The original title is "Harness Engineering" which is very short and vague. I need to rewrite it based on the summary provided.

Short summary

Production AI success depends 90% on the operational harness (observability, evals, rollback, silent-failure detection) and 10% on model choice. Most pilots fail at month nine not because the model degraded, but because the harness wasn't built to handle shifting input distributions and catch silent failures. Successful teams make four specific harness decisions before selecting any model: named ownership of continuous evals, audit-log scope and retention, rollback authority and mechanism, and boundary instrumentation for silent-failure detection.

• •90% of production AI success is operational infrastructure, not model capability
• •Pilots silently fail at month 9 when observability, evals, rollback, or silent-failure detection weren't designed upfront
• •Make four harness decisions before model selection: eval ownership, audit scope, rollback authority, silent-failure instrumentation

Generated with AI, which can make mistakes.