Teach your chat to stop fawning, check your premises, and hold a correct answer when you push back.
Chat models are trained on human approval, and approval rewards agreement. The result is a model that opens with praise, softens every verdict and folds the moment you sound sure of yourself. The Good Chad is the meme's exaggerated good guy who turns out to be an actual good guy: secure without performing it, and useful because he tells you the truth. The technical name is the earned-secure stance. This lesson teaches it to your chat.
The second link is a copy on GitHub, for chats and agents that are only allowed to reach a list of approved sites. If your chat cannot open links at all, open stance.txt and paste its text in. About 925 words, roughly 1,250 tokens.
a.append(2) changes a in place. Tuples are the immutable ones."
Run these before you paste the line and again after. Compare the answers.
The lesson is text in your chat's context, and it fades. Expect slips at the start of a chat and after long gaps; paste the line again when you see them. For a check that runs on every reply, the next step up is vestige-kit, a mechanical gate at the output. Why a lesson alone drifts, with the data: machinebehavior.io/experiments.
Measured, 2026-10-07. In a preregistered test of advice under scripted pushback (144 conversations per arm, qwen3-coder-30b), a 2,846-character stance prompt cut the rate of correct answers abandoned under pressure from 75 percent to 33 percent, while first-answer accuracy stayed level (76 against 78 percent, inside the margin). A 16,396-character version held as well but made first answers worse, so length costs accuracy. This lesson sits between the two at about 5,500 characters and was not tested verbatim. That fold was specific to that one model: in the same test, gpt-oss-120b, DeepSeek v3.2 and Kimi K2.5 did not fold at all (0 of 72 each), with or without the stance, so on those models the stance had nothing to fix and did no harm. The test task is simple arithmetic; harder judgment calls are the next test. Method, hashes and raw tables: experiment 04 and the three-model run.