LLMTraceFX Introduces Independent Auditor to Verify That Cache Hits Do Not Prove Computation Omission

LLMTraceFX has introduced an independent verification auditor to address the issue that cache hit does not prove actual computation omission. Since self-reporting by existing engines is unreliable, it verifies engine attestation across five dimensions: independent oracle, observed prompt work, output identity, evaluator correctness, and checksum-bound evidence.

Verification Results and Classification

Deterministic synthetic token-level control tests on 10 cases showed that output identity and evaluator passed in all cases, and attestation matched.

• exact-reuse: 8 out of 9 tokens reused, 1 recorded as verified_hit with prompt work
• partial-reuse: Clearly classified similarity-based reuse boundaries
• verified-miss: 0 reused, 9 recorded as prompt work due to candidate absence, ID mismatch, namespace isolation, etc.
• eviction: Determined as evicted in capacity-pressure and revisit scenarios

These results demonstrate that matching token counts does not imply matching token identities, and show that reuse boundaries are based on prefix identity rather than similarity.

Verification Limitations and Security Enhancements

Since this verification was conducted in a deterministic synthetic control environment, it does not prove MLX/vLLM speedup, production cache correctness, GPU performance, or latency reduction. Additionally, block-level cache behavior is not covered as the standard is token granularity.

Regarding security, the TOCTOU (Time-of-Check to Time-of-Use) risk of the standalone auditor was identified and fixed by generating a private verified script using O_EXCL semantics before execution. The final PR #74 completed verification with 4,499 tests passing and a 6-job matrix build.