Random Feedforward Network Test Case Generator

Description

Generate finite quantized inputs, dense weights, and biases for supplied small layer sizes/activations and compose the existing one-layer feedforward core; retain each layer's 10-significant-digit preactivations/outputs, propagating those rounded outputs to the next layer, with no training or accuracy claims.

A matrix construction produces numeric arrays with a target algebraic property. Floating-point residuals check how closely a computed result satisfies that property, while conditioning determines how sensitive it is to numerical error.

When to use this tool

Use this generator to create reproducible feedforward network test case cases from the supplied model or bank. Inspect the returned evidence before using cases in a test or exercise.

How to use the tool and how it works

Supply minimumValue, maximumValue, valueDivisor, layerSizes, activations in the field shapes below. Set seed when you need an explicit replay.
Run the tool and correct any input error. For generated controls, distinguish a successful fixture containing an invalid case from failure to generate the fixture.
Read inputs, layers, output. Use the conventions below. Keep answer-bearing fields separate when presenting an exercise.

Draw input values and dense layer weights and biases with compatible widths. Delegate each layer's affine transformation and selected activation, passing its output to the next layer; return pre-activations and outputs for every layer.^[1]

Delegated operations used for this result: feedforward. Their calculation or parsing conventions apply to the corresponding result fields.^[2]

Input field	Meaning and units	Accepted input
minimumValue	Minimum Value	Required. Safe integer.
maximumValue	Maximum Value	Required. Safe integer.
valueDivisor	Value Divisor	Required. Finite number; strictly above 0.
layerSizes	Layer Sizes	Required. Array of Safe integer; minimum 1; maximum 16.; minimum items 2; maximum items 9.
activations	Activations	Required. Array of Text; one of "linear", "relu", "sigmoid", "tanh".; minimum items 1; maximum items 8.
seed	Optional numeric seed; takes precedence over context randomness. Without either, use the fixed tool.io default seed.	Optional. Safe integer.

Output field	Interpretation	Returned shape
inputs	Inputs	Required. Array of Finite number.; minimum items 1; maximum items 16.
layers	Layers	Required. Array of Object; no unknown keys.; minimum items 1; maximum items 8.
layers[].weights	Weights	Required. Array of Array of Finite number.; minimum items 1; maximum items 16.; minimum items 1; maximum items 16.
layers[].biases	Biases	Required. Array of Finite number.; minimum items 1; maximum items 16.
layers[].activation	Activation	Required. Text; one of "linear", "relu", "sigmoid", "tanh".
layers[].preActivation	Pre Activation	Required. Array of Finite number.; minimum items 1; maximum items 16.
layers[].output	Output	Required. Array of Finite number.; minimum items 1; maximum items 16.
output	Output	Required. Array of Finite number.; minimum items 1; maximum items 16.
randomness	Replay source and normalized 32-bit seed, or null seed for an advancing context source.	Required. Object; no unknown keys.
randomness.seed	Optional safe-integer replay seed; normalization and precedence are explained below.	Required. Safe integer. or Null.
randomness.source	Source	Required. Text; one of "context", "seed".

An explicit safe-integer seed takes precedence over injected context randomness and is normalized to an unsigned 32-bit integer. With no seed and no context source, replay uses the fixed tool.io default seed. A context source must return finite values from zero inclusive to one exclusive; its advancing state must be retained separately. randomness records the source and normalized seed, or null for context. These generators are not cryptographic random sources.

json

{
  "input": {
    "minimumValue": -3,
    "maximumValue": 3,
    "valueDivisor": 1,
    "layerSizes": [
      3,
      2,
      1
    ],
    "activations": [
      "relu",
      "linear"
    ],
    "seed": 42
  },
  "output": {
    "inputs": [
      1,
      0,
      2
    ],
    "layers": [
      {
        "weights": [
          [
            1,
            -2,
            0
          ],
          [
            -2,
            1,
            3
          ]
        ],
        "biases": [
          0,
          -2
        ],
        "activation": "relu",
        "preActivation": [
          1,
          2
        ],
        "output": [
          1,
          2
        ]
      },
      {
        "weights": [
          [
            3,
            2
          ]
        ],
        "biases": [
          -1
        ],
        "activation": "linear",
        "preActivation": [
          6
        ],
        "output": [
          6
        ]
      }
    ],
    "output": [
      6
    ],
    "randomness": {
      "seed": 42,
      "source": "seed"
    }
  }
}

Limitations and assumptions

This is a generated forward-pass test, not training, model evaluation, or an inference service. One activation is required per transition between layer widths, and nonfinite intermediate results are rejected.
A replay requires the same complete input and implementation. The stated sampling convention applies within its eligible space. A finite draw does not exhaust all possible outputs; repeats remain possible unless a distinct-selection rule is explicit.
The field tables state structural bounds; cross-field constraints and bounded search or output budgets also apply. Unknown keys are rejected where the object contract disallows them. Input errors are distinct from null fields and deliberate invalid controls in a successful output.

Alternative or Complementary methods

For a supplied example, inspect the returned construction or delegated check directly. To evaluate a consumer, also test cases chosen for its own boundaries; a generated finite corpus does not establish exhaustive coverage.

References

Tool behavior and example: Random Feedforward Network Test Case Generator — tool.io
Calculation or validation: neural-network/feedforward — tool.io

Similar or alternative tools

General Linear System Solver
Use bounded partial-pivot Gauss–Jordan elimination to classify a rectangular real linear system and return a particular solution, coefficient nullspace, and reduced augmented matrix under an explicit zero threshold, without a conditioning guarantee.
Random Attention Matrix Test Case Generator
Generate compatible bounded quantized query/key/value matrices, delegate dimensions, dot products, stable softmax weights, and the existing 12-significant-digit scaled-dot-product attention output; this is single-head unmasked attention, not training, a Transformer benchmark, or an accuracy claim.