# ICLR 2027 replan: local learning under structured bias

## Decision

Keep the title **Learning from the Unexpected: Somato-Dendritic Innovations
for Local Credit Assignment**, but replace the old clean-scaling claim with a
more defensible claim:

> Two-state local learners can average away zero-mean noise, but not a
> differential bias in their local teaching measurements. This bias produces
> error floors, drift and task-switching cycles that can worsen as more local
> populations are biased. A somato-dendritic innovation removes the locally
> predictable component without backpropagation.

The method remains frozen:

```text
r = a - P(z)
Delta w = eta * r * eligibility
```

`P` is an affine or fixed-basis local adaptive filter trained by LMS on an
instruction-off observation. No backbone-specific recovery stage, copied clean
update, centered target, global loss or backpropagated signal may train it.

The debiasing mechanism is already de-risked in synthetic experiments. The
paper is not yet de-risked: the main remaining question is whether the same
locally identifiable operation beats strong bias-specific corrections on a
measured physical model and transfers unchanged across local-learning
backbones.

## What the paper must and must not claim

The paper may claim that structured differential bias does not disappear with
averaging and that its aggregate estimator norm can grow with the number of
biased local blocks. It may claim task-level scaling only under explicit
curvature, bias-alignment and identifiability assumptions, followed by the
corresponding experiment. It must not claim that EP, Dual Propagation or local
learning in general cannot scale cleanly.

The comparison target is each backbone's strongest correction for the stated
bias, not BP accuracy in a clean setting. Clean BP and clean native training
are reference points. A result counts as SDIL only after it passes
`HARDWARE_LOCALITY_CONTRACT.md`.

## Critical path and hard gates

### Gate 0: freeze the measurement and adapter contract

Deliverables:

- one adapter API exposing local neutral state, local teaching measurement and
  native local eligibility;
- one shared affine/fixed-basis predictor and neutral schedule across comparable
  digital backbones;
- no-grad, local-replay, downstream-independence, instruction-isolation,
  common-mode and phase/storage tests;
- a cost ledger counting phases/observations, stored local scalars, local
  multiplies, wall time and peak memory;
- exact definitions of common-mode bias, differential fixed bias,
  state-dependent bias, slow drift and same-RMS zero-mean noise.

Do not start expensive scale runs until this gate passes. This prevents a
software-only implementation or a backbone-specific correction from becoming
the reported method.

### Gate 1: P1 measured-physical mechanism

Recreate the released small-network coupled-learning dynamics using the
published equations and measured bias traces. Freeze the released task pairs,
bias traces, neutral cadence and predictor basis before comparing:

1. standard coupled learning;
2. same-RMS zero-mean noise;
3. constant per-edge calibration;
4. overclamping;
5. unchanged SDIL;
6. oracle subtraction.

Sweep cycle period and bias/drift strength over every compatible released
small-network instance. Report error floor, null-space drift, cycle span,
forgetting and charged cost. Use simulator initializations for uncertainty;
never present them as new independent hardware samples.

Pass conditions:

- structured bias remains after temporal averaging while matched zero-mean
  noise decreases;
- SDIL closes a substantial fraction of the raw-to-oracle gap on every
  identifiable state-dependent-bias setting;
- SDIL beats constant calibration when the bias varies with local state;
- SDIL is nondominated with overclamping on error/cycle span versus charged
  observation and circuit cost;
- common-mode bias gives no artificial SDIL advantage;
- task leakage and neutral-to-task shift fail in the direction predicted by
  the identifiability theory.

If SDIL loses to constant calibration or overclamping without a cost or
generality advantage, stop the physical-first paper and reassess before using
more GPU time.

### Gate 2: frozen-adapter transfer

Use the identical predictor family, update rule and neutral schedule in all
paper-facing digital experiments. Only the native local teaching measurement
and eligibility change with the backbone.

**Dual Propagation.** Complete the frozen five-seed miniCNN confirmation, then
run MiniCNN, VGG-style middle scale and author VGG16. The main conditions are
native clean, raw differential bias, constant calibration, SDIL and oracle;
same-RMS noise is the causal control. Use one seed to screen mechanics, three
seeds for the accept matrix and five only for final claims. One existing author
VGG16 run takes about 23,120 seconds (6.4 hours) on a GTX 1080, so VGG16 cells
should move to the A6000 after the one-seed gate.

**Equilibrium propagation.** Use the author DCHN implementation unchanged for
native dynamics and hyperparameters, with a hand-written paper-facing local
adapter. First test measurement/readout bias, for which constant calibration
and oracle subtraction are matched baselines. Separately test finite-nudge
bias against random-sign beta and centered EP. Do not claim SDIL solves
finite-nudge bias unless the neutral observation actually identifies it.

**Coupled-learning DCHN.** After EP passes, transfer the same adapter within the
same codebase. Compare with constant calibration, centered coupled learning and
an overclamping analogue. This is the cleanest digital bridge to the measured
physical experiment.

The accept matrix stops at three seeds and three families: measured physical
coupled learning, Dual Propagation, and EP. Digital coupled learning is added
as soon as the same adapter transfers; it is not allowed to delay the first
complete paper draft.

Pass conditions:

- the raw biased endpoint worsens with the declared scale axis in at least two
  independent families;
- neutral predictability, rather than corruption RMS, predicts recovery;
- SDIL improves every biased raw endpoint and beats constant calibration on
  state-dependent bias;
- SDIL is competitive with each family's strongest correction after cost is
  charged;
- the frozen adapter passes the executable BP-free audit in every family;
- clean/common-mode controls show that the gain is not ordinary regularization
  or a changed optimizer.

