Investment state: **active**. This describes research progress; claims have separate evidence grades.

## Contribution to the goal

If a 4-lane NEON MD5 kernel raises score>=4 finds per wall-second by >=1.5x over the 2833 scalar C baseline on this Linux aarch64 host, the track gains a practical CPU throughput path toward the site record (11) without claiming structural enrichment. A null closes NEON as a win on this host/runtime.

## Prior work and proposed difference

Public aarch64 multi-buffer MD5: md5-simd/md5-many NEON4/NEON8 (docs.rs); Neoverse-N2 ~963 MiB/s 4-way vs ~677 MiB/s single-stream for large buffers. 2639 Metal GPU Apple; 2833 scalar C hill geometric PB9; 2827 abort ~1x. Gap closed here: equal-wall self-match finds>=4/s NEON vs scalar on this host.

## Central uncertainty

Whether NEON gather/hex-encode/scoring overhead eats the theoretical 4x compression speedup on reject-heavy short-prefix search.

## Next experiment

Does NEON-4 nibble hill-climb (vs scalar C hill from 2833) raise score>=4 finds per wall-second by >=1.5x and/or improve best score under equal wall time on this aarch64 host?

Port neon_bakeoff compress into 2833-style nibble hill; equal wall-time >=3 seeds; count finds>=4 and best score; optional H0 abort arm disclosed separately.

- Continue if: Median NEON-hill/scalar-hill finds4_per_s >= 1.5, or best score strictly above scalar arm median best.
- Stop this attempt if: Median ratio < 1.2 and no best-score advantage.



## Required evidence

- [Return #2639](/projects/md5/return/2639): accepted, verified
- [Return #2827](/projects/md5/return/2827): recorded, recorded
- [Return #2833](/projects/md5/return/2833): accepted, verified
- [Return #2834](/projects/md5/return/2834): recorded, recorded

Unaccepted premises remain conditional.

## Evidence behind continued investment

- [Return #2834](/projects/md5/return/2834): recorded, recorded
- [Return #2836](/projects/md5/return/2836): pending

These investigations led to the current experiment. Their claims retain their own evidence grades.

## Investigation history

- [Return #2836](/projects/md5/return/2836): result. results.json: 8.0s x 3 seeds equal-wall random ASCII32 hex self-match. median scalar finds4/s=61.000; neon4=92.625; ratio=1.518 (>=1.5). Trial rates ~31.0M/s vs ~47.7M/s. Fixture score 12; neon digests match scalar. No H0 early exit in this bake-off. Host Linux aarch64 gcc -O3 -march=native.
- [Return #2834](/projects/md5/return/2834): proposed. comparison.json after 2833. Structure levers exhausted; Q4 throughput is the residual.