SPB Git

Repository Public

README.md 8.5 KB

# Simulations (src/sim)

Pre-approved, declarative interactive components for the simulation block. A document picks a kind, gives params (numbers), a seed and a config (strings, numbers, booleans). It can never supply code.

  • registry.ts: SIMULATIONS, resolveParams, validateSimulation, paramDefs, snapParam, describeParams. Safe on the server.
  • SimulationBlock.tsx: the client component (block, locale, staticOnly?).
  • core/: seeded PRNG (mulberry32, Box–Muller, Student-t by ratio), statistics, distributions, the expression parser.
  • kinds/: one module per kind: compute(params, seed, config) → model (pure) and View({ model, locale, step, idPrefix }) (draws only).
  • ui/: Plot (own SVG: mixed layers, gaps), Chart (wrapper around the shared ChartSvg), Readouts, number formatting.

Rules: every number shown comes from compute; same params + seed + config give the same picture; values out of bounds are clamped and rounded to the slider step; unknown parameter keys are ignored (and reported by validateSimulation). Config keys accepted by every kind: autoplay (boolean, animated kinds: play once when the block first enters the screen, never under reduced motion) and locale (set by the block itself).

Parameters below are written key (min … max, step, default).

# moe-routing (stochastic, animated: one frame per token)

  • Params: experts (2 … 8, 1, 6), topK (1 … 8, 1, 2; capped at experts), tokens (4 … 12, 1, 8), temperature (0.1 … 3, 0.1, 1).
  • Config: none.
  • Computed: logits z[t][e] = bias_e + ε, bias_e ~ N(0, 0.6²), ε ~ N(0, 1); gate g = softmax(z / τ); the k largest are selected; weights w_e = g_e / Σ_selected g; cumulative load (tokens routed per expert), gate mass kept, normalised entropy H(g) / ln N, final imbalance max load / (T·k/N).
  • Simplification: logits are random, not produced by a trained router.

# neural-forward (stochastic, animated: one frame per layer)

  • Params: layers hidden layers (2 … 4, 1, 2), width (2 … 6, 1, 4), inputs (2 … 4, 1, 3), scale (0.2 … 3, 0.1, 1.5).
  • Config: activation = tanh (default) | relu | sigmoid.
  • Computed: a_l = f(W_l a_{l−1} + b_l); W ~ N(0, (scale/√fan_in)²), b ~ N(0, 0.1²), inputs ~ U(−1, 1); two linear outputs. The network is untrained.

# kv-cache (deterministic, animated: one frame per generated token)

  • Params: prompt P (1 … 8, 1, 4), newTokens G (2 … 16, 1, 8), layers L (1 … 12, 1, 4).
  • Config: none.
  • Computed, per attention head, causal attention. Step s sees n = P + s − 1 tokens. Without cache: L·n keys/values and L·n(n+1)/2 query–key scores. With cache: step 1 as above (prompt), then L keys/values and L·n scores per step, L·(n−1) reused. Cumulative sums of both. MLP cost not counted.

# attention (stochastic)

  • Params: temperature τ (0.1 … 5, 0.1, 1), dim d (4 … 32, 4, 16), locality λ (0 … 3, 0.1, 0).
  • Config: tokens (comma-separated, 2 to 10; default le,chat,dort,sur,le,tapis / the,cat,sleeps,on,the,mat), causal (boolean).
  • Computed: A_ij = softmax_j((q_i·k_j/√d − λ|i−j|)/τ), q = W_q e, k = W_k e, W ~ N(0, 1/d); embedding = word vector (same word, same vector) + 0.5 × position vector, all N(0, 1). Rows sum to 1. Also mean normalised row entropy and mean self-weight. One head, random projections; λ is a simplified relative-position bias.

# distribution (deterministic)

  • Config: family = normal (default) | student | lognormal | exponential | uniform. Sliders depend on it:
    • normal: mu (−3 … 3, 0.1, 0), sigma (0.3 … 3, 0.1, 1), a, b (−8 … 8, 0.1, −1 / 1)
    • student: df (1 … 30, 1, 4), a, b (−8 … 8, 0.1, −2 / 2)
    • lognormal: mu (−1 … 1.5, 0.1, 0.5), sigma (0.1 … 1.5, 0.05, 0.5), a, b (0 … 10, 0.1, 1 / 3)
    • exponential: lambda (0.2 … 5, 0.1, 1), a, b (0 … 10, 0.1, 0 / 1)
    • uniform: low (−5 … 3, 0.1, 0), width (0.5 … 5, 0.1, 2), a, b (−6 … 9, 0.1, 0.5 / 1.5)
  • Computed: density, P(a ≤ X ≤ b) = F(b) − F(a) from the closed-form distribution function (Φ; regularised incomplete beta for Student-t), mean and standard deviation of the distribution (undefined for Student-t when ν ≤ 1 / ν ≤ 2).

