https://dmitryserikoff.com/tags/llm/Recent content in Llm on Dmitry SerikovHugoen-usThu, 23 Jul 2026 00:00:00 +0300https://dmitryserikoff.com/projects/lynxcontract/Thu, 23 Jul 2026 00:00:00 +0300https://dmitryserikoff.com/projects/lynxcontract/<p><strong>LynxContract</strong> gives developers control over LLM-generated code by anchoring every change to a contract, a requirement, and a verification check. Git records <em>what changed</em>; LynxContract records <em>why the code exists at all</em>.</p> <p>Every code unit carries a <code>//@contract</code> block — signature, pre/post conditions, behavior — enough for an agent to generate the implementation without inventing anything. Tests are derived from contracts, not from implementations, so silent bugs have nowhere to hide. When a contract can&rsquo;t cover a requirement, the honesty protocol stops the agent, records the gap, and marks the location instead of improvising.</p>https://dmitryserikoff.com/posts/how-do-you-take-back-control-of-ai-generated-code/Thu, 23 Jul 2026 00:00:00 +0300https://dmitryserikoff.com/posts/how-do-you-take-back-control-of-ai-generated-code/<p>AI-generated code has a strange failure mode: it can be functionally correct and still unaccountable. Nothing connects the requirement to the implementation. Git records <em>what</em> changed; nothing records <em>why the code exists at all</em>.</p> <p>My take: control doesn&rsquo;t come from restricting the agent — it comes from a construction in which the machine simply has no way to change the system silently. Machine-readable contract blocks ahead of the code (Eiffel-style, but used prospectively), tests derived from contracts rather than implementations, explicit gap-marking when the agent lacks context, and a deterministic graph that answers &ldquo;why does this line exist?&rdquo;</p>