Convergence Gate-Check: interactive demo

Illustrative, not empirical. All inputs are assumptions from the paper's benchmark, and the paper is AI-written with no recorded human review. This demo shows the maths, not a validated detector.

1. Try it

Prior belief the claim is true before asking any model: Chance two models agree when the claim is true p_c: Chance two models agree when the claim is false p_b: Shared training bias ρ, makes false claims look agreed-on: Models in the panel:

2. What the agreement is worth

Bayes factor per added model

posterior odds

models needed for 95%

two-model posterior

Solid: your settings. Dashed: the paper's benchmark. Green line: 95% target.

3. Does the maths hold? Simulate it

Draw 200,000 imaginary claims. Each is true with your prior. Models agree with the chances above. Of claims where all models agreed, what share was true?

Not run yet.

This checks the formula against its own assumptions. It cannot tell you whether real models behave this way.

4. How fragile is the answer?

Two-model posterior as the false-agreement chance changes (p_c and prior from your sliders).

5. Partial agreement (paper §E.2)

An argument has S steps; per-step agreement is q_c if true, q_b if false.

Steps in the argument: Steps the two models matched: q_c: q_b:

6. Reproduces the paper

Limits

Models share training data, so agreement is partly manufactured. The shared-bias slider shows how little it takes to erase the gain.
The Bayes factors come from one illustrative benchmark, not measured data. Technology readiness level 1–2.
Agreement on retrieval questions carries no evidence (factor 1).