Parametric memory auditing evaluates unprompted brand perception by systematically probing what major foundation models have memorized during pre-training, without access to live web search or URL grounding. The engine initiates multi-round, ungrounded prompts across providers like Google, OpenAI, and Anthropic, employing token-level logprob extraction via Treewalker to measure probability confidence and explore alternative token branches.

These raw outputs are parsed into ranked, line-by-line statements that are processed through an automated curation pass using structured JSON schemas. This stage groups related assertions, calculates appearance rates, and collapses variations into a canonical fact ledger that maps consensus across platforms. The system surfaces key perceptual themes through cross-model narrative analyses, flags emerging hallucinations or brand weaknesses, and records human editorial rulings on individual claims to maintain an authoritative benchmark of organic model recall.