ZTC
A classification and decision paradigm based on embedding similarity. Decisions are made by computing the cosine similarity between a query embedding and candidate option embeddings, generating no tokens in the process. The approach is characterized by high determinism and low computational cost.
Timeline (last 90 days)
ZTC offers three sizes—4B, 9B and 27B—progressing from CPU-friendly to highest quality.
ZTC (Zero-Token Confidence) takes a question-answer pair generated by any model and outputs whether the answer is trustworthy in a single forward pass, without generating any text.
The ZTC decision engine directly reads a certain aggregation vector after only a single forward pass, outputs a classification judgment, and produces no text tokens at all, belonging to discriminative models.
The author calls this local form ZTC (Zero-Token Confidence).
VIDRAFT Darwin-27B-ZTC-v2 is an open-weight model that uses ZTC (Zero-Token Confidence) to score decisions, generating no tokens during the decision process.
Because no tokens are generated, ZTC's decision outputs are deterministic and inexpensive.
ZTC completes decisions in a single forward pass by applying calibrated probes to the model's final hidden state to produce results.
All Facts (7)
ZTC offers three sizes—4B, 9B and 27B—progressing from CPU-friendly to highest quality.
50%UnverifiedZTC (Zero-Token Confidence) takes a question-answer pair generated by any model and outputs whether the answer is trustworthy in a single forward pass, without generating any text.
50%UnverifiedThe ZTC decision engine directly reads a certain aggregation vector after only a single forward pass, outputs a classification judgment, and produces no text tokens at all, belonging to discriminative models.
50%UnverifiedThe author calls this local form ZTC (Zero-Token Confidence).
50%UnverifiedVIDRAFT Darwin-27B-ZTC-v2 is an open-weight model that uses ZTC (Zero-Token Confidence) to score decisions, generating no tokens during the decision process.
50%UnverifiedBecause no tokens are generated, ZTC's decision outputs are deterministic and inexpensive.
50%UnverifiedZTC completes decisions in a single forward pass by applying calibrated probes to the model's final hidden state to produce results.
50%