Subjective Evaluation and Consensus in the Taloton
The Taloton does not leave a scenario result to a single model verdict. The High Council layer combines different reasoning styles so quality, consistency, and explainability signals can be reviewed in one report.
1. Council Structure
The council has three independent seats. All three score with the same rubric; the order of the two answers is shuffled separately for each seat, and which seat made which call is never shown:
- Gemini 2.5 Flash Lite
- GPT-4o Mini
- GPT-OSS 120B (on Groq; falls back to Gemini 2.5 Flash)
2. Decision Mechanism
- Each seat says which of the two answers is better (or that it is a tie) and scores the gap from 0 to 30.
- The winner is decided by majority vote; if the votes are split evenly, the round is a tie.
- The round’s margin is the median of the gap scores from the seats that did not call a tie — a single extreme vote cannot drag the result.
- If a seat fails to answer, the margin is reduced, so the uncertainty carries into the damage.
- If the margin is very small, nobody is treated as the winner: both sides take a small hit.
3. Dissent
If the seats do not pick the same winner, or their gap scores are far apart, the system marks the round as "Council split" and reduces that round’s damage.
"True justice arises not from absolute numbers, but from the consensus of diverse intelligences."