id stringlengths 3 3 | title stringlengths 11 56 | severity stringclasses 3
values | category stringclasses 8
values | engineId stringclasses 5
values | page stringlengths 34 78 | provenance stringclasses 3
values | detail unknown |
|---|---|---|---|---|---|---|---|
G01 | Impact preview | warn | blast radius | impact | https://neurarch.com/r/impact-preview | null | {
"trigger": "Any destructive action (delete_component, delete_components_matching, rename_components_matching, clear_canvas, replace_model) where downstream impact > 4 layers, or any of them are shape-changing.",
"surfaces": "Number of downstream layers affected, how many would reshape, how many carry weights (reb... |
G02 | Param explosion | block | resource | param-explosion | https://neurarch.com/r/param-explosion | null | {
"trigger": "Estimated total parameter count grows by ≥ threshold × baseline (default 5×, slider 2-20×). Blocks at 2× threshold, warns below.",
"why": "A misread \"scale this 10×\" prompt should not silently turn a 60M model into a 6B model that won't fit anywhere.",
"setting": "useProviderStore.paramExplosionTh... |
G03 | Cycle introduction | block | structure | cycle | https://neurarch.com/r/cycle-introduction | null | {
"trigger": "An add_connection closes a cycle in the existing DAG. DFS check on the union of current edges and proposed edges.",
"why": "NN forward graphs are acyclic. Recurrence lives inside the layer type (LSTM, GRU), not as a graph cycle. A connection that closes a cycle is almost always a mis-step.",
"source... |
G04 | Orphan layers | warn | structure | orphan | https://neurarch.com/r/orphan-layers | null | {
"trigger": "add_component without afterName, and no follow-up add_connection referencing the new component as either endpoint.",
"why": "A new layer that's not wired is dead code. The agent occasionally proposes these when it forgets to chain edits; the warning flips the decision back to the user."
} |
G05 | Shape inference (new) | block | shape | shape | https://neurarch.com/r/shape-inference-new | head-dim-divisibility | {
"trigger": "Propagates tensor shapes through the sandboxed post-action graph and surfaces only the issues the action would introduce. Sub-checks: attention embedDim % numHeads, GQA numHeads % numKVHeads, elementwise merge parent equality, concat axis compatibility, explicit linear inFeatures vs upstream, computeOut... |
G06 | Param range | warn | param-range | null | https://neurarch.com/r/param-range | null | {
"trigger": "A single parameter value falls outside its conventional range on add_component or update_params: e.g. dropout above 1, stride of 0, numHeads of 0. Uses the same convention-based ranges the inspector shows inline.",
"why": "These are values the shape gate cannot see: the graph still propagates, the mod... |
R01 | No Input node | block | structure | null | https://neurarch.com/r/no-input-node | null | {
"trigger": "Model has ≥ 1 component but no component of type input.",
"why": "Without an Input layer, tensor shapes can't be propagated and generated code is incomplete.",
"fix": "Drag an Input layer from the I/O section of the palette."
} |
R02 | No Output node | warn | structure | null | https://neurarch.com/r/no-output-node | null | {
"trigger": "Model has > 1 component but no component of type output.",
"why": "Code generator can't tell where the forward pass terminates."
} |
R03 | Isolated components | warn | structure | null | https://neurarch.com/r/isolated-components | null | {
"trigger": "Component has no inbound and no outbound connections.",
"why": "Most often a stranded layer left over from a refactor. Code generator skips it but it inflates the visual graph."
} |
R04 | Dead-end layer | warn | structure | null | https://neurarch.com/r/dead-end-layer | null | {
"trigger": "A non-output node has inbound connections but no outbound connections.",
"why": "Compute happens but the result is discarded. Gradients flow into a black hole."
} |
R05 | BatchNorm / LayerNorm after activation | warn | ordering | null | https://neurarch.com/r/batchnorm-layernorm-after-activation | null | {
"trigger": "A normalization layer is connected directly downstream of an activation (relu, gelu, swish, etc.).",
"why": "Normalization is meant to stabilize the pre-activation distribution. Applying it after activation breaks the assumption and shifts already-nonlinear features back toward zero mean.",
"source"... |
R06 | Dropout directly before BatchNorm | info | ordering | null | https://neurarch.com/r/dropout-directly-before-batchnorm | null | {
"trigger": "A dropout layer is connected directly upstream of a BatchNorm.",
"why": "Dropout at train time changes the activation variance; BN's running statistics get distorted. A known train-vs-eval mismatch.",
"source": "Li et al. 2018, \"Understanding the Disharmony Between Dropout and Batch Normalization\"... |
R07 | Softmax / Sigmoid directly before Output | info | ordering | null | https://neurarch.com/r/softmax-sigmoid-directly-before-output | null | {
"trigger": "An explicit Softmax or Sigmoid layer is wired directly into Output.",
"why": "PyTorch's nn.CrossEntropyLoss applies LogSoftmax internally; an explicit Softmax before it double-applies and slows training. BCEWithLogitsLoss has the same issue with Sigmoid.",
"source": "torch.nn.CrossEntropyLoss docs."... |
R08 | Normalization at output | warn | ordering | null | https://neurarch.com/r/normalization-at-output | null | {
"trigger": "A normalization layer is wired directly into Output without a final linear / classification head.",
"why": "Normalization zero-centers the logits, breaking calibration. The model's output scale is no longer meaningful."
