23/23
Audi Bench tests passed
PASS
7/7
BMW Bench tests passed
PASS
13/13
BMW Bench probe cases passed
PASS
The Assayer's verdict
9/11 claim groups PROVEN on the Assayer's own re-run — the grader re-ran everything itself, including its own hand-built clients and attacks.
The other two groups — C1 (the two session-recording implementations matching the written format) and C2 (documentation honesty) — were graded FLAWED in the full report, fixed the same day, and re-verified: the BMW session writer was rewritten to the spec's letter and re-ran at probe 13/13 and tests 7/7 under the stricter verifier, and the wording items were corrected. A later spot re-grade also amended A3 from FLAWED to PROVEN after its fix.
Full record — read it before quoting any number from this page:
The other two groups — C1 (the two session-recording implementations matching the written format) and C2 (documentation honesty) — were graded FLAWED in the full report, fixed the same day, and re-verified: the BMW session writer was rewritten to the spec's letter and re-ran at probe 13/13 and tests 7/7 under the stricter verifier, and the wording items were corrected. A later spot re-grade also amended A3 from FLAWED to PROVEN after its fix.
Full record — read it before quoting any number from this page:
- reports/BENCH-ADDENDUM-2026-10-22.md — what was built, the verdict table, the fixes, and the honest limits.
- reports/BUILD-REPORT-2026-10-22.md — the whole car-tooling build (Bench, BMW ID Kit, SQ5 Dash, and the Phase-2 machinery), stage by stage.
What the Bench proves — and what it never proves
It proves tool behaviour: the tools parse correctly, refuse write operations, fail loudly on bad or tampered data, and seal their session recordings so tampering is detected.
It never proves car behaviour: no Bench result says how the 540i or the SQ5 will answer. Every Bench value is synthetic, every ECU in the Bench is a simulation, and the real cars are only reached through the owner gates (G-1 / G-2) — separately, read-only first, with the owner present.
It never proves car behaviour: no Bench result says how the 540i or the SQ5 will answer. Every Bench value is synthetic, every ECU in the Bench is a simulation, and the real cars are only reached through the owner gates (G-1 / G-2) — separately, read-only first, with the owner present.
What was tested, in plain words
Audi Bench — a virtual Audi ECU that only answers read requests (service 0x22), refuses writes, and a probe that reads a 12-record identification set. Fault modes (wrong answers, short answers, garbage) must fail loudly — and did.
BMW Bench — the BMW diagnostic library (EdiabasLib) in its own simulation mode, with a fixed read-only job list. Write jobs were proven unreachable at the runner layer, and the adversarial cases (bad checksums, wrong addresses, truncated answers, and more) behaved exactly as documented.
BMW Bench — the BMW diagnostic library (EdiabasLib) in its own simulation mode, with a fixed read-only job list. Write jobs were proven unreachable at the runner layer, and the adversarial cases (bad checksums, wrong addresses, truncated answers, and more) behaved exactly as documented.