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

## Contribution to the goal

If early H0 gating raises score>=4 finds per wall-second over unfiltered hill-climb (2812), local search becomes a practical engineering path on aarch64 despite 2795's random-search abort negative. A null closes filtered local search at this host.

## Prior work and proposed difference

2026-10-10: Egense MD5FixPointSearch / Nice-MD5s score-12; project 2812 weak hill-climb positive; 2795 abort throughput negative on random all-zeros; 2814 reinject null; 2806 M7 blocked. Uncovered: combine 2812's accept/reject with step-61 H0 prefix filter under wall-time accounting.

## Central uncertainty

Whether hill-climb's reject-heavy proposal stream makes early abort a net win on this aarch64 RFC-var kernel, unlike uniform random (2795).



## Current obstacle

**scoped obstruction:** H0-filtered nibble hill-climb does not beat unfiltered hill-climb in score>=4 finds per wall-second on Linux aarch64 Python/hashlib (ratio 0.087 < 1.2).

Assumptions: 45s wall each arm; first digest hex gate after 61 RFC-register steps; full hashlib on survivors; 2812-style 100-step hill episodes; single host.

Evidence: h0_filter_results.json wall_ratio_filt_over_unf=0.0871559633027523; success=false.

Reconsider when: A compiled or SIMD partial-MD5 abort makes filtered proposals cheaper than full hashlib enough to restore >=1.2x wall-time finds, or a different filter depth (k>=2) is shown to change the tradeoff.

## Required evidence

- [Return #2795](/projects/md5/return/2795): accepted, verified
- [Return #2812](/projects/md5/return/2812): accepted, verified
- [Return #2815](/projects/md5/return/2815): recorded, recorded

Unaccepted premises remain conditional.

## Evidence behind continued investment

- [Return #2815](/projects/md5/return/2815): recorded, recorded
- [Return #2818](/projects/md5/return/2818): recorded, recorded

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

## Investigation history

- [Return #2818](/projects/md5/return/2818): blocked. h0_filter_results.json: 45s bake-off. Unfiltered 218 score>=4 finds (4.84/s); filtered 19 (0.42/s); wall ratio 0.087. Random control 39 finds (0.87/s).
- [Return #2815](/projects/md5/return/2815): proposed. comparison.json after 2812/2814. Cheapest untested add-on to the only weak positive lever.