} |
R09 | Deep network with no residual connections | warn | pattern | null | https://neurarch.com/r/deep-network-with-no-residual-connections | null | {
"trigger": "≥ 8 weight-carrying layers (conv2d / linear / etc.) and zero residual / skip / add nodes.",
"why": "Gradient signal degrades through depth without skip connections.",
"source": "He et al. 2015, ResNet."
} |
R10 | Attention without positional encoding | warn | pattern | null | https://neurarch.com/r/attention-without-positional-encoding | null | {
"trigger": "Model contains attention layers but no positional encoding (sinusoidal, learned, RoPE, ALiBi).",
"why": "Self-attention is permutation-invariant. Without PE, the model is a bag-of-words and can't learn order.",
"source": "Vaswani et al. 2017, Attention Is All You Need, Section 3.5."
} |
R11 | Sigmoid / Tanh in deep networks | info | pattern | null | https://neurarch.com/r/sigmoid-tanh-in-deep-networks | null | {
"trigger": "Sigmoid or Tanh activation appears in a network with ≥ 5 weight-carrying layers.",
"why": "Saturating activations vanish gradients in deep stacks. ReLU / GELU / SiLU are the modern defaults.",
"source": "Glorot & Bengio 2010 on the vanishing gradient problem."
} |
R12 | Deep network with no normalization | info | pattern | null | https://neurarch.com/r/deep-network-with-no-normalization | null | {
"trigger": "≥ 7 non-I/O layers and zero normalization layers (BN / LN / IN / GN / RMSNorm).",
"why": "Without normalization, deep nets converge slowly and are sensitive to initialization scale."
} |
R13 | Very high dropout rate | warn | performance | null | https://neurarch.com/r/very-high-dropout-rate | null | {
"trigger": "Any dropout layer with p > 0.65.",
"why": "Common dropout rates are 0.1 to 0.5. Anything above 0.65 usually indicates a typo or a confused regularization strategy."
} |
R14 | Very large activation tensor | warn | performance | null | https://neurarch.com/r/very-large-activation-tensor | null | {
"trigger": "Any layer's estimated activation tensor exceeds 50 M elements (~200 MB per sample at float32).",
"why": "Activations live in GPU memory through the backward pass. One bloated layer forces tiny batch sizes or OOM."
} |
R15 | MoE without auxiliary loss | info | pattern | null | https://neurarch.com/r/moe-without-auxiliary-loss | null | {
"trigger": "Model contains an MoE layer but no auxiliary load-balance loss is referenced.",
"why": "Without an aux loss, experts collapse into a few dominant ones. Most papers add a 0.01 weight router-balance term.",
"source": "Shazeer et al. 2017, Sparsely-Gated MoE, Section 4."
} |
R16 | GQA numHeads not divisible by numKVHeads | block | structure | null | https://neurarch.com/r/gqa-numheads-not-divisible-by-numkvheads | gqa-head-divisibility | {
"trigger": "A groupedQueryAttention component has params where numHeads % numKVHeads != 0.",
"why": "GQA groups query heads; the group size must divide the head count. Mis-set ratios crash on the first attention forward.",
"source": "Ainslie et al. 2023, GQA."
} |
R17 | SwiGLU intermediateSize convention | info | performance | null | https://neurarch.com/r/swiglu-intermediatesize-convention | null | {
"trigger": "A swiGLU / gated FFN has intermediateSize that doesn't follow the common ~(2/3) × 4 × hidden convention used in LLaMA / Mistral.",
"why": "Param-count parity with standard FFN; mis-sized SwiGLU silently changes the FLOP and param budget.",
"source": "Shazeer 2020, GLU Variants."
