3-Stage Model Pipeline

Logit Readout Extraction vs. Auto-Regressive Generation Path

Stage 1 // Evidence Input Stage 2 // Single Forward Pass Stage 3 // Decision Output TOKENS LOGITS PROBS LONG DOCUMENT Context Evidence (N tokens) QUESTION BLOCK Target Query Prompt DECLARED OPTIONS Candidate Choice Tokens A B C D MODEL CORE 1x Forward Evaluation ● Parallel Self-Attention LOGIT FILTER Extract Slots: [A, B, C, D] RESTRICTED SOFTMAX Normalize Declared Set FULL VOCAB LOGITS Unconstrained (32,000+ slots) AUTOREGRESSIVE SAMPLER Token-by-Token Loop RANKED DECISION Target Token Probabilities A [#1032] 78.2% B [#1033] 14.1% C [#1034] 5.3% D [#1035] 2.4% TEXT DECODER Free-form Generation "Answer is A..."
System Overview
Select or hover over any node or option token to inspect the single-pass logit extraction mechanism.
MODE: LOGIT READOUT
FORWARD PASSES: 1