# ICLR 2027 paper plan

This document maps the frozen evidence to a defensible paper narrative. It is
not permission to promote development results or to hide failed gates.
The current prose realization is `paper/MANUSCRIPT.md`; its 34 central numeric
claims, figure set, positive D4 gate, and seven retained R2 failures are checked
by `experiments/audit_manuscript.py`.

## One-sentence claim

Per-neuron subtraction of the apical component predictable from neutral
somatic activity isolates an innovation signal that can preserve useful local
credit when raw feedback is contaminated by ordinary soma-coupled traffic;
when paired with causally learned feedback, the resulting local learner
preserves accuracy and alignment over depth substantially better than fixed
direct feedback in the audited regime.

This sentence deliberately says neither that cortex implements
backpropagation nor that the current model explains Harnett's temporal-error
signature. The frozen D3 validation endpoint supports a near-BP standard
ResNet-20 result, but “scales on standard residual networks” may be added only
after the untouched multi-seed/depth confirmations pass.

## Working title

**Learning from the Unexpected: Somato-Dendritic Innovations for Local Credit
Assignment**

If Oral-A A4 passes cleanly, “Scalable” may be added to the title. If A3 or A4
fails, the title stays mechanism-focused and the standard-scale result appears
as a limitation rather than a title claim.

## Abstract logic

1. Biological and local-learning models often treat raw apical activity as a
   teaching signal even though it also carries ordinary state/context traffic.
2. Inspired by the per-neuron residual measured by Harnett and colleagues,
   define teaching as apical activity minus its neutral-period prediction from
   the same neuron's soma.
3. Show that this is the orthogonal conditional innovation within the chosen
   somatic function class; norm matching cannot reproduce its directional
   effect. Calibrate the inherited feedback map using causal perturbations and
   update forward weights with local eligibilities.
4. Report the frozen residual-necessity panel: at strong predictable traffic,
   innovation retains `97.35%` while raw and norm-matched raw reach about
   chance, with an `87.038 +/- 0.607`-point paired gain over norm-matched raw.
5. Report the frozen depth panel: SDIL changes by only
   `-0.214 +/- 0.349` points over 12x hidden depth while DFA alignment falls to
   `0.047`; qualify that flattened CIFAR is depth-flat.
6. Report the untouched standard-ResNet confirmation: dynamic innovation
   reaches `91.584%` mean test accuracy under four-times-RMS predictable
   traffic versus clean KP's `91.388%`, with a `0.131`-point one-sided deficit
   bound, `0.999687` early alignment, zero task-loss queries, and `1.326x` BP
   MACs. Attribute reciprocal KP credit to Akrout et al. and disclose the
   paired neutral microphase. End with the supported scope: ResNet-20
   robustness is established, but positive added-depth utility and the broader
   Harnett-like population signature are not.

The abstract should not mention the failed desired-velocity hypothesis unless
the paper explicitly positions that falsification as a contribution. It must
not imply that no-traffic scaling establishes the necessity of residualization.

## Contribution statements

1. **Mechanism.** A per-cell neutral-period innovation rule for separating
   task instruction from ordinary apical traffic, with raw and norm-matched
   controls.
2. **Theory.** A conditional-projection account of what residualization can
   optimally remove, its failure boundary, a descent condition separating
   direction/gain/curvature, and exact simultaneous-perturbation variance.
   The conditional-expectation theorem itself is standard and is not claimed
   as mathematical novelty.
3. **Evidence.** Frozen, seed-complete tests of innovation necessity, depth
   preservation, alignment, query/MAC/memory cost, and direct-versus-amortized
   causal feedback.
4. **Scientific negatives.** Broad top-down traffic, useful-depth learned
   vectorization, the Harnett desired-velocity/full population signature, and
   the original learned-vectorizer ResNet branch fail their frozen gates,
   defining the method's actual boundary.
5. **Standard scale.** D3 passes all frozen checks at `91.18%` validation on
   ResNet-20 versus BP's `91.62%`; the independent paired D4 panel then reaches
   `91.584%` mean test accuracy versus clean KP's `91.388%`. This supports
   ResNet-20 robustness, not a positive depth-scaling claim. The separately
   frozen dynamic-depth panel in `ORAL_A_RECOVERY.md` was the only permitted
   route to the latter, but its oral-B R2 prerequisite fails, so no new
   ResNet-32/56 endpoint is opened.

