Treewalker measures recall by running repeated unprompted queries against a model, analyzing token-level logprob tree-walking to discover associative paths above a set probability threshold, and pairing ungrounded and grounded sessions to identify gaps between memory and live search.
sequenceDiagram
participant App as Treewalker Engine
participant Model as Gemini (Parametric)
participant Ground as Google Search
App->>Model: Unprompted Association Query
Model-->>App: Raw Logprobs & Alternative Tokens
App->>Model: Branch Walking (Alts > Threshold)
Model-->>App: Completed Association Trees
App->>Ground: Grounded Search Query
Ground-->>App: Live Web Findings
App->>App: Delta & Holistic ComparisonReferenced by