The Mira Playground · falsify-eval

1 · Pick a student strategy

Real studentalways knows the answer (100% accurate)

Miraalways writes "C" — never studied

Slackerpicks one of A/B/C/D/E at random

Popularity-onlypicks each label proportional to its frequency

Mostly studiedaccuracy slider below — partial real knowledge

Studied accuracy p: 70%

3 · Live results seed = 2026

Student score

—

vs the four nulls (Δ = student − null)

Δ_A · gold permuted

—

Δ_B · uniform random

—

Δ_C · random retrieval

—

Δ_D · marginal-matched NEW

—

GATE: PASS Computing…

The numbers above run the four-null gate algorithm in your browser, exactly the way the Python library does on a real benchmark — just with a tiny synthetic exam instead of 3,633 documents. Δ_D (marginal-matched random) is the new contribution of this work; the other three nulls are standard. Try Mira first, then the Studied slider, and watch which Δ collapses for which strategy.

Pick a strategy. Beat the four nulls.

1 · Pick a student strategy

2 · Test setup

3 · Live results seed = 2026