RelayAdversarial
Regressions from the adversarial audit, on its own reproducers
5/5 passed in 0ms
| Test | What it does | Result |
|---|---|---|
test_audit_domain | At tau = (0.9, 0.5), omega = 2, g = 1.72 puts chi at 5.696, between the bound 4.356 and the pure-loss 6.222, and cvmdi_point converges onto cvmdi_rate. | ✅ pass |
test_audit_helps | A HELPING correlation raises cvmdi_rate above pure loss, which is legal: Alice and Bob never get to assume the attack was kind. | ✅ pass |
test_audit_modulation | cvmdi_point converges as 1/va to 2.6e-6 at the guard's va = 1e5 ceiling and raises past it, where the va^2 blocks put it above the bound. | ✅ pass |
test_audit_nan | chi at 3.9, 3.9999999996 and 4.0 on lossless equal arms raises, a NaN comparing false against every bound. | ✅ pass |
test_audit_quadratures | Off the bisector chi does not determine the rate: cvmdi_rate reports -0.155 where the per-attack rate is +0.067, never the other way. | ✅ pass |