For work you may have to defend later. The answer that gets you in trouble will sound confident.

Single-model AI fails quietly. It skips the caveat, misses the contradiction, invents the source, or agrees with your framing. debate.tellodb runs high-stakes prompts through rival models before you ship them, then returns a verdict with the dissent and transcript attached.

When you prompt a single LLM, you get one reasoning trace, one stream of consciousness, generated by one training distribution. If that trace goes slightly wrong at token 500, tokens 501 through 50,000 will confidently reinforce the error. The model doesn't know it's wrong because its own context window now contains a wrong premise, and LLMs are designed to be coherent, not correct.

This isn't a bug. It's a consequence of autoregressive generation. Chain-of-Thought helps but doesn't solve it, it's still one model talking to itself. Peer review, the gold standard of human knowledge verification, introduces independent agents who have no stake in the first answer and are explicitly instructed to find faults.

Every debate produces a chronological, timestamped transcript. You can see exactly what each model said, what it disagreed with, and why the final synthesis took the shape it did. This isn't a black box, it's a paper trail that holds up to post-mortems and regulatory scrutiny.