## Figure order in the manuscript

The current filenames reflect generation history rather than final paper
numbering.

1. **Mechanism and necessity:** current `figure3_innovation`. Lead with the
   subtraction diagram, accuracy under traffic, and used-signal alignment.
2. **Standard-network confirmation:** current
   `figure4_resnet_confirmation`. Show untouched paired test endpoints,
   layerwise raw-versus-innovation direction, the complete credit-tracking
   trajectory, and the explicit resource audit together.
3. **Accuracy/cost:** current `figure1_pareto`, augmented by the completed
   native-author table in the text or supplement. Do not place unmatched
   architectures on one purported equal-compute frontier.
4. **Depth scaling:** current `figure2_scaling`. Retain the MLP result as a
   controlled depth-preservation diagnosis; report the failed ResNet-20 A3
   trajectory in the limitations or supplement, not as a scaling figure.
5. **Mechanism/cost anatomy:** query-budget retention, direct-NP diagnosis,
   predictor timescale, and failure boundaries. Keep
   `figureS_dynamic_stability` in the supplement because the D4 main figure now
   shows the complete stable trajectory; cite it for the failure-to-recovery
   sequence.
6. **Biological signatures:** the complete failed Oral-B screen belongs in the
   supplement/limitations unless the paper foregrounds falsification. Report
   its positive decodability and negative causal-role results together.

## Section skeleton

### 1. Introduction

- Mixed apical traffic is the unaddressed problem, not merely transporting an
  output error to a dendrite.
- Harnett motivates a per-neuron residual operation without proving that it is
  a plasticity signal.
- Existing dendritic, feedback-alignment, Dual Prop, BurstCCN, and learned
  synthetic-feedback methods define the prior-art boundary.

### 2. Somato-dendritic innovation learning

- Forward dynamics and local eligibility.
- Neutral predictor and innovation.
- Causal perturbation calibration, clearly attributed to learned synthetic
  feedback prior art.
- Online-control variant described only as a tested hypothesis, not as the
  successful method.

### 3. Theory and resource accounting

- Conditional projection and task-period absorption.
- Multiplicative residual coupling, the two-sided momentum/Jury stability
  window, and fast neutral projection as a local empirical certificate.
- Hidden alignment versus parameter descent.
- Simultaneous perturbation bias/variance.
- Logical loss queries, MACs, memory, and wall time as separate axes.

### 4. Does innovation matter?

- Frozen 60-run predictable-traffic panel and matched-raw control.
- Endogenous C1 near-miss/failure and FashionMNIST recovery failure.
- The allowed claim is restricted to the predictor's somatic information.

### 5. Does causal local credit survive depth?

- Five-depth BP/FA/DFA/SDIL panel and local Pareto frontier.
- Direct-NP useful-depth diagnosis and the failed learned-vectorizer branches.
- Low-query confirmation.
- Standard ResNet results only according to the frozen A1--A4 branch.

### 6. Relation to biological signatures

- Residual decorrelation, decoder, and plasticity-lesion positives.
- Wrong sign-inversion, error-magnitude dominance, and weak acute online
  lesion; desired velocity is unsupported.
- Keep the original failed screen intact. The separately frozen recovery
  factorizes learned causal role from within-episode performance innovation;
  its D4-gated development screen passes, but the untouched 6-task-by-5-model
  confirmation fails the joint population-vectorization/longitudinal gate.
  Report the learning, lesion, sign-inversion, and velocity positives as a
  bounded diagnostic, not as a passed oral-B claim.
- Even a complete recovery pass supports innovation-guided plasticity, not
  online neural control: the recovery fixes `kappa=0` by construction.

