Same input, different verdicts. A public key taken from the 8-point small-order torsion subgroup of Ed25519, paired with a signature an attacker can compute without knowing any scalar, is accepted by some mainstream verifiers and rejected by others — on the same machine, with the same vectors.

We published the full dataset: verifier-divergence-corpus — pinned vectors + replay harness + CI, so every claim below is re-runnable in minutes.

What diverges

Cross-verifier verdict divergence on the same forged vector: 8/2/2/2/0 across the matrix (accept / reject / reject / reject / crash-style failure). Zero secret needed — the forgery is constructed purely from the torsion subgroup structure.

v1.1: encoding changes verdicts

v1.1 adds a non-canonical encoding class (vectors 9–11): the same group element re-encoded with y' = y + p (mod 2^255):

  • tweetnacl flips reject → accept on 2/3 of the re-encodings
  • noble {zip215:true} accepts 3/3

The same mathematical point changes verdicts depending on its byte encoding. Any enumeration/denylist-style defense is incomplete by construction — only a decode-then-subgroup property test (8P = identity / isSmallOrder) is complete.

Scope and honesty

  • Everything is library-side behavior, no service-endpoint claims.
  • No CVE numbers here; this is a published conformance dataset, not a disclosure event.
  • Vectors are pinned; the harness is idempotent; CI replays on every push.

Replay

Clone, run the harness, watch CI. Third-party re-verification is the whole point — if your verifier disagrees with our matrix, that's a finding about your verifier (or ours), and we want to know.

Cross-verification welcome: file issues with your verifier's verdicts on the pinned vectors.

Correctover builds independent authorization verification for AI agents (CCS — signed, offline-verifiable action evidence; IETF draft-correctover-ccs). This crypto divergence work feeds the same discipline: evidence a third party can re-verify, not claims they must trust.