ADIC Red Teaming Report (Recomputable Verification)

The architecture of ADIC is designed not to rely on post-hoc explanations, but on recomputability—the principle that any third party can derive the same judgment from the same inputs. Therefore, it is a mandatory prerequisite for this protocol not only to propose the theory but also to independently design attack scenarios and verify the detection capabilities and the cryptographic binding of evidence.

In this red team execution, we tested the detection capabilities against the following attack vectors to verify the structural boundaries where a judgment switches from OK to NG:

• Fingerprint Tampering (Layer-A: changes in data, configuration, code, and environment fingerprints)

• Semantic Disturbance (Layer-B: destruction of time-series data via localized spikes, missing value injections, phase shifts, etc.)

• Data Contract Violation (schema destruction where parsing succeeds but requirements are not met)

• Boundary Exploration (continuous application of minimal perturbations aimed at the precision margin of the threshold)

• Post-hoc Simultaneous Tampering of certificates, ledgers, and evidence

As a result of the verification, we confirmed that detection (NG verdict) functioned accurately for all attack scenarios as designed. Furthermore, through an independent verification process, the cryptographic hash consistency among the generated certificates, evidence files, and the global ledger was confirmed via recomputation. This constrains post-hoc reinterpretation and verifies local consistency.

This test is re-executable on the public repository and is intended for independent verification by third parties.

• Target Commit: 815e61d763a13962b8823f990924967393433896

• Repository: https://github.com/GhostDriftTheory/ghostdrift-adic-audit

ADIC judges legitimacy not by explainability, but by recomputability.