Release-bound evidence
These are current Invarra-run results bound to the selected Phalanx release. They are regression and evaluation evidence, not independent certification or a promise that bypass is impossible.
870 / 870
Direct attacks governed or held
0
Observed direct pass-throughs
2,948 / 2,948
Typed attack controls contained
0
Unauthorized effects in declared runs
Current release evidence
Direct jailbreak governance and typed prompt-injection containment are separate contracts. Their numbers should not be blended into one undefined accuracy score.
870 / 870
Attack rows governed or held; 0 observed pass-throughs. Ordinary controls: 56/60 released. Benign lookalikes: 152/190 handled without hard block.
Created and run by Invarra. Excluded from neural-head training under the documented process, but repeatedly used as a runtime release gate; not a pristine untouched holdout.
949 / 949
Attack cases contained; 97/97 benign tasks preserved at the containment boundary and 84/97 passed stricter end-to-end continuation.
External benchmark-derived, Invarra-run. Requires trusted typed provenance and protected sinks. Preservation at the containment boundary is not identical to final-answer correctness.
1,999 / 1,999
Attack-control cases contained; 3,001/3,001 benign-data cases remained available at the authority boundary.
External benchmark-derived, Invarra-run. This measures authority containment, not proof that every injected sentence was semantically identified or every downstream task was completed.
Interpretation
Adaptive evidence
Accepted attempts are scored under the arena protocol. Submitted, accepted, inconclusive, and reviewed outcomes are distinct; the public ledger is evidence infrastructure rather than a customer-runtime guarantee.
