DeepSeek Harness Router Repo Retracts Its Own Theory Paper

A DeepSeek Harness router repo retracted its own theory paper on August 15, keeping the result but voiding the explanation, per its GitHub README.

The repo, yjh051108/dsh-router-standard, hit 292 stars in three days. Its companion, dsh-routing-suite, passed 3,292. Both sell the same idea: a DeepSeek Harness routing layer built on the claim that a model's 'We need' versus 'Let me' outputs mark two distinct reasoning modes.

That claim survives the retraction. What doesn't is the theory behind it. The author marked the paper's A1-A4 dual-attractor hypothesis formally void, along with the 'god/ghost' framing and the claim that self-routing is impossible. The paper stays posted, now with a linked statement and an apology letter attached.

The revised explanation drops the deliberate dual-mode design. The router's 'We need' versus 'Let me' split reads as a fracture between a native deep reasoning path and an under-converged minimal path. That fracture is a post-training defect, not a chosen mode, and the router exploits it as a routing signal.

That's a narrower claim than the one it replaced, and a more useful one. A routing layer built on a real defect is worth reusing. One sold on a 'god/ghost' duality was worth reading once and forgetting. Fork the code and test the routing behavior against your own runs. Skip the withdrawn theory: the retraction says the measured gain reproduces, not that the explanation does.

Covered in the August 17 briefing, two days after the repo's August 15 retraction.