Log probabilities are converted from log space into percentages, aligned to the parsed response items by cumulative character offsets, then traversed to spawn parallel completion prompts for any alternative token whose probability exceeds the configured threshold.
Withheld — proprietary.
flowchart LR
A[Raw Model Output] --> B[Extract Logprobs]
B --> C[Token & Probability Stream]
C --> D[Align Tokens to Items]
D --> E[Item Confidence & Token Alignment]
E --> F[Find Branches]
F -->|prob >= threshold| G[Spawn Branch Prompts]
F -->|prob < threshold| H[Prune Path]