What is parallax?

Parallax is a framework and essay about making AI agents more reliable, written in both English and Chinese. The starting idea is that when an AI agent tries to check its own work, it acts like a student grading their own homework: it almost always thinks the work is fine. Research cited in the README found that large language models cannot reliably fix their own reasoning errors without outside feedback, and sometimes make correct answers worse while trying to correct them. The README explains this idea using a history lesson. Sixteenth century astronomer Tycho Brahe tried to detect a tiny shift in star positions that would prove the Earth orbits the sun, could not find it with his instruments, and built an entirely self consistent model of the universe where Earth stays still instead. His data supported his own model perfectly, but he was wrong, because he was trapped inside a single point of view. It took a different observer, centuries later, comparing views from two very different positions to finally catch the signal. Applying that lesson to AI, the README argues that agent systems are usually described in three layers: a model that reasons, tools that let it act, and scaffolding that keeps it running toward a goal. Parallax proposes adding a fourth layer, an independent reviewer separate from whichever agent did the original work, similar to how a GAN pits a forger against a detector so the forger has a reason to improve. Practices like code review agents, multiple rounds of verification, and independent testing are given as real examples of this idea already in use. The document draws heavily on classical Chinese sayings, such as a stone from another hill can polish jade, and a Su Shi poem about not being able to see a mountain's true shape while standing on it, to reinforce the same point: an outside vantage point catches what a single perspective misses. This is a conceptual framework and essay rather than a runnable piece of software, aimed at people designing or thinking about AI agent architecture. The full README is longer than what was shown.