[{"data":1,"prerenderedAt":4},["ShallowReactive",2],{"article-body-the-ai-act-deferred-the-hard-part":3},"\nBrussels gave European compliance teams the sentence they wanted.\nThe AI Act's high-risk obligations are deferred, to December 2027\nfor standalone systems and August 2028 for embedded ones\n([Gibson Dunn's analysis of the postponement](https://www.gibsondunn.com/eu-ai-act-omnibus-agreement-postponed-high-risk-deadlines-and-other-key-changes/)).\nMost boardrooms drew the natural conclusion, that they now have\nmore time.\n\nThey misread what was deferred. The deadline moved. The burden of\nproof did not, and the evidence that proof demands accumulates in\nreal time or not at all. The duties already in force stayed in\nforce, with general-purpose model duties applying since August 2025\nand transparency duties since August 2026.\n\n## What proof will be made of\n\nWhen a high-risk system faces scrutiny, the questions will be\narchival. How was this model validated before deployment, and by\nwhom? What has its error rate been, measured how, over what period?\nWhen it was wrong, who was told, and what changed?\n\nEvery answer is a record with a date on it. Start keeping records\nin mid-2027 and you can account for six months of a system's life.\nAny assessor will read the silence before that for exactly what it\nis, and you cannot buy the missing months at any price. If you own\na system that will one day be classified high-risk, that\nconversation is already on your calendar.\n\n## The backfill temptation\n\nExpect a small industry of retrospective documentation as 2027\napproaches (e.g. validation reports written long after deployment\nand approval chains formalised backwards). Some of it will be\nsincere. All of it proves the ability to write documents, not the\npractice of running a governed system. Regulators are trained to\ntell the two apart, the same skill that separates\n[an evaluation from a pilot](/insights/an-evaluation-is-not-a-pilot).\n\n## What to log from today\n\nFive habits, none of which waits on legal certainty about the final\nstandards:\n\n- **Version everything.** \"Which model decided this\" should be a\n  lookup, not an investigation.\n- **Backtest before deployment** on frozen data and keep the\n  result, including every prediction the model got wrong.\n- **Record approvals** with names and dates.\n- **Score outcomes against predictions** continuously, so\n  performance in production is a curve, not a claim.\n- **Route every material error** to a named owner and keep the\n  outcome.\n\nDo these five things from today and you arrive at your compliance\ndate with an unbroken record that predates the obligation. The same\nevidence also tells you whether your models deserve the trust they\nare getting.\n\nThis is what [Model Passports](/insights/the-model-passport) are.\nA passport holds the version history, validation record, backtests,\napprovals and scored outcomes of every model, accumulated as the\nsystem runs rather than assembled for an audience. Evidence\nproduced as a by-product of operating cannot be bought in 2027,\nwhich is why it is worth starting now. To start the record,\n[start here](/contact).\n",1786984937325]