Skip to content

README.md

Latest commit

 

History

History
131 lines (90 loc) · 3.16 KB

README.md

File metadata and controls

131 lines (90 loc) · 3.16 KB

aibenchlab-verify

Standalone command-line tool for verifying the integrity of AiBenchLab MBX v2 files.

This tool recomputes the SHA-256 content hash of an MBX file and compares it to the stored value. If the hashes match, the file has not been modified since export. If they differ, the data has been altered.

No AiBenchLab installation required. This is an independent verifier that anyone can build and run.

How it works

This tool implements Section 6 (Verification Procedure) of the MBX v2 Specification:

  1. Parse the .mbx file as JSON
  2. Verify the format identifier (AiBenchLab-MBX) and version (2.0)
  3. Save the stored content_hash, then set it to ""
  4. Normalize all JSON numbers to 6-decimal-place strings (e.g., 42 becomes "42.000000")
  5. Serialize to compact canonical JSON (no pretty-printing, preserving key order)
  6. SHA-256 hash the UTF-8 bytes
  7. Compare the computed hash to the stored hash

The normalization and hashing logic is identical to the implementation in AiBenchLab's mbx.rs.

Build

Requires Rust 1.70 or later.

cd aibenchlab-verify
cargo build --release

The binary is at target/release/aibenchlab-verify (or aibenchlab-verify.exe on Windows).

Cross-compilation

# Linux (from any platform with the target installed)
cargo build --release --target x86_64-unknown-linux-gnu

# macOS
cargo build --release --target x86_64-apple-darwin
cargo build --release --target aarch64-apple-darwin

# Windows
cargo build --release --target x86_64-pc-windows-msvc

Usage

Verify a single file

aibenchlab-verify benchmark.mbx

Output on success:

VERIFIED: benchmark.mbx
   Hash: a1b2c3d4e5f6...
   Session: "My Benchmark Run"
   Models: llama3.1:8b, mistral:7b
   Tests: 142
   Exported: 2026-03-02T14:30:00Z
   App Version: 0.7.0-beta

Output on failure:

FAILED: benchmark.mbx
   Stored hash:   a1b2c3d4e5f6...
   Computed hash:  x9y8z7w6v5u4...
   Data has been modified since export.

Verbose mode

aibenchlab-verify --verbose benchmark.mbx

Shows everything from default mode plus hardware details, benchmark configuration, score distribution, anomaly count, and session duration.

Batch verification

aibenchlab-verify --batch ./exports/

Verifies all .mbx and legacy .mbx.json files in the directory and prints a summary:

Results: 12 verified, 1 failed, 0 errors

Help

aibenchlab-verify --help
aibenchlab-verify --version

Exit codes

Code Meaning
0 All files verified successfully
1 One or more files failed verification, or an error occurred

Dependencies

Crate Purpose
clap Command-line argument parsing
serde_json JSON parsing and serialization
sha2 SHA-256 hashing

No network access. No telemetry. No AiBenchLab installation check.

Running tests

cargo test

Tests cover: valid file verification, tamper detection (score, model name, app version), malformed JSON handling, v1 format rejection, batch counting, float normalization correctness, and hash format validation.

License

MIT