} |
R18 | Conv feeds Linear with no flatten / pool | block | structure | null | https://neurarch.com/r/conv-feeds-linear-with-no-flatten-pool | null | {
"trigger": "A convolution (conv1d/2d/3d or a depthwise/separable/transpose variant) connects directly into a linear layer.",
"why": "Conv outputs a multi-dimensional feature map; Linear expects a flat [batch, features] tensor. The forward pass raises a shape error, or silently mis-multiplies the spatial dims. Ins... |
R19 | Back-to-back activations | warn | ordering | null | https://neurarch.com/r/back-to-back-activations | null | {
"trigger": "An activation node (relu/gelu/silu/sigmoid/tanh/softmax/…) connects directly into another activation node with no Linear/Conv/Norm between them.",
"why": "Two activations in a row add compute but no expressivity (same-type is a no-op duplicate). A common slip is ReLU → Softmax, which clips logits to ≥... |
R20 | Linear → Linear with no activation | info | pattern | null | https://neurarch.com/r/linear-linear-with-no-activation | null | {
"trigger": "A linear layer connects directly into another linear layer with no activation between them.",
"why": "Two stacked linear maps collapse into one (W₂·W₁), so the extra layer costs parameters but adds no representational power, usually a forgotten ReLU. Deliberate low-rank / factorized projections (down-... |
R21 | Dropout immediately before Output | warn | ordering | null | https://neurarch.com/r/dropout-immediately-before-output | null | {
"trigger": "A dropout layer is the last node before the output node.",
"why": "In training this randomly zeroes the final logits fed to the loss; at eval it is a no-op, so train and eval behaviour diverge. Dropout belongs before the final projection (Linear/Conv), not after it.",
"source": "Srivastava et al. 20... |
R22 | Conv stride larger than kernel | warn | structure | null | https://neurarch.com/r/conv-stride-larger-than-kernel | null | {
"trigger": "A downsampling convolution (conv1d/2d/3d, depthwise, or separable) has stride > kernelSize.",
"why": "The kernel jumps past its own footprint each step, so a band of the input is never read, a silent loss of information. Non-overlapping patches use stride == kernel (e.g. ViT 16/16); stride > kernel is... |
R23 | Non-spatial tensor into Conv | warn | structure | null | https://neurarch.com/r/non-spatial-tensor-into-conv | null | {
"trigger": "A flatten or linear layer connects directly into a (non-transpose) convolution.",
"why": "Mirror of R18: a Conv needs a [channels, …spatial] feature map, but Flatten / Linear emit a flat [batch, features] vector. The forward pass raises a shape error unless the spatial dims are rebuilt with a Reshape ... |
R24 | Back-to-back normalization | info | pattern | null | https://neurarch.com/r/back-to-back-normalization | null | {
"trigger": "A normalization layer connects directly into another normalization layer (e.g. LayerNorm → BatchNorm).",
"why": "Normalizing an already-normalized tensor is redundant; the second layer mostly re-centres/re-scales what the first produced and just burns its own learnable parameters.",
"source": "Idemp... |
R25 | Duplicate positional encoding | info | pattern | null | https://neurarch.com/r/duplicate-positional-encoding | null | {
"trigger": "More than one positional-encoding layer (positionalEncoding, learnedPositionalEmbedding, rope, alibi) in the model.",
"why": "Position is normally injected once. Stacking absolute + rotary, or two of the same, double-counts position and tends to hurt more than help.",
"source": "Positional-encoding ... |
R26 | Pooling feeds Linear with no flatten | warn | structure | null | https://neurarch.com/r/pooling-feeds-linear-with-no-flatten | null | {
"trigger": "A spatial pooling layer (maxpool, avgpool, adaptive pool) connects directly into a linear layer. Global pools are exempt: they already collapse spatial dims.",
"why": "Sibling of R18: pooling still emits a [channels, …spatial] map, but Linear expects a flat [batch, features] tensor, so the forward pas... |
R27 | Flatten before attention | warn | structure | null | https://neurarch.com/r/flatten-before-attention | null | {
"trigger": "A flatten layer connects directly into an attention layer.",
"why": "Flatten collapses the sequence dimension into one long vector, leaving a length-1 sequence, so attention has nothing to relate. Keep the [sequence, dim] layout and flatten only after the attention stack.",
"source": "Attention oper... |
R28 | ConvTranspose checkerboard risk | info | pattern | null | https://neurarch.com/r/convtranspose-checkerboard-risk | null | {
"trigger": "A ConvTranspose2d has kernelSize not divisible by stride (with stride > 1).",
"why": "Uneven kernel overlap during upsampling deposits more weight on some output pixels than others, producing checkerboard artifacts. Make kernelSize a multiple of stride, or upsample with Upsample + Conv (resize-convolu... |
