[alikahwaji]

Phase 4a of the redesign: the two cards designed in the mockups next to Regularisation. Based directly on main.

Overfitting risk

• A Low / Moderate / High meter comparing trainable parameters with training points (e.g. 65 parameters · 160 training points).
• Noisy data (> 30%) raises the risk by one band. Any L1, L2 or dropout lowers it by one.
• Shows a one-line explanation and a suggestion that points to the Regularisation settings above it.
• Updates live as layers, classes, samples, noise, validation split or regularisation change, and when data loads. Before any data is loaded, it estimates the training set from the sliders.
• Level and meter come from one effective ratio, so the fill always sits inside the band its label names (covered by a property test).
• The Regularisation section now starts open, because the card's suggestion refers to it.

Reproducibility

• Seed field with a new-seed button, a weight-initialiser choice (He normal, Glorot uniform, LeCun normal) and a Lock seed checkbox.
• New core/domain/Random.ts: a seedable Mulberry32 stream that now drives dataset generation, the train/validation split and batch order. When unseeded it falls back to Math.random, so existing test spies still work.
• TF.js initialisers get a per-layer seed, so layers of the same shape don't start with identical weights.
• Locked: the same seed repeats a run exactly. Unlocked: each dataset load draws a new seed, so the seed shown always describes the current run, and you can lock it after a run you liked.
• Dropout is deliberately left unseeded: a seeded TF.js dropout reuses one mask on every batch, which would quietly turn dropout into a fixed, smaller network. The card's hint says so.
• Seed and initialiser count toward the stale-model badge. An invalid seed is rejected with a validation message instead of silently falling back to random.

Testing

• Unit: new suites for Random (including repeatable shuffles and datasets), OverfittingRisk, seeded TFNeuralNet initialisation, ReproducibilityPanel and OverfittingRiskCard. Full suite: 1,217 passed.
• Browser check (CI build):
  • Seed 1234 locked, spiral, 8 epochs, run twice with a page reload in between: identical point positions, loss, validation loss and accuracy.
  • Seed 999 gives a different run.
  • Unlocked runs drew different seeds.
• E2E (CI mode): 73 passed, 5 skipped. Visual baselines unchanged.
• Screenshots checked at desktop and phone sizes. tsc and eslint are clean.

🤖 Generated with Claude Code