# SDIL novelty audit This file records the closest-prior-art boundary as of 2026-07-22. It is a claim constraint, not a literature-survey completeness claim. New evidence may narrow the allowed claims further. ## Exact overlap with learned node-perturbation feedback Lansdell, Prakash, and Kording, [Learning to solve the credit assignment problem](https://arxiv.org/abs/1906.00889) (ICLR 2020), already train synthetic feedback matrices from node-perturbation estimates. Their direct-feedback form maps the output error to every hidden layer and was evaluated in a CIFAR convolutional network. Their public implementation is [`benlansdell/synthfeedback`](https://github.com/benlansdell/synthfeedback). With no ordinary apical traffic and `P=0`, SDIL reduces exactly to this idea (up to sign and tensor conventions): ```text c = output error r_l = A_l c q_l = node-perturbation estimate of -dL/dh_l A_l <- A_l + eta_A (q_l - A_l c) c^T ``` Independent noise at multiple hidden layers during a shared noisy forward pass is also present in the earlier method. SDIL's antithetic Rademacher directions, vectorized direction batching, and explicit cost accounting are useful estimator/implementation choices, but are not the core algorithmic novelty. The existing no-traffic depth results therefore establish the strength of a learned-node-perturbation backbone; by themselves they do not establish a Harnett-specific contribution. The closest exact baseline is consequently the same implementation with `traffic_mode=none`, `learn_P=0`, and `use_residual=0`. It should be named **learned direct feedback by node perturbation (Lansdell et al.)**, not SDIL, in comparisons intended to isolate novelty. ## Overlap with dendritic credit assignment Apical/basal segregation, neuron-specific dendritic teaching signals, and local three-factor or burst-dependent plasticity substantially predate SDIL. In addition, Greedy et al., [Cell-type-specific cortical feedback coordinates hierarchical credit assignment](https://www.biorxiv.org/content/10.64898/2026.06.16.732595v1) (BurstCCN, 2026), explicitly connects a scalable dendritic credit-assignment model to the Francioni/Harnett data. BurstCCN already: - models deviations from balanced dendritic feedback as neuron-specific error signals; - reproduces the opposite residual signs of the P+ and P- BCI populations and their disruption by NDNF activation; - trains convolutional models on CIFAR-10 and ImageNet with plastic feedback; and - provides an [author implementation](https://github.com/neuralml/BurstCCN-journal). Thus neither "apical dendrites carry vector errors", "the model reproduces Harnett Fig. 5", nor "a dendritic local rule scales to ImageNet" is a sufficient novelty claim. Likewise, recursive spatial feedback, learned feedback convolutions, and local ReLU-gated error propagation are not new by themselves. The post-failure hierarchical oracle deliberately reconstructs a BP-like local adjoint and may guide engineering, but it cannot be presented as an SDIL contribution. Any trainable version must be compared directly with hierarchical FA, learned feedback alignment/weight-mirroring methods, and BurstCCN. The Harnett-specific claim remains the neutral-period somato-dendritic innovation operation under mixed traffic. The implemented feedback-parameter perturbation rule is task-causal rather than a copied-weight or forward-reconstruction rule, but this distinction does not by itself make recursive learned feedback novel. It is treated as a credit-path engineering component unless the complete innovation model shows a load-bearing residualization result beyond those mandatory comparators. Likewise, the normalized local response-mirror path is explicitly inherited from weight mirroring/weight estimation. Even if it recovers BP-level scale, that performance belongs to the baseline. A paper-level SDIL gain requires the somato-dendritic innovation subtraction to remain necessary on top of the mirrored hierarchical path under a frozen mixed-traffic intervention. Residual response mirroring changes the inherited baseline's estimator but not this attribution. Its zero-noise fixed point and any resulting scale belong to local predictive weight estimation; they become relevant to SDIL only as the feedback substrate on which innovation is separately ablated. The modified Kolen--Pollack reciprocal-plasticity path is likewise inherited from Akrout et al. Equal local activity products and matched decay in reciprocal synapses are not SDIL. Even if the audited implementation reaches BP-level accuracy, it only supplies a stable feedback substrate. The candidate novelty still requires neutral somato-dendritic subtraction to outperform raw and norm-matched raw apical activity under the same mixed-traffic intervention. The post-MT-1 fast neutral projection does not make reciprocal KP novel. It adds a second-timescale operation to the innovation mechanism: ```text e0_l = a0_l - P_l(h_l) neutral residual q_l = Cov(e0_l,h_l) / Var(h_l) current local coupling r_l = s_l + e0_l - E[e0_l] - q_l(h_l-E[h_l]) stabilized innovation ``` The coefficient fit is local and instruction-off, so it does not use a task loss, task-nudged state, downstream weight, or reverse pass. Its defensible novelty is not “two phases” or generic normalization; it is the explicit operator-stability role of a per-cell somato-dendritic innovation measured from paired neutral observations. The paper must nevertheless count and disclose the neutral microphase. It cannot describe this variant as single-phase, and must discuss proximity to contrastive/phase-separated local-learning methods. ## Candidate SDIL contribution The candidate contribution is the combination ```text a_l = A_l c + t_l mixed apical traffic P_l(h_l) ~= E[t_l | h_l, neutral] per-cell neutral predictor r_l = a_l - P_l(h_l) somato-dendritic innovation