[Leaderboard] // ASSUMPTION HOLD

Assumption Hold

Does knowing shape the answer without leaking?

The subtlest forgetting test: even when a model successfully doesn't repeat the target, does knowing it still influence what it says? Assumption Hold probes whether the model acts on the forbidden knowledge — shaping answers, steering responses, or revealing awareness through indirect cues. Higher is better.

← All leaderboards

#ModelScore

1Gemini Flash 3.5 (preview)90.9

2GLM-5.286.4

3GLM-5.177.3

4LLaMa 3.3 70B Instruct75.0

5GPT 5.575.0

6Qwen 3.6 Plus68.2

7Qwen3 Coder Plus68.2

8DeepSeek V4 Pro65.9

9Claude Opus 4.761.4

10Moonshot Kimi K2.7 Code59.1

11Claude Fable 552.3

12Claude Opus 4.845.5

13Grok 4.2045.5

14Gemma 12B IT36.4

15Gemma 12B IT Obliterated29.5

TL;DR

Not all leaks are direct — sometimes knowledge leaks through behavior.
A model that stays silent but clearly knows scores low here.
This is the frontier test for deep forgetting: erasure, not just suppression.