### 7. Limitations and discussion

- Predictor observability and distribution invariance.
- Perturbation variance and feedback amortization.
- Standard architecture/energy cost versus biological plausibility.
- No inference from algorithmic utility to the causal role of dendritic
  residuals in cortex.

## Reviewer objection map

| Objection | Evidence that addresses it | Residual risk |
|:--|:--|:--|
| “This is Lansdell synthetic feedback renamed.” | `NOVELTY.md`; no-traffic method explicitly attributed; raw/matched/innovation panel isolates the new operation | Novelty remains narrow and needs clear writing |
| “The residual only clips an oversized signal.” | Per-example norm-matched raw control; positive scaling preserves cosine; Figure mechanism panel | Artificial predictable traffic is still a controlled construction |
| “Depth does not help this task.” | Claim says preservation; C2 reports the learned-vectorizer failure; Oral-A A3 is retained as a failed standard-scale test | Fatal to a broad scaling claim; the paper must remain mechanism-focused |
| “Local methods hide enormous extra work.” | Logical queries, MACs, peak memory, wall time, C3 confirmation, EP/native protocols | Hardware implementations are not all equally optimized |
| “Weak baselines define the win.” | BP/FA/DFA/direct NP/FF/PEPITA/EP plus native BurstCCN and Dual Prop | Native reproductions are one seed and method-native, not equal compute |
| “Harnett already proves the biological story.” | Original oral-B negative result plus the separately frozen role/velocity recovery and untouched 30-record confirmation | Recovery fixes learning and sign but fails the joint outcome/longitudinal gate; `kappa=0` cannot support desired velocity |
| “The fast controller is just an unreported contrastive phase.” | Paired neutral observation count and elementwise/wall cost are explicit; no task-nudged state, loss query, or reverse pass is used | The instruction-off microphase is a real assumption and weakens the single-phase biological claim |
| “The gates were selected after results.” | Git-frozen protocols, untouched confirmation seeds, failed branches retained | Early inherited pilots predate the strict boundary and must remain labeled |

## Result-dependent branch

### If A3/A4 pass

Lead with the standard ResNet result, promote empirical support, and retain the
MLP useful-depth failure as a diagnosis of the earlier vectorizer. The paper
can plausibly claim a rare local-learning method that retains performance and
alignment with standard depth under an audited cost budget. Reviewer score
should be reassessed from the frozen confirmation only.

The original A3/A4 branch is permanently failed, so this paragraph can now
apply only to the separately frozen dynamic-depth recovery. Its full 60-cell
pass was required before a standard-depth abstract/title claim or a 9/10
score; D4 alone remained a single-architecture robustness result. The
calibrated-BCI-gated oral-A-v2 panel subsequently passed all 60 record,
accuracy, paired depth-gain, alignment, mechanism, query, MAC, and memory
checks. The manuscript may therefore make the standard-depth claim, while
attributing clean-task transport substantially to inherited reciprocal KP.

### If A3 or A4 fails

Do not soften the threshold or replace the seed panel. Keep a mechanism paper:
innovation is load-bearing under identifiable mixed traffic and the inherited
causal-feedback backbone preserves depth on the controlled task, but standard
useful scaling remains unresolved. The correct response is a narrower title,
claim, and score—not another post-hoc ResNet tuning branch.

This is the realized outcome for the original A3/A4 branch: A3 SDIL became
nonfinite at epoch 89 and ended at chance, and A4 remained untouched. It is
retained as a failed direct-vectorizer path rather than treated as the final
standard-depth outcome. The independently authorized dynamic-innovation v2
recovery later passed ResNet-20/32/56, so the remaining limitation is matched
baseline and architecture breadth rather than positive depth utility itself.

## Post-A3 accept recovery: innovation on a strong inherited substrate

The failed direct-vectorizer A3 branch remains closed. The post-failure
baseline audit found two sharply different outcomes: intermittent residual
response mirroring ends at chance with NaN validation loss despite 0.999998
endpoint Q/W cosine, while modified Kolen--Pollack passes its frozen 20-epoch
screen at 82.66%. Both mechanisms are inherited and receive no novelty credit.

