Context Budget is an Ops Console view of the prompt every panelist evaluation is built from: how much of the model's context window each part reserves, and where to change it.
An evaluation prompt combines the system instructions, the study subject, the sampled persona, prior-result memory, creative assets, and brand attributes. Each is authored in a different place. The chart shows what each part is allowed to cost against the tightest enabled grading model's window, and every row links to the surface that owns it — the prompt catalog, Brand Settings sections, or study setup.
The numbers are worst-case reservations made at admission time, before a persona is sampled or memory retrieved — real prompts are smaller. The budget line names which grading model sets the context window and which sets the output cap, since under overrides they can differ.
An over-reserved configuration (for example, an unrecognized model id budgeting at the conservative default) renders as more than 100% reserved rather than an error — that reading is the diagnostic.
The memory row covers both the panelist's prior-result history and the text that introduces it — how the model is told to read that history, and the header it arrives under. That framing is an editable template, Evaluation memory context, in the Ops Console prompt catalog. Its default is the wording the platform shipped with, so an untouched tenant behaves exactly as before. How much history the framing wraps is a separate control — see below.
The template must keep its memory placeholder — an edit that drops it is rejected on save, so no override can silently stop sending the history it was meant to frame.
The size of the memory reservation comes from the Prior-result memory pair in LLM settings: how many prior studies a returning panelist looks back over, and how much prompt that history may occupy. The two only mean anything together — entries past the byte budget are dropped, so raising the lookback alone is silently undone. Admission reserves the byte budget for every study before a persona is sampled, which is why raising it spends headroom that asset-heavy studies would otherwise use.
The real limit is the tightest enabled grader's context window, not the input field's ceiling. A value the active ensemble cannot admit is rejected when you save, naming the setting rather than failing later on each study.