https://www.oceanheart.ai/tags/regulated/ Recent content in Regulated on oceanheart.ai Hugo en-gb Mon, 15 Jun 2026 00:00:00 +0000 https://www.oceanheart.ai/projects/fail-closed-llm-engine/ Mon, 15 Jun 2026 00:00:00 +0000 https://www.oceanheart.ai/projects/fail-closed-llm-engine/ <a class="claude-coach" href="https://claude.ai/new?q=Read&#43;https%3A%2F%2Fwww.oceanheart.ai%2Fprojects%2Ffail-closed-llm-engine%2F&#43;and&#43;interrogate&#43;the&#43;architecture&#43;with&#43;me%3A&#43;where&#43;fail-closed&#43;validation&#43;could&#43;still&#43;leak%2C&#43;how&#43;the&#43;audit&#43;trace&#43;holds&#43;up&#43;under&#43;adversarial&#43;input%2C&#43;and&#43;what&#43;you&#43;would&#43;probe&#43;first&#43;in&#43;a&#43;design&#43;review." target="_blank" rel="noopener noreferrer" aria-label="Open a prepared Claude conversation"> <span class="claude-coach-mark" aria-hidden="true">&#10035;</span> <span class="claude-coach-copy"> <span class="claude-coach-eyebrow">Claude coach</span> <strong class="claude-coach-title">Interrogate this architecture</strong> <span class="claude-coach-description">Open a Claude conversation primed to stress-test the fail-closed design with you: validators, audit traces, and where you would attack it first.</span> <span class="claude-coach-action">Stress-test it with Claude <span aria-hidden="true">&#8599;</span></span> </span> </a> <h2 id="what-it-is">What it is</h2> <p>LoanSlam is a fail-closed conversation engine for a regulated lending domain. An untrusted language model proposes each turn; deterministic code decides what the customer actually sees. Every turn is audit-traced.</p> <p>The design goal is narrow and unforgiving: let a model help customers in a domain where a wrong answer is a regulatory problem, not a bad review. So the model never gets the last word. The LLM is an untrusted planner. The compliance boundary is plain code.</p>