Integrity
Claims, quotes, and weights
The model may propose attribution, but code verifies quotes and computes contribution weights.
Two separate steps
Synthesis is creative and writes the answer from paid source material. Attribution is deterministic and asks for atomic claims, source ids, and verbatim supporting quotes.
Those two steps are split because money should not ride on unconstrained prose.
Substring gate
Every supporting quote must be an exact substring of the stored paid source text. If the model changes a word, invents a quote, or cites unpaid material, that edge is rejected and contributes no weight.
Weighting
Weights are computed in code from the verified claim-to-source graph. The LLM never outputs payout weights or payout amounts.