nx_nofloat_induction_gate.nx

nx_nofloat_induction_gate.nx -- CONTENT-BASED INDUCTION (CAP-NF-INDUCTION), REBUILT after the 3rd adversarial review showed the old fixed-repeat-position version could be solved by a fixed-offset head (no content match). FIX: the repeated "key" token now appears at a VARYING earlier position p; the target is its FOLLOWER S[p+1]. Context tokens 0..T-2 are DISTINCT (shuffled) so the key matches a unique earlier position. Because p varies, a fixed-offset(-k) head and a fixed-absolute-position head BOTH score ~chance; the ONLY way to win is true content-induction: at the last position, match the current token to its earlier occurrence BY CONTENT, then copy what followed it. Every example fresh-random -> no memorization. 2-block attention (the induction circuit). T1 induction acc (predict S[p+1]) >> chance(1/V) = content-induction works on unseen data. T2 (teeth) acc > copy-back-1 rate (predict S[T-2]) = it is content-match, not a fixed offset. HONEST: research-hard for a tiny fixed-point model; if it only ties chance/copy-back-1 it has NOT earned the word and stays downgraded. The measured numbers are the evidence. Sovereign. expect_exit: 0