Yesterday I argued that we need to move from "Prompt Engineering" to "Logic Engineering" to fix LLM hallucinations in critical infrastructure.
Instead of just writing about it, I built this artifact. It is a single HTML file that contains:
The Architectural Whitepaper (Operational Safety via Recursive Constraint Satisfaction).
The Actual Engine running interactively in the browser (Tab: "Run Interactive Demo").
It demonstrates how to decouple Neural Planning (LLM) from Symbolic Execution (SMT) using a self-extending JSON-DSL. The demo simulates the SMT-Solver logic in client-side JavaScript to prove that validation can be portable and serverless.
Hi HN,
Yesterday I argued that we need to move from "Prompt Engineering" to "Logic Engineering" to fix LLM hallucinations in critical infrastructure.
Instead of just writing about it, I built this artifact. It is a single HTML file that contains:
It demonstrates how to decouple Neural Planning (LLM) from Symbolic Execution (SMT) using a self-extending JSON-DSL. The demo simulates the SMT-Solver logic in client-side JavaScript to prove that validation can be portable and serverless.The code is the paper. The paper is the code.
Happy to answer questions about the architecture.