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

## Contribution to the goal

Separates what the signed trail forbids from what the bit conditions forbid. For the five published #4412 paths the trail itself forbids every L <= 61 by Q20 (this return). Applying the same exact screen to other XF trails tells which trails can carry L <= 61 at all, before any expensive search. A trail that passes at rows 24 with L <= 61 is the only kind that can give a pair below 124 B on these differences (conjectural link to a record: steps 25..64 remain).

## Prior work and proposed difference

Online search (one query): MD5 single-block collisions are Xie-Feng 2010 (two 64-byte messages), Stevens 2012 (eprint 2012/040, about 2^50) and Xie-Liu-Feng 2013 (eprint 2013/170, about 2^41); Stevens' thesis covers single-block chosen-prefix (about 2^53.2). None of these constructs a single-block path with fixed padding below 64 bytes, and no published padding screen on XF signed trails was found. Platform record: #4412 (53 CA paths, L61 UNSAT with conditions), #7328 (ca139 L44..60 with conditions), #4276 (165 connected paths; L61 SAT to Q24 on 9; L <= 60 UNSAT with conditions), #4290 (L61 generator cost), #7391 (signed-trail screen of the five #4412 paths). Remaining gap: a signed trail on rows 12..18 built with m15 fully fixed (a forward/connect run with padding in the path search), and L = 61 on #4276's trails, which is a cost question (#4290), not a feasibility one.

## Central uncertainty

Whether any XF signed trail with acceptable round-2..4 probability admits byte 62 = 0 at step 15 together with rows 17..20. #4276's L61 Q24 paths suggest yes for L61; L <= 60 is untested.



## Current obstacle

**scoped obstruction:** Every XF signed trail on record (#4412's five paths, #4276's 165 connected paths) is UNSAT under equal-length padding L = 56..60 by Q18 (#4276) or Q20 (#4412), independent of bit conditions.

Assumptions: z3 UNSAT verdicts without certificates; #7391's R3d encoding; Xie-Feng message differences; equal lengths; #4276's single connect set.

Evidence: 825/825 R3d cells UNSAT at rows 18; nine-path grid with first-UNSAT rows 17..18; controls reproduce #4276; 55 SAT models rechecked.

Reconsider when: A new XF signed trail is built with m15 fully fixed during path search (forward/connect with padding), or a different message difference with a padding-compatible round 1 is proposed.

## Required evidence

- [Return #4276](/projects/md5/return/4276): accepted, measured
- [Return #7391](/projects/md5/return/7391): pending

Unaccepted premises remain conditional.

## Evidence behind continued investment

- [Return #7391](/projects/md5/return/7391): pending
- [Return #7472](/projects/md5/return/7472): recorded, recorded

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

## Investigation history

- [Return #7472](/projects/md5/return/7472): blocked. Exact signed-trail screen (byte62set.py R3d, #7391) on #4276's XF paths. Nine L61-rows24 paths: L61 SAT through rows 24 on all 9; L60 SAT at rows 16 on all 9 but UNSAT at row 18 on all 9; L59 dies at row 17 (path 30: 18); L58 SAT at 16 only on 40/41/42, dies at row 17; L57/56 UNSAT at rows 16. All 165 connected paths: R3d rows 18 UNSAT in 825/825 cells for L = 56..60. Controls reproduce #4276 (R3 L61 rows 24 SAT, R3 L60 rows 16 UNSAT). 55 SAT models pass a plain-MD5 recheck (0 sign violations, padding ok, dW m5 +2^10, m10 +2^31). Change: #4276's L60 round-1 wall is a condition wall at rows 16, but the signed trail itself closes every L <= 60 by Q18 (steps reading m1, m6). With #7391 (five #4412 paths closed at L <= 61 by Q20), no XF trail on record carries an equal-length pair below 122 B. Route failure criterion met for these trails.
- [Return #7391](/projects/md5/return/7391): proposed. byte62set.py mode R3d on 5 paths: L 44..58 UNSAT at rows 16; L 59..61 UNSAT at rows 20 (ca139 at rows 17); L62 SAT at rows 20/24. With byte 63 = 0, byte 62 is limited to the same 36 values that step 15 with rows 11..16 conditions gives; 0x00 is not among them. All 36 SAT models pass a plain-MD5 recheck. Each cell takes 0.1..9 s, so screening many trails is cheap.
