std.random

Status: Slice 45m (closed 2026-07-11). Two tools; draws are traced and substituted under replay, so a program that drew 0.42 draws 0.42 again on every re-run.

Quick reference

import "./std/random" use random_float, random_int

agent roll() -> Int:
    return random_int(1, 6)

The design decision

Randomness is a tool, never a builtin — the same decision as std/time and for the same reason. Tool calls are traced and substituted under replay; the checker rejects tool calls inside @deterministic bodies. A "deterministic" agent that secretly rolls dice is a compile error, not a production incident.

The two tools

random_float() -> Float uses rand_draw

Uniform draw in [0.0, 1.0) from OS entropy (53 uniform bits — the standard f64 recipe).

random_int(min: Int, max: Int) -> Int uses rand_draw

Uniform integer in [min, max] — INCLUSIVE on both ends (Python's randint contract: random_int(1, 6) is a die). Uses rejection sampling, so there is no modulo bias. min > max is a runtime error.

Non-scope

No seeded/reproducible PRNG surface — reproducibility in Corvid comes from replay, not from seed management. No distributions (gaussian, exponential); compose them from random_float or ship a follow-up slice if agent workloads demand them.