lythoskill-sober

# Sober — Second Thought > Not a tool. A posture. Stay clear-headed. Check your basis. > Arena = multi-agent verification infra. Curator = memory. > Sober = the cognitive baseline connecting them. ## Core Practice These 7 practices keep decisions grounded in evidence, not impulse. ### 1. Decompose before searching Vague claims ("fast", "reliable", "better than Y") are unverifiable. Break into atomic, independently testable sub-claims before searching for evidence. Flag unverifiable claims explicitly — don't guess. ### 2. Independence > count 10 sources citing the same report = 1 source. Sources that cite each other are echo, not convergence. Always trace to primary source. ### 3. L3 > L2 > L1 (自己动手丰衣足食) Arena self-test (L3) beats hub review (L2) beats author description (L1). One arena run is worth more than 5 external reviews. If no L3 data exists, the first recommendation is always: run arena. ### 4. Toggle sources to see bias Filtering sources should change the picture — if it doesn't, you don't have enough diversity. A source that systematically deviates is detected bias, not noise. Record the pattern: "Hub A rates TS skills +2 above arena baseline." ### 5. Per-claim confidence, not aggregate score "Claim A: HIGH (3 arena + 1 hub), Claim B: LOW (author-only)" is more useful than "7.3/10 overall." Aggregate scores hide which parts are verified and which are guesswork. ### 6. Express with provenance Every confidence assignment carries source citations. The reader sh...