Kosson et al. is adjacent, not the same failure, so the novelty sentence comes down

[syncytium2]

Closes the last item the blind round (#674) left outside its own claim: role 2 flagged
arXiv:2410.23922 (Kosson et al., NeurIPS 2024) as uncited prior art sitting under the
page's closing sentence, "That it prevents this collapse is this page's result."

Verdict: adjacent, not the same. The novelty sentence is withdrawn. Both candidate
papers were read, not searched.

Kosson et al. 2024 — adjacent. It walks the same chain (too large an early step →
damage training does not undo → warm-up prevents it), and that is enough to retire an
unqualified novelty claim. It differs in two ways that matter:

	Kosson et al. 2024	this page
where the dead-unit result lives	§7 + Appendix A, on a small image network	the whole page
direction	units are alive, the update kills them	units are already silent at the starting weights
when measured	fraction of dead units at the end of training	before the first step, in every replayed fit
activation	ReLU — remedy tested is a leaky ReLU	GELU — no exact zero region, so that remedy is unavailable and no layer is ever "dead" in his sense, only silent

Lu et al. 2020 (arXiv:1903.06733) is the closer neighbour, and was not on the list.
Born dead means dead before training; their theorem sends a gradient method (Adam
among those they name) to a constant function; and their depth-up / width-down ordering
is the order Table 1 finds. Recomputed here, it holds in all four comparisons at lr 0.03
— 69→90% and 46→81% with depth, 69→46% and 90→81% with width. But their theory is ReLU-only,
and it forbids what this page measures: a born-dead network cannot be recovered, while
these fits do train at a lower rate or behind a 200-step ramp.

What the page says now is narrower than what it said: that a head can start almost silent
in a network with no zero region to be stuck in, that whether training wakes it is what decides
the fit, and that the failure is recoverable — shown by replaying one fit bit for bit. It also
names what was searched and what was not, because this repo has already paid once for a novelty
claim that four web searches failed to check (detector_history.md).

Scope held. Related work and two references only. Nothing else on that page — the ranked
list from #674 is Tony's to rule on and the census re-run needs his GPU decision. No JSON
changed; only prose. I widened by one paper (Lu), because the novelty sentence could not be
rewritten honestly while a closer neighbour sat unread in my own review record.

Checks. tests/test_diagnose_chorus_collapse.py 3 passed; sapper clear; the page rebuilds
byte-identically on Python 3.11, 3.13 and 3.14; the darkroom copy moved with it
(0dfdc1e5… → 9083d574…, matching the repo copy) and the claim is released here.

🤖 Generated with Claude Code