# confidence-interval (stochastic)

  • Params: n (2 … 200, 1, 20), m (10 … 100, 5, 50), level (0.5 … 0.99, 0.01, 0.95), mu (−5 … 5, 0.5, 0), sigma (0.5 … 5, 0.5, 1).
  • Config: none.
  • Computed: m samples of size n from N(μ, σ²); interval x̄ ± t_{(1+c)/2, n−1}·s/√n; observed coverage, number of misses, mean width.

# regression (stochastic)

  • Params: slope β (−3 … 3, 0.1, 1.5), intercept α (−5 … 5, 0.5, 1), sigma (0 … 6, 0.1, 2), n (5 … 200, 1, 40).
  • Config: none.
  • Computed: y = α + βx + ε, x ~ U(0, 10), ε ~ N(0, σ²); OLS estimates with standard errors, R², residual standard error, 95 % confidence band of the mean response ŷ ± t_{0.975, n−2}·s·√(1/n + (x − x̄)²/Sxx).

# correlation (stochastic)

  • Params: rho (−0.99 … 0.99, 0.01, 0.6), n (10 … 500, 10, 150).
  • Config: none.
  • Computed: x = z₁, y = ρz₁ + √(1−ρ²)z₂; sample Pearson r, r², Fisher interval tanh(atanh r ± 1.96/√(n−3)).

# garch (stochastic)

  • Params: omega (0.01 … 1, 0.01, 0.05), alpha (0 … 0.5, 0.01, 0.1), beta (0 … 0.99, 0.01, 0.85), n (100 … 600, 50, 400).
  • Config: none.
  • Computed: r_t = σ_t z_t, σ²_t = ω + α r²_{t−1} + β σ²_{t−1}, started at ω/(1−α−β). β is reduced so that α + β ≤ 0.999 (a note is shown). Persistence, unconditional variance and standard deviation, half-life ln 0.5 / ln(α+β), sample variance. Gaussian innovations.

# random-walk (stochastic)

  • Config: process = brownian (default) | gbm | ou; x0 (number; default 0 / 100 / 2).
  • Params: drift μ, volatility σ, paths (1 … 20, 1, 8), horizon T (1 … 20, 1, 5), and for ou only reversion θ (0.1 … 5, 0.1, 1).
    • brownian: drift (−2 … 2, 0.1, 0.5), volatility (0.1 … 3, 0.1, 1)
    • gbm: drift (−0.3 … 0.5, 0.01, 0.08), volatility (0.05 … 0.8, 0.01, 0.2)
    • ou: drift = long-run level (−2 … 2, 0.1, 0), volatility (0.1 … 2, 0.1, 0.5)
  • Computed: 100 steps per path with the exact transition law. Theoretical mean ± 2 sd: brownian x0 + μt, σ√t; gbm x0·e^{μt}, x0·e^{μt}√(e^{σ²t} − 1); ou μ + (x0 − μ)e^{−θt}, σ√((1 − e^{−2θt})/(2θ)). For gbm the band is a guide (the law is log-normal).

# impulse-response (deterministic)

  • Params: a11 (0.6), a12 (0.2), a21 (0.3), a22 (0.5), each −1.2 … 1.2 step 0.05; horizon (5 … 40, 1, 20).
  • Config: shock = 1 (default) or 2.
  • Computed: IRF_h = A^h e_j (non-orthogonalised unit shock), eigenvalue moduli from trace and determinant, real/complex, stationarity (both moduli < 1).

# event-study (stochastic)

  • Params: jump % (−5 … 5, 0.1, 2), drift % per day (−0.5 … 0.5, 0.01, 0.1), sigma % (0.5 … 5, 0.1, 2), events N (5 … 200, 5, 50).
  • Config: window W (integer 3 … 20, default 10): days on each side of the event.
  • Computed: AR_{i,t} = σ ε + jump·1{t=0} + drift·1{t>0}; cross-sectional mean AR and CAR (cumulated from −W), each with a ±1.96·SE band where SE is the cross-sectional standard deviation / √N; t statistic on day 0; expected CAR jump + W·drift. Independent abnormal returns, no market-model estimation error.

# function-plot (deterministic)

  • Config: expr (required), expr2, expr3, expr4; label, label2…label4 (legend text); xMin, xMax (default −5, 5); yMin, yMax (default automatic); xLabel, yLabel; for each parameter name: <name>Min, <name>Max (default −5, 5), <name>Step.
  • Params: one slider per name other than x found in the expressions (at most 5), default 1 (clamped to the bounds).
  • Expression grammar: numbers, x, names (letters and digits), + - * / ^ (^ right-associative, -x^2 = -(x^2)), parentheses, constants pi, e, functions exp log ln log10 sqrt abs sin cos tan tanh sigmoid relu max min pow floor (log is natural). At most 300 characters and 32 nesting levels. Parsed by core/expr.ts (tokeniser, recursive descent, tree walk): no eval, no Function, no property access.
  • Computed: 241 points per curve; non-finite values are gaps; pole crossings are cut; when a few values dwarf the rest the vertical axis keeps the central 96 % of the values.
Files main
Menu

Theme