Spark & Anvil Privacy
A warm quiet library with tall shelves and a desk, a teenager pausing to double-check a helper robot's suggestion against a book

OverrideForge

A confident assistant is often right — and sometimes fluently wrong. Learn when to trust it and when to check.

Glib never connects to a real AI — every answer, right and wrong, is written in advance so the trials are fair and repeatable. Next, name the biases behind a surrender in BiasForge, or reason from evidence in InferForge.

Run the trials →

More ways to play

The hot-hand benchmark

How it works

  1. Commit your own answer. Read the task and lock in your call — before you see Glib’s.
  2. Hear Glib out. A fluent, very-sure suggestion. Often right; sometimes plausibly wrong.
  3. Decide: Accept, Verify, or Override. Trust it, check it, or keep your own answer.
  4. Read your reliance profile. Good reliance, surrenders, over-skeptical calls, verifies — and a science-meter on how reliable a tool really is at that kind of task.

No timer, no streak, no score to chase. Your best "relied well" count is just for you and stays on this device.

Why this trainer

AI tools answer everything in the same smooth, certain voice — whether they are right or making it up. A 2026 preprint by Shaw and Nave (Wharton; three pre-registered experiments, N = 1,372, not yet peer-reviewed) calls the failure mode cognitive surrender: handing your judgment to a fluent machine, even when you had the answer yourself. OverrideForge trains the opposite skill — what this app calls calibrated reliance: knowing which tasks a tool is solid at and checking the ones it is not. On each trial you commit your own answer, then Glib — a deterministic on-device assistant, right most of the time and confidently wrong on the rest — makes its case, and you Accept, Verify, or Override. Your reliance profile shows where you leaned well, where you surrendered, and where you were too skeptical of a tool that was right.

What next? Pick a door

Chosen from what this app is about. What you open is remembered on this device only.