The modern engineering workflow is currently caught in a dangerous tension between the velocity of AI-assisted coding and the absolute rigidity of physical reality. For a software developer, a hallucinated function name is a bug that a compiler catches in seconds. For a mechanical design engineer, a hallucinated part number or a slight arithmetic error in a load-bearing calculation is not a bug, but a potential procurement disaster or a structural failure. The industry has spent the last year attempting to solve this through prompt engineering, yet the fundamental nature of large language models means that confidence does not equal correctness.
The Architecture of the Verification-Harness
To bridge this gap between probabilistic AI and deterministic engineering, a mechanical design engineer has released verification-harness, a specialized framework designed specifically to audit the outputs of Claude Code. The system is built around 13 distinct verification gates that act as a filter between the AI's suggestion and the final implementation. This is not a wrapper that asks the AI to double-check its own work, but a systemic harness that enforces external validation.
One of the primary failure points in industrial AI usage is the fabrication of identifiers. The verification-harness explicitly prohibits the estimated generation of part numbers and Digital Object Identifiers (DOI), treating any unverified identifier as a critical failure. When it comes to mathematics, the framework bypasses the LLM's internal reasoning entirely. Instead, it triggers arithmetic operations to be re-executed via Bash, using the Unix shell and command-line interface to generate a ground-truth result that is then contrasted against the AI's output.
For data that cannot be verified by a simple script, the system implements a three-level confidence rating for all key numerical values and citations. This creates a transparency layer that tells the human operator exactly how much trust to place in a specific figure. In scenarios involving safety-critical conclusions, the harness employs a parallel induction strategy. It sends the same query to a diverse set of models, including Gemini, Groq, and the OpenAI CLI, to determine if the conclusions converge. If the models disagree, the output is flagged for manual review.
Security is handled through a strict supply chain intake pipeline for external plugins. Rather than allowing unrestricted access, every plugin must undergo a static scan and a human review process before being integrated into the environment. The tool is designed for modularity, allowing users to select specific hooks and checkers rather than deploying the entire suite. Configuration is managed through a `settings.example.json` file, which allows engineers to wire the harness into their specific local environment. For those looking to validate the system's own reliability, the provided Python code allows for immediate regression testing via commands such as `eval/tests/test_g9_regression.py`.
From Prompt Engineering to Systemic Verification
What makes this approach fundamentally different from standard AI safety layers is the shift from linguistic trust to computational verification. Most teams attempt to reduce hallucinations by adding phrases like be accurate or think step-by-step to their system prompts. This is an attempt to fix a probabilistic problem with more probability. The verification-harness treats the LLM as an untrusted proposal engine and the harness as the actual authority.
By routing math through Bash and safety conclusions through a multi-model consensus, the framework introduces a system of checks and balances that mirrors traditional aerospace or automotive quality assurance. The tension here is between the fluidity of Claude Code and the rigidity of the verification gates. When the AI suggests a part number, it is no longer a suggestion to be blindly accepted, but a hypothesis to be tested against a database. When it performs a calculation, it is not providing an answer, but a draft that must be signed off by a deterministic shell script.
This creates a new paradigm where the AI is not the expert, but the draftsperson. The real intelligence shifts from the model's ability to guess the right answer to the engineer's ability to define the correct verification gate. The result is a pipeline where the cost of a hallucination is shifted from the production phase—where it could cost thousands of dollars in wrong parts—to the verification phase, where it costs a few milliseconds of compute time.
This transition toward verification-driven AI marks the end of the honeymoon phase with LLMs in high-stakes industries and the beginning of a rigorous, audit-first era of deployment.
