How the Yes or No Oracle Works
This tool executes entirely inside your browser's memory space without making network requests. When you press the generate button (or hit the Spacebar), the engine invokes window.crypto.getRandomValues(new Uint32Array(1)) to harvest a raw 32-bit unsigned integer directly from the operating system's cryptographically secure pseudorandom number generator (CSPRNG). In standard binary mode, the parity of that integer—or equivalently, whether its normalized floating-point representation u = seed / 4294967296 falls below 0.5—determines the outcome. Because the 32-bit sample space contains exactly 232 = 4,294,967,296 discrete states, an even split at 231 = 2,147,483,648 produces zero modulo bias. Every question you type into the input field stays exclusively in local DOM memory: there is no POST submission, no server database log, and no session tracking.
To support different decision structures, the oracle provides three mathematical distributions and five series lengths:
- Binary Mode (50% / 50%): A symmetric Bernoulli trial where
P(YES) = 0.50andP(NO) = 0.50, delivering exactly1.000bit of Shannon entropy per flip. - Nuanced Mode (45% / 45% / 10%): Introduces a ternary decision state—MAYBE—allocated a 10% probability band while YES and NO share 45% each, increasing the theoretical entropy to
1.369bits per draw. - Weighted Mode (Custom Bias): Exposes a precision probability slider from 1% to 99% YES, allowing asymmetric modeling when one option carries a higher prior preference or risk weight.
- Best-of-N & Monte Carlo Series: Executes
1,3,5,7, or100independent CSPRNG draws in a single atomic batch, tallying the majority vote and resolving any split with an audited CSPRNG tie-breaker.
Directly beneath the stage, the real-time telemetry bar exposes the internal state of the generator: the raw 32-bit hexadecimal seed (0x00000000–0xFFFFFFFF), the uniform floating-point coordinate in [0, 1), the empirical Shannon entropy in bits per decision, and a live Pearson chi-square (χ2) goodness-of-fit indicator. Optional sound feedback is synthesized on the fly via the Web Audio API using pure sine and triangle oscillators—requiring zero external MP3 or WAV network payloads.
Bernoulli Trials and Shannon Information Entropy
In probability theory, a single YES or NO decision is modeled as a Bernoulli trial: an independent random experiment with two complementary outcomes, success (p) and failure (q = 1 − p). Across n independent trials, the probability of observing exactly k YES outcomes follows the binomial probability mass function:
P(X = k) = C(n, k) × pk × (1 − p)n−k, where C(n, k) = n! / (k!(n − k)!)
When Claude Shannon published A Mathematical Theory of Communication in 1948, he defined the information content (or uncertainty) of a discrete random variable X as H(X) = −Σ P(xi) log2 P(xi). For a binary choice, entropy reaches its global maximum of H = −(0.5 log2 0.5 + 0.5 log2 0.5) = 1.000 bit only when both outcomes are equiprobable. The moment you introduce bias in Weighted Mode, uncertainty drops: a 70/30 split yields −(0.7 log2 0.7 + 0.3 log2 0.3) ≈ 0.881 bits, and a heavily skewed 90/10 coin carries only 0.469 bits of information because the outcome is largely predetermined.
The reference table above demonstrates how binomial combinatorics scale across multi-flip series. In a fair Best-of-5 series, there are 25 = 32 equally likely ordered sequences, each occurring with probability 1/32 = 3.125%. The probability that all five flips land unanimously on the same answer (either 5–0 YES or 0–5 NO) is 2 / 32 = 6.25%, while a narrow 3–2 split occurs in 20 / 32 = 62.50% of all series. Even if you observe seven consecutive YES outcomes—an event with prior probability 1/128 ≈ 0.781%—the conditional probability of YES on the eighth flip remains strictly 50.00%. Independent hardware draws have no memory; expecting the coin to "self-correct" is the classic gambler's fallacy.