`MIXED_TRAFFIC.md` defines the only current route that can revise the
standard-ResNet conclusion. It keeps reciprocal KP fixed as the instructional
substrate and crosses raw, norm-matched raw, and neutral-period innovation
under identical four-times-RMS soma-predictable traffic. One short screen, one
full seed-0 validation panel, and one untouched five-seed test panel were
frozen before any mixed-traffic task endpoint. Only the complete final panel
can raise the reviewer estimate to 6/10.

If it passes, revise the abstract and contribution language narrowly:

- claim that somato-dendritic innovation is load-bearing on a standard
  ResNet-20 with a strong zero-query local reciprocal credit path;
- attribute credit transport and reciprocal plasticity explicitly to Akrout
  et al.; do not describe KP performance as SDIL novelty;
- report the 376,832 predictor parameters, elementwise arithmetic, affine
  MACs, peak memory, and wall time rather than presenting a MAC-only frontier;
- retain failed natural/top-down traffic and desired-velocity results, so the
  claim remains predictable-traffic removal rather than a cortical model.

If MT-1, MT-2, or MT-3 fails, preserve the current mechanism-only narrative and
the score of 5/10. No lower traffic ratio, deleted seed, or replacement
confirmation panel is permitted.

This stop rule has now fired at MT-1. All three signal conditions become
nonfinite in epoch 1 and end at chance despite passing the calibration,
predictor-warmup, query, and cost invariants. MT-2 and MT-3 remain untouched.
The paper therefore retains the mechanism-focused title and 5/10 assessment;
the controlled standard-ResNet recovery is a disclosed negative result, not an
active acceptance claim.

## Post-MT-1 operator-stability branch

MT-1 remains failed and its sealed MT-2/MT-3 panels are never reopened. The
separate post-failure theory identifies the predictor error as the
multiplicative operator `D W C`; S0 then falsifies fixed sign-only margins.
`DYNAMIC_INNOVATION.md` defines a new evaluation boundary: a slow neutral
predictor plus a fast paired instruction-off projection of the remaining
neutral residual. This is closer to a local apical shunting/homeostatic
controller than to Dual Prop's two oppositely nudged task states, but the paper
must disclose the extra microphase rather than call the method single-phase.

D1 passes its training-only stability gate. D2 then reaches `83.58%` at 20
epochs versus clean KP's `82.66%`, while the failed raw/matched controls remain
at `10%`; all projection, query, MAC, memory, and evaluation-boundary checks
pass. This is the first standard-ResNet evidence for load-bearing innovation,
but it is one short validation run and therefore leaves the score at 5/10.

The predeclared D3 run passes all 19 gates at `91.18%` validation accuracy and
opens the independently frozen D4 confirmation.  D4 then passes without seed
replacement or threshold changes: dynamic innovation reaches `91.584%` mean
test accuracy across seeds 10--14 versus clean KP's `91.388%`, with a
clean-minus-dynamic one-sided 95% upper bound of `0.131` points.  All mechanism,
query, cost, memory, provenance, split, and endpoint-isolation checks pass.
The strict reviewer score is therefore 7/10 and the narrow accept claim is now
active: somato-dendritic innovation is load-bearing and robust on a standard
ResNet-20 with an inherited zero-query reciprocal credit path.

The paper must still attribute reciprocal KP plasticity to prior work, count
the paired neutral microphase and elementwise work, and retain the failed
arbitrary-top-down, desired-velocity, online-control, and earlier unstable
branches.  D4 alone cannot support positive added-depth utility. Oral-B R1
selects `eta=0.1` with 98.05% worst-task development success, but the untouched
R2 confirmation fails the broader outcome-vectorization and longitudinal
signature checks despite 99.53% mean success and 30/30 positive sign
inversions. The score therefore remains 7 and the frozen
ResNet-20/32/56 oral-A panel stays closed. Do not repair thresholds or route
around the failed prerequisite.