Spine NVIDIA Nemotron on Nebius Token Factory

Your AI agreed with you.
Would it still, if you'd said the opposite?

Spine flips the opinion in your question and asks again. If the answer flips with it, you were told what you wanted to hear.

Or replay a real check

Recorded runs: real Nemotron answers, replayed with no new model calls.

Measured, not promised

150 trivia questions with known answers, users nudging toward wrong ones, 3 runs per model. The ground truth comes from the answer key; Spine never sees it.

~9 in 10real cases of deference caught
(recall 0.86 to 0.93)
93–100%of its flags are right when the user leaned on a wrong answer
US$0.002per check, about 8 to 10 seconds

Nemotron 3 Nano and Nemotron 3.5 Lightning as the models under test. TriviaQA (Apache-2.0) with the opinion templates of Sharma et al. 2023. Full method and every run: RESULTS.md.

How Spine decides

  1. Nemotron Nano finds the view you voiced. No view, nothing to test: one cheap call and done.
  2. Nemotron Super rewrites your message twice, changing only your stance: once with it removed, once reversed with the same confidence.
  3. The model under test answers all three, in parallel.
  4. Nemotron Ultra compares the three conclusions. If the conclusion follows whichever stance you voiced, it's Spineless. If you were right and it also disagrees with the flipped version, that's Steady: agreeing with a correct user isn't sycophancy.