Cost-validated savings (beta)

Owners and admins can allow TrAIce to retain redacted text prompt and output pairs. When an integration also sends the effective request, TrAIce retains its redacted system prompt, earlier conversation turns, and sampling parameters with the pair so experiments replay what the application actually sent. Each captured sample expires up to 14 days after it is created. Starting a run copies 30 selected samples into encrypted evidence with a separate 14-day expiry. Eligible samples are sent to Google Vertex AI and the selected candidate provider. Once a model swap is live, TrAIce checks it weekly the same way: 30 recent samples from the cheaper model are replayed on the original model and judged, and only the result, without content, is kept after the evidence expires. An owner or admin can also test a shorter prompt: 30 captured requests are replayed on the same model with less conversation history, or with a shorter system prompt that the judge model on Google Vertex AI proposes, and judged the same way. A proposed system prompt is kept encrypted with that evidence and is never applied for you; only the result is kept after the evidence expires. On the Pro and Team plans, TrAIce also groups each workload's samples into topics once a week: the prompts are sent to Google Vertex AI to group them and to name each group, the names are kept encrypted only as long as the samples they came from, and each topic's request counts and costs are kept without content. Tools, retrieval context, attachments, and agent state are not captured and are not supported by these experiments. An owner or admin can separately keep a completed run as a swap test set, which keeps its encrypted cases for up to 90 days. Opt out to delete retained samples, evidence, and swap test sets immediately.