Failure on one family narrows the scope. It must not trigger a new
backbone-specific SDIL variant.

## Theory package

The theory and experiments share one bias model,

```text
a_l = s_l + b_l(z_l) + epsilon_l.
```

The accept version needs five results:

1. a bias--variance decomposition showing that averaging removes the
   conditionally zero-mean term but leaves the predictable differential term;
2. an aggregate-block result, carefully stated as estimator scaling rather
   than a universal loss theorem;
3. local quadratic dynamics predicting displaced optima, null-space drift and
   alternating-task cycle span;
4. a residual bound decomposing predictor approximation, sample error, drift
   and neutral-to-task shift;
5. an instruction-preservation condition plus explicit leakage and
   distribution-shift counterexamples.

Each theoretical quantity must appear on an experimental axis. A theorem that
does not predict a plotted transition, slope or failure control stays in the
appendix or is removed.

## Main figures

1. **Problem and real evidence:** Harnett residual motivation, the two-state
   measurement model, and the released physical error plateau/cycle drift.
2. **Mechanism on the physical model:** matched noise versus bias; raw,
   calibration, overclamping, SDIL and oracle; error/cycle span against charged
   cost.
3. **Scaling across backbones:** performance gap and residual bias versus depth,
   width, biased-block count and task difficulty for physical learning, DP and
   EP/CpL.
4. **What makes recovery possible:** neutral predictability, drift rate,
   task leakage and distribution shift, followed by an accuracy--cost Pareto
   summary.

The key scaling plot must show both the failure term and its removal. Plotting
only final accuracy across larger clean architectures is not evidence for the
new claim.

## Three publication bars

### Accept bar: target reviewer score 6--7

- Gates 0--2 pass for the measured physical model, DP and EP;
- all main comparisons have three seeds or every released physical replicate;
- the five theory results predict the main experiments;
- the method remains two lines and executable without BP;
- strong baselines, cost accounting, negative controls and failure cases are
  present;
- the complete manuscript makes the conditional claim, not a universal
  anti-EP claim.

At this point the paper has a new problem, a biologically motivated local
operation, real measured evidence, causal controls and cross-backbone transfer.

### Oral-B bar: target reviewer score 7--8

- extend EP and coupled-learning DCHNs across FashionMNIST, SVHN and CIFAR-10,
  multiple depths/biased-block counts and five seeds;
- finish five-seed VGG16 DP and the full cost Pareto comparison;
- cross fixed, state-dependent and slowly drifting bias with matched noise,
  predictability and neutral/task shift;
- keep one frozen predictor configuration across comparable backbones;
- show that measured or independently fitted bias parameters predict the
  observed scaling curves without post-hoc tuning.

This bar is breadth plus prediction: the same mechanism must explain when SDIL
works, how much it recovers and when it fails.

### Oral-A bar: target reviewer score 8--9

- demonstrate the local filter and subtraction path in circuit simulation or
  on physical hardware, or validate on a second independently measured
  substrate;
- complete long large-scale runs such as CIFAR-100/large DCHNs only after the
  adapter is frozen;
- compare end-to-end phase, storage, precision, time and energy costs rather
  than only optimizer steps;
- obtain a prospective prediction: select bias/predictability conditions from
  small systems, then correctly predict recovery and failure on the held-out
  larger system.

These long runs may use spare GPUs in parallel after Gate 0, but their design
must not change in response to their final endpoints.

## Execution order and rough budget

1. **Now (hours):** let the existing DP C1 run finish; audit all five seeds and
   commit only the completed frozen result.
2. **Next 1--3 working days:** implement and test the released small physical
   model, finish Gate 0, and run the cheap P1 CPU sweep.
3. **Following week:** write theory v0 beside the P1 plots; run one-seed DP
   scale and EP screens. Reject broken adapters before replication.
4. **Following 1--2 weeks:** run the three-seed accept matrix on the 1080s and
   OscarWX's A6000; build the four main figures and write a full paper draft.
5. **After a reviewer-style audit reaches 6:** launch five-seed/multi-dataset
   Oral-B jobs. Start Oral-A long runs opportunistically on otherwise idle
   cards.

Expected GPU exposure is roughly 40--100 hours for screened accept-scale
digital runs and several hundred GPU hours for Oral-B breadth. These are
planning ranges, not promises: actual per-cell time must be recorded from the
one-seed screens before the full launch. P1 and theory are the immediate
critical path; adding GPUs cannot compensate for failure there.

## Reviewer-score ledger

Update one evidence table after every completed gate:

| item | evidence | strongest alternative | unresolved failure | score effect |
|---|---|---|---|---|
| real bias exists | released physical P0 | overclamping paper | no SDIL fix yet | necessary, not sufficient |
| physical mechanism | P1 | calibration/overclamping | pending | unlocks 6 |
| general transfer | DP + EP/CpL | native correction | pending | supports 6--7 |
| scaling prediction | theory + held-out scales | descriptive curve | pending | supports 7--8 |
| hardware realization | circuit/second substrate | primitive argument | pending | supports 8--9 |

Every progress report should state positive results, negative results, compute
and a fresh reviewer score. Accuracy without the matched strongest baseline,
cost coordinate or frozen protocol does not move the score.