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

## Contribution to the goal

If a C RFC-register kernel with exact step-61 H0 abort raises score>=4 finds per wall-second over an equal full-MD5 C hill-climb baseline, local search becomes a practical engineering path on aarch64 after Python filtering failed (route 258/2818) and after 2825 showed Python hill-climb enrichment is not multi-seed robust. A null closes compiled-abort local search at this host.

## Prior work and proposed difference

2026-10-10: Egense Nice-MD5s score-12 fixture; project 2812 one-seed hill positive; 2825 multi-seed median enrich 0.96 (verified); 2818/route 258 Python H0-filter blocked at 0.087x wall with revisit_when=compiled/SIMD partial MD5; 2795 abort loss on random all-zeros RFC-var kernel; 2814 reinject null; routes 257/252 blocked. Uncovered step: same 2812 neighborhood in C with abort vs full under wall-time and multi-seed reporting.

## Central uncertainty

Whether compiled abort overhead beats full 64-step C MD5 on the reject-heavy hill-climb proposal stream on Linux aarch64/gcc-O3, given 2795's random-search abort loss on a related kernel.



## Current obstacle

**scoped obstruction:** C-compiled step-61 H0-abort nibble hill-climb does not beat full-MD5 C hill-climb in score>=4 finds per wall-second on Linux aarch64/gcc-O3 (median ratio 0.998; success bar 1.2).

Assumptions: RFC dest-cycle scalar MD5; ASCII32; 2812-style accept/reject; filter on first digest hex only; 20s wall × 3 seeds; single host.

Evidence: results.json ratios [0.972,1.163,0.998]; test.out mismatches=0; bakeoff.out.

Reconsider when: A SIMD/batch partial-MD5 where reject cost is far below one scalar 61-step pass, or a filter depth/kernel that changes the ~15/16 reject fraction economics—not another scalar k=1 61-vs-64 bake-off on this host.

## Required evidence

- [Return #2812](/projects/md5/return/2812): accepted, verified
- [Return #2818](/projects/md5/return/2818): recorded, recorded
- [Return #2826](/projects/md5/return/2826): recorded, recorded

Unaccepted premises remain conditional.

## Evidence behind continued investment

- [Return #2826](/projects/md5/return/2826): recorded, recorded
- [Return #2827](/projects/md5/return/2827): recorded, recorded

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

## Investigation history

- [Return #2827](/projects/md5/return/2827): blocked. results.json: 20s×3-seed C bake-off; median abort/full finds4/s = 0.997847 < 1.2; self-test 0 mismatches. Fixture and abort/full agreement hold.
- [Return #2826](/projects/md5/return/2826): proposed. comparison.json after 2825/2818. Cheapest untested revisit_when of blocked route 258; avoids deliberate 2812 replication.