Delta W_l proportional to r_l * eligibility_l ``` where the subtractive per-cell innovation, rather than raw apical activity, is the teaching variable. The feedback vectorizer `A_l` may be learned using Lansdell-style causal perturbations; that component is inherited, not claimed as new. The new empirical question is whether residualization is load-bearing when teaching signals share an apical channel with ordinary somatic/contextual feedback. Neutral-period fitting is an algorithmic adaptation rather than a direct statement of the Harnett paper. It addresses an identifiability problem: task-period regression can subtract the predictable part of `A_l c` whenever output error correlates with somatic state, while neutral-period regression identifies the traffic relationship without observing the teaching component. The formal justification is an application of the standard conditional-expectation projection identity, not a new mathematical theorem. Among all predictors measurable from the chosen somatic statistic, the neutral conditional mean uniquely minimizes residual nuisance power and leaves an innovation orthogonal to every such predictor. The implemented per-cell affine model realizes the restricted projection onto `{1,h_i}`. Positive norm matching cannot change the raw signal's cosine direction at a fixed state, so the matched-raw ablation specifically distinguishes subtraction from gain control. `THEORY.md` states the assumptions and the executable finite-sample identities. ## Claims currently allowed - The inherited learned-feedback backbone preserves performance with depth better than fixed DFA in the audited flattened-CIFAR MLP setting. - Frozen controls show that per-cell residualization protects learning from soma-coupled traffic. It mitigates but does not fully remove endogenous top-down traffic, so arbitrary contextual innovation is not an allowed teaching-signal claim. - Neutral residualization is the minimum-power nuisance removal within the predictor's somatic function class. For the implemented diagonal affine class this is a narrow, testable statement, not a claim of optimal population-level denoising. - Simultaneous batched perturbations have a frozen hardware-independent advantage: K1/every4 retains 112.9% of the K16/every4 gain over DFA with 16x fewer logical loss queries, 11x less calibration work, and 5.3x less total forward-equivalent work on CIFAR-10 d20/w64. - Direct node perturbation solves the frozen useful-depth task while the amortized context vectorizer does not. This identifies feedback amortization as the current engineering bottleneck; it is not a novelty claim and the direct estimator's 68.4x work precludes presenting it as the scalable algorithm. - On the inherited reciprocal-KP substrate, dynamic paired-neutral innovation first passes a frozen 20-epoch ResNet-20 validation gate at 83.58%, versus 82.66% for clean KP and 10% for the failed raw/norm-matched mixed-traffic controls. The subsequent 200-epoch D3 validation reaches 91.18%, and the independently frozen paired D4 test panel reaches 91.584% across seeds 10--14 versus clean KP's 91.388%, with a 0.131-point one-sided upper bound on the clean-minus-dynamic deficit. This supports load-bearing innovation and robustness on ResNet-20. It does not establish positive utility from adding standard-network depth. ## Claims currently forbidden - Node perturbation training of `A_l` is novel. - Simultaneously perturbing all hidden layers is novel. - The no-traffic scaling panel demonstrates the necessity of somato-dendritic residuals. - SDIL is the first dendritic method to scale to deep vision tasks. - Dynamic neutral projection is a single-phase rule or has zero observation cost merely because it adds no convolutional MAC. - A single ResNet-20 architecture establishes scaling with added standard depth, or licenses the sealed ResNet-32/56 panel after its oral-B prerequisite failed. - SDIL reproduces the complete Harnett population outcome-vectorization, longitudinal, desired-velocity, or online-control signature. The untouched oral-B R2 panel fails its joint gate despite successful task learning and sign inversion. - Reproducing the already-published Harnett/BurstCCN qualitative signatures is an oral-level biological contribution. ## Evidence that can restore a strong paper 1. Show across independently generated, endogenous traffic families that residualization beats raw and norm-matched raw feedback, while the no-traffic methods coincide. 2. Compare directly with Lansdell-style learned feedback and BurstCCN under validation-selected, disclosed compute/query budgets. 3. The conditional-projection/SNR result and task-period absorption prediction are now proved and executable. A broader paper would still benefit from a separately frozen failure test where traffic depends on the teaching signal after conditioning on the available soma statistic. 4. For a biological contribution, make and test a prediction not used by Harnett or BurstCCN. Raw event-level data and analysis code require author coordination; the paper exposes only plotted source data and states that full data/code are available on request. 5. For a scaling contribution, target standard architectures and a simplicity/cost regime not already occupied by BurstCCN, rather than treating ImageNet alone as novelty. 6. D3 and D4 now establish the narrow ResNet-20 result. A stronger scaling contribution requires an independently frozen, genuinely depth-necessary standard-network panel; the existing ResNet-20/32/56 panel remains sealed because oral-B R2 failed. 7. A biological contribution beyond Harnett/BurstCCN requires a new mechanism and prediction frozen independently of the failed R2 metrics. Successful task learning, lesion sensitivity, and sign inversion alone cannot be relabelled as the failed joint population-signature claim.