 10k-example, 20-epoch ResNet-20 validation jobs.
Copy WM-2 exactly: hidden LR `{0.03,0.1}`, output LR 0.1, batch 128, cosine
decay, no warmup, momentum 0.9, weight decay `1e-4`, 20 mirror warmup
observations, then one batch-1 observation every 16 task updates. The extra Q
response-prediction convolution is charged explicitly.

Select accuracy, then total MACs/rate. Both records must be finite; selected
RRM must reach 65%, lie within 10 points of matched BP 74.94%, retain early
alignment 0.50, use zero task-loss queries for feedback learning, and cost no
more than 1.15x matched BP MACs. Failure closes RRM.

## RRM-3: conditional full baseline

Only an RRM-2 pass opens one 200-epoch seed-0 validation run, copying selected
settings and the A1 epoch-100/150 step drops on all 45,000 development-training
examples. It must be finite, reach 88%, retain early alignment 0.50, and cost
at most 1.15x BP. It does not authorize test access. A pass instead freezes the
same scalable substrate for a raw-versus-innovation mixed-traffic experiment.

RRM is a baseline throughout. RRM-1/RRM-2 cannot raise the reviewer score;
only a later load-bearing innovation result on top of it can do so.

## Audited RRM-1 outcome

The frozen-forward capture gate passed on source revision `18aff5b`. The full
rate sweep is retained rather than reporting only the selected record:

| method / rate | early alignment | all-layer alignment | feedback/forward cosine | norm range |
|:--|--:|--:|--:|:--|
| fixed HFA | -0.000263 | 0.011279 | 0.001614 | [0.959, 1.055] |
| RRM 0.03 | 0.075016 | 0.235001 | 0.622965 | [0.695, 0.792] |
| RRM 0.10 | **0.665070** | **0.730253** | **0.953680** | **[0.884, 1.018]** |
| RRM 0.30 | 0.003489 | 0.053744 | 0.740396 | [0.999, 30.132] |

The preregistered selector chooses `eta_M=0.1`. All records have finite
reported metrics, and the selected record passes every alignment, cosine,
norm, and zero-query check. Its 20 observations cost `2.4458e9` MACs. The
high-rate instability remains visible and is not used to widen the rate grid.
RRM-2 is therefore opened at the frozen selected rate.

Relative to normalized estimate-then-average mirroring at the same selected
rate and observation count, residual mirroring improves early alignment by
`0.218929`, all-layer alignment by `0.188094`, and feedback/forward cosine by
`0.038200`. This is evidence that removing estimator noise at the fixed point
matters. It remains inherited local predictive weight estimation, so the
strict reviewer score stays `5/10`.

## Audited RRM-2 outcome

The frozen short accuracy gate passed on source revision `6206483`:

| hidden LR | validation accuracy | early alignment | all-layer alignment | feedback/forward cosine |
|--:|--:|--:|--:|--:|
| 0.03 | 62.32% | 0.999604 | 0.999692 | 0.999970 |
| 0.10 | **65.08%** | **0.999128** | **0.999377** | **0.999902** |

The accuracy selector chooses hidden LR 0.1. It clears the 65% gate by only
0.08 points and lies 9.86 points below matched BP at 74.94%, clearing the
within-10-point gate by 0.14 points. Both runs are finite. The selected run
uses zero task-loss queries, 119 mirror events, `2.4256703e13` MACs
(`0.9969x` matched BP), 1.731 GB peak allocated memory, and 93.0 s on a GTX
1080. Every frozen RRM-2 check passes, so RRM-3 opens without changing the
selected mirror rate, cadence, or hidden learning rate.

The nearly exact directional and weight cosines do not erase the short-horizon
accuracy gap. This makes the result more useful as a strict baseline: direction
alone is not a surrogate for trajectory equivalence, especially when feedback
tracks changing forward weights at discrete intervals. Because RRM is inherited
predictive weight estimation, this pass does not change the reviewer score.

## Audited RRM-3 outcome

The sole frozen full run from source revision `8a5c104` fails decisively:

| metric | RRM-3 | frozen requirement/reference |
|:--|--:|--:|
| validation accuracy | **10.00%** | at least 88%; BP is 91.62% |
| validation loss | **NaN** | finite |
| early / all-layer endpoint alignment | 0.877435 / 0.848738 | early at least 0.50 |
| endpoint feedback/forward cosine | 0.999998 | diagnostic only |
| feedback/forward norm range | [1.000157, 1.002447] | diagnostic only |
| MAC ratio to BP | 0.996857x | at most 1.15x |
| task-loss queries for feedback | 0 | exactly 0 |

The task trajectory, not the endpoint tracking snapshot, identifies the
failure. Mean training loss grows from `2.6107` at epoch 1 to a peak of
`1.5808e16` at epoch 92 and remains `9.6549e15` at epoch 100. Stepwise learning
rate drops subsequently reduce it to `8.9503e13` by epoch 200 without restoring
the classifier. Thus an almost perfect final Q/W cosine and controlled norm do
not certify that feedback tracked the fast task-weight trajectory when the
destructive updates occurred.

The ideal recurrence in `THEORY.md`,
`E_(m+1)=(1-eta_M)E_m-U_m`, explains the audit: intermittent tracking error is
driven by cumulative forward motion between mirror events, but a quiet or
low-rate tail can erase endpoint error geometrically. RRM-3 therefore closes
without cadence/rate rescue, opens no innovation experiment, touches no test
seed, and leaves the strict reviewer score at 5/10.