This example uses the public MASSIVEMAGNETICS/truth-compiler-ai repository. The useful part is not that the repository can point to successful evidence. The useful part is that the audit preserves a hard boundary where evidence does not justify a stronger claim.
Claim → evidence → state
Real implementation exists
PRESENT
src/ and include/ contain a C++17 implementation; CMakeLists.txt defines the build.
Substantive tests exist
PRESENT
The public assessment records XOR regression, SHA-256 known-vector, and capability-denial tests.
Learning behavior is evidenced
TESTED
XOR convergence is the bounded behavior asserted by the test. The assessment does not inflate that into a claim of general intelligence.
CI and quality controls exist
PRESENT
The assessment records GCC/Clang CI, sanitizers, Valgrind, clang-tidy, and Docker support.
Victor integration grants deployment authority
NOT GRANTED
victor_adapter.hpp is described as bounded and emits a hash-chained verification receipt. Production authorization remains explicitly false.
The model can turn its own prediction into proof
PROHIBITED
The governing law states PREDICTION != PROVEN. The heuristic ranker may prioritize review but cannot manufacture evidence or issue PASS/FAIL/authorization decisions.
Why this is materially different from “AI code review”
A normal model can produce a plausible narrative. Truth Compiler is designed around a stricter object: every material conclusion must retain the evidence basis, the scope, the contradiction state, and the unknowns. The most important result in this self-audit is not a PASS—it is NOT GRANTED. The system is not allowed to promote evidence of bounded learning into deployment authority.
This evidence pattern translates directly to a customer's repository: instead of paying for “looks good,” the buyer gets an inspectable map of what is present, what is tested, what is unsupported, what conflicts, and what authority is not justified. That changes technical due diligence from a vibe into a bounded evidence record.