the object that would not move

A new research line looked unusually good on paper. Take a family of algebraic objects that can be recognised cheaply because they carry a certificate. Let a hidden group action walk over a congruence level of that family. Read a shared secret off the fact that two parties' walks commute. Exact agreement, no noise, small messages — the properties the noisy designs cannot have.

We wrote the protocol sketch, the hardness analysis, and a list of open questions before proving that the walk produced anything new.

It does not. The action permutes the splittings of a fixed object rather than producing new objects, so every step of the walk returns the same underlying thing. What a secret could hide in, then, is a symmetry of an object the attacker already holds, and the only non-trivial symmetry the geometry offered over the fields we cared about had order two. A group of order two is not a key space.

The refutation was a page of algebra, not a failed implementation, which is why it cost a week rather than a quarter. What survived was demoted deliberately: the certificate stopped pretending to be a group action and became a one-time proof that a starting point has the structure the protocol needs.

We kept the write-up and its open questions rather than retracting them, with a correction at the front: the reasoning that led into the dead end is the reasoning someone else would repeat.

The cost is an artifact whose premise is wrong, read from now on with its correction. What it proves is that this team can kill its most attractive line with an invariant, and knows the difference between a certificate and an action.