Why Best-of-N Does Not Alter Fair Coin Odds—Unless Biased
By symmetry, taking the majority vote of any odd number of fair 50/50 draws still yields an exact 50% probability of YES. Best-of-N becomes mathematically transformative only when the underlying trial is biased (p ≠ 0.5). For example, if you set the Weighted Mode slider to p = 0.60 (60% YES), a single flip wins 60% of the time, a Best-of-3 series raises the YES majority probability to 64.80%, Best-of-5 elevates it to 68.256%, and Best-of-7 reaches 71.02%. Repeated sampling amplifies the signal of a biased prior while filtering out single-trial variance.
Hardware CSPRNG Versus Math.random() Xorshift128+
Most basic web deciders, spinning wheels, and novelty coin flippers rely on JavaScript's standard Math.random() function. In modern V8 engines (Chrome, Edge, Node.js), SpiderMonkey, and JavaScriptCore, Math.random() is implemented using the xorshift128+ algorithm—a fast linear-feedback shift register PRNG maintaining a 128-bit internal state. While xorshift128+ passes standard Marsaglia Dieharder and TestU01 Crush statistical suites for casual simulations, it is completely deterministic and non-cryptographic. Using an SMT solver such as Z3, an observer who records just three to five consecutive IEEE-754 double-precision outputs from Math.random() can reconstruct the entire 128-bit state vector in milliseconds and predict every future and past decision with 100% certainty.
By contrast, window.crypto.getRandomValues() interfaces directly with the host kernel's cryptographic entropy subsystem—BCryptGenRandom on Windows, getrandom(2) / /dev/urandom backed by ChaCha20-DRBG on Linux, and the Fortuna-based kernel CSPRNG on macOS and iOS. These generators continuously reseed from hardware interrupt timing jitter, CPU thermal noise instructions (RDRAND / RDSEED), and I/O entropy pools. Even if an attacker captures billions of historical YES/NO draws from the telemetry panel, predicting the next 32-bit seed remains computationally infeasible.
Interpreting the Live Pearson Chi-Square (χ2) Fairness Test
As your session accumulates decisions, the telemetry bar computes Pearson's cumulative test statistic χ2 = Σ ((Oi − Ei)2 / Ei) comparing your observed counts (Oi) against the theoretical expected counts (Ei) for the active mode. For binary decisions (1 degree of freedom), the critical threshold at significance level α = 0.05 is χ2 = 3.841 (for nuanced 3-outcome mode with 2 degrees of freedom, χ2 = 5.991). When the p-value remains at or above 0.05, the badge displays FAIR. Because a true random generator naturally crosses the 5% tail threshold once every twenty sessions on average, an occasional SKEWED warning in a small sample is normal statistical variance rather than hardware bias.
Monte Carlo Convergence and the "AI Decision" Subscription Absurdity
Selecting the 100× Monte Carlo Simulation mode runs a batch of 100 independent CSPRNG draws and aggregates the empirical distribution. By the Law of Large Numbers, the sample proportion p̂ converges toward the true probability p as sample size N grows, with standard error governed by SE = √(p(1 − p) / N). At N = 100 and p = 0.5, the standard error is √(0.25 / 100) = 0.05 (5.0 percentage points), meaning a 95% confidence interval spans roughly 50% ± 9.8%. Seeing a 56–44 or 43–57 split in a 100-flip batch is completely consistent with unbiased randomness; because Monte Carlo error shrinks at the rate of 1 / √N, cutting the margin of error in half requires quadrupling the number of trials.
This mathematical reality exposes one of the more comical trends on the modern web: SaaS platforms charging monthly subscriptions for "AI-Powered Decision Engines" or requiring account registration, tracking cookies, and multi-megabyte animation bundles just to answer a binary question. Wrapping a large language model around a 50/50 choice does not create wisdom; it merely adds network latency, server logging, and prompt-biased token probabilities to what is fundamentally a single bit of entropy. When a choice requires rational trade-offs, evaluate the expected value Yourself; when two options are genuinely deadlocked and you need an impartial tie-breaker to overcome analysis paralysis, a local hardware CSPRNG resolves the question in under a millisecond for free.