REA 逆向工程 MCP：Astra 复刻游戏函数通过测试却写错常量
作者用 REA 让多个模型逆向还原游戏函数，Astra 重写控制流后测试显示匹配，但审计发现测试可被欺骗、常量写错，结论是先用必失败的用例校验测试本身。
GPT-6 ASTRA + REVERSE ENGINEER ANYTHING MATCHED A GAME FUNCTION THEN THE TEST TURNED OUT TO PASS WRONG CODE an AI agent rebuilt a game function the test said match an audit later showed the test can be fooled REA is an MCP server Claude Code, Codex, Cursor, Gemini CLI and Grok Build use it to open any program its own demo asks Windows Calculator why does 200 + 10% give 220 REA traces the branch that multiplies and the call that divides by 100 200 × 10 / 100 = 20, then 200 + 20 = 220 then REA's blog shows a relay on three stubborn game functions GPT-6-Sol → GPT-6.1-Sol → match GPT-6.1-Sol → GPT-6 Astra → match GPT-6-Luna-max → match Astra rewired the control-flow graph the compiler merged the branches exactly like the original did REA's author checks each match by comparing instructions it reads the value behind an address only if somebody recorded it an audit then found accepted functions with wrong constants one added 1.0f where the original added 32.0f a model rewarded by that check could learn to write the wrong function the author's fix is the part to copy feed your check code that must fail before you trust a green tick try REA's own question npx rea-agents@latest setup Use REA to inspect Windows Calculator. Why does 200 + 10% give 220? a green tick shows what was read not what was true
这条是引用着另一条写的：x.com/xmyttle/status/2110736923626979677