> ## Documentation Index > Fetch the complete documentation index at: https://docs.omnia-voice.com/llms.txt > Use this file to discover all available pages before exploring further. # Self-improvement (RL) > Your model improves itself on your own traffic, judged by your own validated evaluators. Self-improvement closes the loop between your logged traffic, your evaluators, and your models. Instead of curating a training set by hand, Omnia generates candidate answers to your real prompts, scores them with a **validated judge** (an aligned criterion in the API), and trains on what passes: automatically, with hard budget ceilings enforced server-side. There are two halves, and they feed each other: For each logged prompt, the platform samples several candidate answers, keeps the ones your judge passes, and fine-tunes on the winners. Prompts that fail every attempt are set aside rather than forced. The prompts self-training gave up on become the **candidate queue**, and online reinforcement learning trains directly against your judge's graded signal on exactly those prompts, where offline sampling can't help. ## Why the judge must be aligned Training a model toward an evaluator's opinion is only safe when the evaluator is measured. Omnia refuses to start any self-improvement work unless the judge is **aligned**: validated against your human labels with true-positive and true-negative rates at or above the platform bar. An unvalidated judge as a training signal is reward hacking by construction; the platform will not do it. ## Safety rails on every run * **Reward budget**: a hard ceiling on judge spend, enforced by the platform's ledger (never self-reported by the trainer). A run that hits its ceiling is refused further scoring and marked over-budget. * **GPU-hour budget**: a wall-clock ceiling on the training side. * **Hacking tripwire**: a run whose policy collapses onto degenerate judge-pleasing answers (entropy collapse, vanishing reward variance, drifting divergence) is stopped automatically and flagged. * **Held-out gate**: a slice of prompts is reserved at run start that the trainer never sees; adoption decisions use held-out judge pass-rate, not training reward. * **Stop button**: stopping a run takes effect at its next scoring call. ## The candidate queue Every self-training round records the prompts your model failed even with extra attempts. The dashboard shows this as **candidates waiting**, the strongest possible signal for when an online run is worth its cost, and a run started with *use candidate queue* trains on exactly those prompts. ## API Runs are scriptable with the same shapes as fine-tuning; see the [GRPO runs API](/fine-tuning/grpo-api).