R29 | GroupNorm channels not divisible by numGroups | block | structure | null | https://neurarch.com/r/groupnorm-channels-not-divisible-by-numgroups | compute-error | {
"trigger": "A GroupNorm layer's channel count is not an exact multiple of numGroups.",
"why": "GroupNorm splits channels into equal groups; a non-divisible count raises at construction. Parallel to the GQA / attention head-dim divisibility checks. Set numGroups to a divisor of the channel count.",
"source": "Wu... |
R30 | Very large Linear layer | warn | performance | null | https://neurarch.com/r/very-large-linear-layer | null | {
"trigger": "A single Linear exceeds ~1B parameters (inFeatures × outFeatures).",
"why": "A dense layer that large (~4 GB float32) almost always means a feature map was flattened without pooling first. Add a Global Average Pool / more downsampling, or factorize the layer. Embedding and vocab-projection heads are t... |
R31 | Full multi-head attention at LLM scale | info | performance | null | https://neurarch.com/r/full-multi-head-attention-at-llm-scale | null | {
"trigger": "A transformer stack (≥6 attention layers) at LLM width (embedDim ≥ 2048) uses full multi-head attention everywhere, with no grouped-query or latent attention present.",
"why": "At real model scale the KV cache, not the weights, dominates serving memory. Full per-head K/V caching is what production LLM... |
R32 | Default init assumes ReLU, saturating activation follows | info | pattern | null | https://neurarch.com/r/default-init-assumes-relu-saturating-activation-follows | null | {
"trigger": "A Linear/Conv layer feeds a sigmoid or tanh activation directly.",
"why": "PyTorch's default Kaiming (He) init is derived for ReLU-family activations. Feeding a saturating activation from a He-initialized layer starts training in the saturated tails and shrinks early gradients. Use Xavier init with th... |
R33 | Deep attention stack without depth-scaled init | info | pattern | null | https://neurarch.com/r/deep-attention-stack-without-depth-scaled-init | null | {
"trigger": "≥ 8 attention layers stacked in one model.",
"why": "Residual-branch outputs add up with depth; unscaled init lets activation variance grow layer over layer. GPT-2/LLaMA-family models scale residual output projections by depth: N(0, 0.02 / √(2L)).",
"source": "GPT-2 (Radford et al. 2019) init ... |
R34 | LM head width disagrees with embedding vocab | info | structure | null | https://neurarch.com/r/lm-head-width-disagrees-with-embedding-vocab | null | {
"trigger": "The model has an embedding with a numeric vocabSize and attention layers, but the final Linear feeding the Output projects to a different width.",
"why": "A language model's head must project back to vocab size (and is usually weight-tied to the embedding); a mismatch means the model cannot emit token... |
R35 | KV cache exceeds serving budget at reference context | warn | performance | null | https://neurarch.com/r/kv-cache-exceeds-serving-budget-at-reference-context | null | {
"trigger": "Total fp16 K/V cache across all attention layers (GQA-aware: counts numKVHeads, skips MLA) exceeds 4 GB for a single 8,192-token sequence.",
"why": "That much cache for ONE sequence is gone before weights or activations load; on a 24 GB GPU it caps concurrency at a handful of requests. Raise the GQA r... |
Neurarch structural check catalogue
The 41 structural checks Neurarch runs on a model graph, as data: id, severity, category, the condition that triggers it, why it costs something, and the fix. These are the checks that run in the editor as a design is built, in CI through the GitHub Action, and over the wire at POST /api/v1/check, so the catalogue is the vocabulary any of those three surfaces grades in.
- Size: 41 checks
- Licence: Creative Commons Attribution 4.0
- Canonical page: https://neurarch.com/d/checks.html
- On the site: https://neurarch.com/rules.html
What the viewer shows
The viewer shows one row per check. checks.json is the same content with the index wrapper, which is the shape the API returns.
Fields
- check id
- severity
- category
- trigger condition
- provenance
What this dataset is not
Severity is our editorial judgment, not a measured error rate. 3 of the 41 carry a non-null provenance field naming the study that grounds them; the rest follow from the shape algebra or from convention, and the entry says which. Reading the whole catalogue as measured would overstate it by 38 checks.
Verify a design of your own
These rows describe neural network graphs that were checked by a deterministic verifier, and the same verifier is callable:
curl -X POST https://www.neurarch.com/api/v1/check \
-H "Authorization: Bearer $NEURARCH_API_KEY" \
-H "Content-Type: application/json" \
-d @graph.json
Cite
Neurarch. Neurarch structural check catalogue. https://neurarch.com/d/checks.html
- Downloads last month
- 38