🏔️ #17 Parameter-plateau vs needle GROUNDED

Does the feature's performance form a broad plateau across its parameter grid (robust — neighbours also work) or a single-node needle (overfit — only one lucky setting works)? A real signal has a contiguous plateau of good IC around the centre; an overfit one has an isolated needle. All 3 gates ground, robust across 3 seeds.

3 / 3

gates · 3 seeds

8 / 9

median plateau run

0.0

needle false-plateau

19 / 20

terminal (18 G · 1 R)

The three gates

Gate (§7 row 17)	Result (3 seeds)
Admit / power (plateau)	detect 1.0 — R_agg≥.7 & neighbour-fraction P≥.8 & contiguous run ≥5 (median run 8 of 9)	PASS
False-plateau (needle + noise)	needle 0.0 (median run 1) · noise ≤0.05 — neither called a plateau	PASS
Neighbour-correlation preserved	adjacent-node corr 0.83 (∈[.6,.95]) — the real knob-neighbour structure	PASS

The plateau/needle separation is a contiguous-run-through-centre. A real robust signal spreads good IC across adjacent parameter settings (a broad plateau); an overfit needle concentrates it on one lucky node. The 1-D cos-θ node grid is built from a deterministic mix of two real gaussianised features (√.85 correlated direction + √.15 per-node idiosyncrasy), so adjacent nodes carry the real .7–.9 knob-neighbour correlation; the null permutes only the target and preserves the cross-node feature-correlation. A needle — good only at the centre node, driven by that node's private idiosyncrasy — never forms a contiguous run ≥5 and so is correctly rejected. Real-data-only (03c).