Mubin Attar — Writing https://mubin-attar.vercel.app/writing Essays, guides, and notes on AI systems, evaluation, and honest ML. en A validator is not a better prompt https://mubin-attar.vercel.app/writing/a-validator-is-not-a-better-prompt https://mubin-attar.vercel.app/writing/a-validator-is-not-a-better-prompt Sat, 20 Jun 2026 00:00:00 GMT essay Why DBWhisper gates every generated query with deterministic code instead of a stricter system prompt — 'less often' is not a safety property. Trust is not a safety model https://mubin-attar.vercel.app/writing/trust-is-not-a-safety-model https://mubin-attar.vercel.app/writing/trust-is-not-a-safety-model Wed, 20 May 2026 00:00:00 GMT essay Why I gate an LLM's SQL behind a deterministic, fail-closed validator instead of trusting the model to behave. Look-ahead-free by construction https://mubin-attar.vercel.app/writing/look-ahead-free-by-construction https://mubin-attar.vercel.app/writing/look-ahead-free-by-construction Sun, 12 Apr 2026 00:00:00 GMT essay The most common way a backtest lies to you is look-ahead bias. Here's how I designed it out of TradePulse instead of hoping to catch it later. Retrieval beats stuffing the schema https://mubin-attar.vercel.app/writing/retrieval-beats-stuffing https://mubin-attar.vercel.app/writing/retrieval-beats-stuffing Sun, 08 Mar 2026 00:00:00 GMT guide A practical approach to natural-language-to-SQL on real databases: document the schema, embed it, and retrieve per question instead of pasting the whole thing into the prompt.