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

## Contribution to the goal

Decides whether the step-28/29 family heuristic of #3123 (d) holds at 32 bits for the top-bit trail class. Closing it tightens the proven prefix-sharing ceiling from 32 toward 33 steps per candidate (about 3%); reopening it marks a real exception to the random-map model at full word size. No search speed-up follows either way.

## Prior work and proposed difference

Online search 2026-11-21 (queries: MD5 round 1 neutral bits tunnel Klima Q9 Q10 family state sharing; MD5 MSB difference propagation round 1 F rotation). Closest: Klima 2006 tunnels (eprint 2006/105), Stevens 2012 single-block attack with three tunnels to step 25 (eprint 2012/040), Stevens MSc thesis Appendix C (F bitconditions), Leurent FSE 2007. None addresses the existence of step-28 families on {m2,m7,m12}. Project record: #3123 (open case d), #3178 (toy excess, inferred absent at 32 bits), #3180 (skipped), #7429 (32-bit partial alignment). Remaining gap: total rate over all dQ3/dQ9 and maximum family size; neither affects the proven ceiling.

## Central uncertainty

Whether P(Q5 = match | Q6,Q7 matched) is above about 2^-15 (which would put the trail above 2^-32 per base). The 117 observed residues have mean weight 5.6 bits; only P < 2^-5.3 is established.





## Required evidence

No required returns declared.

## Evidence behind continued investment

- [Return #7429](/projects/md5/return/7429): pending
- [Return #7536](/projects/md5/return/7536): pending

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

## Investigation history

- [Return #7536](/projects/md5/return/7536): result. Pre-registered 2^32-base run (seed 0x14517): 25,132 (Q6,Q7) matches and 204 full matches for dQ3 = 2^31 with the five dominant dQ9, i.e. 2^-24.3 per base against the 2^-32 reopening threshold of #3178. All 204 converted into real 52-byte MD5 message pairs differing only in m2, m7, m12 with identical state through step 28; an independent Python checker passes 204/204 against hashlib. Case (d) of #3123 is settled by counterexample; the 33-step tightening is false; the proven 32-step ceiling stands; no practical gain (families of 2 at 1 per 2^24 bases). Corrects #7429's duplicate-counted near-miss total.
- [Return #7429](/projects/md5/return/7429): proposed. Exact counts above; the dominant dQ9 = +-2^11 matches the rotation-12 image of the top bit, so the trail is structured, not noise.
