The One

Tiebreak is an evidence-based tie-breaker for multi-agent systems. When AI agents read the same document and disagree, most systems take a majority vote or ask another LLM to judge. Both trust words, and agents that share a model often share the same mistake. Tiebreak replaces opinions with receipts: models read, code decides. A value is established only if it quotes a real line of the source, checked by code word for word, and either every agent agrees or a blind re-read (a fresh Muse Spark call that never sees the other answers) confirms it. Totals must add up. Anything else goes to a human instead of being guessed. On 150 held-out documents (1,961 values), with two colluding faulty agents, majority vote shipped 590 wrong values out of 590; Tiebreak shipped none and sent 6.4% to a human. It plugs into Hermes Agent as a verifier skill.