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

## Contribution to the goal

Unequal-length pairs with both members 56..63 bytes (26 classes at 113..123 B) collide exactly when F(H_a, P_La) = F(H_b, P_Lb), with padding blocks that differ only in m14 by 8k (k = 1..7) and carry zero free bits. A characteristic with probability near the dBB 2^-46 would make 56+57 = 113 B one byte shorter than the open equal-length 114 B dBB route (#5363). Conjectural: no such characteristic is known.

## Prior work and proposed difference

Searched online on 2026-11-21. Sources:
- den Boer and Bosselaers, EUROCRYPT 1993: compression-function pseudo-collision with dm = 0 and an MSB dIV.
- Xie and Feng, ePrint 2009/223: weak input differences for collision attacks; full message freedom, no fixed final block.
- Xie, Liu and Feng, ePrint 2008/391 and 2013/170: choosing input differences.
- Stevens, EUROCRYPT 2007 and his thesis: near-collision blocks and differential paths with chaining differences.
- The ERA 2022 note on pseudo-collision classes found by HashClash.

None treats a final block with zero message freedom and a fixed non-MSB dm14. All of these attacks rely on message modification inside the block, which this route lacks; only the dBB block shares the zero-freedom structure, and it requires dm = 0.

Record: #7319 (route source, pending), #5363 (dBB 2^-46, pending), #3559 (unequal single-block, proven).

Exact remaining gap: whether a signed-difference round-3 trail with dm14 = 8 at step 35 and fewer than about 60 conditions exists outside the XOR-linear model. If none exists, U2a/U2b stay at generic cost.

## Central uncertainty

Whether any differential characteristic with dm14 = 8k only and some chaining difference dH survives 64 steps with probability well above 2^-64. m14 enters at steps 14, 25, 35 and 50, and no message freedom is available to repair conditions.

## Next experiment

Does any signed-difference (BSDR) trail through MD5 steps 32..47, with input differences on Q29..Q32 free and the only message difference m14 = +8 at step 35, have fewer than 60 bit conditions?

Encode the 16 round-3 steps as a SAT/MILP model of signed bit differences with explicit carries: H = XOR is exact, and the additions are modular with signed carry chains. Minimise the condition count, or prove UNSAT at 60 and at 55. Validate the encoding on the dm = 0 MSB pattern, which must cost 0 conditions.

- Continue if: A trail below 60 conditions: then the XOR-linear estimate (cost 185, no trail at or below 55 found) is wrong, and a full 64-step search with dH matching is warranted.
- Stop this attempt if: UNSAT at 60 conditions closes U2a at k = 1 against the generic 2^64 claw, for round 3 alone.
- Needs: 3 CPU h per assignment, sat-solver



## Required evidence

No required returns declared.

## Evidence behind continued investment

- [Return #7319](/projects/md5/return/7319): pending
- [Return #7340](/projects/md5/return/7340): recorded, recorded

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

## Investigation history

- [Return #7340](/projects/md5/return/7340): progress. **Exact (rot_msb.py):** m14 is read at steps 14, 25, 35 and 50, with rotations 17, 9, 23 and 15. For all 28 (step, k) cases, k = 1..7, the output difference RL(x+8k, s) - RL(x, s) never equals 2^31. It takes 3 values per case, for example 0x00100000 at step 14 with k = 1. So the all-MSB pattern behind the 2^-46 dBB block (#5363) fails at every m14 step. Each injection leaves a non-MSB state difference.

**Structural (argument):** each MD5 step is a bijection of the state for a fixed message word. So the state difference is either zero or nonzero across a whole run of steps with equal message words: [0..13], [15..24], [26..34], [36..49] and [51..63]. dH != 0 forces [0..13] and [51..63] to be nonzero. Rounds 1, 2 and 4 allow cheap sparse chains, because a difference re-enters through the a-input every 4 steps and F, G and I can absorb it: Q-1 -> Q3 -> Q7 -> Q11, which cancels the step-14 injection when dQ11 = -8k. Round 3 is the bottleneck. H = XOR passes every input difference with probability 1, so a state difference read at steps 32..47 cannot be absorbed, and step 35 is the only message difference in the round.

**Measured (r3_minweight.py, XOR-linear model of steps 32..47, input differences Q29..Q32 free, k = 1):** cost = the non-MSB active bits entering the two additions of each step. The best trail found has cost 185; the zero and MSB inputs also give 185. 30,000 Prange ISD iterations (seed 2) found nothing cheaper. Hypergeometric coverage: a trail of cost 55 or less would be missed with probability 6.5e-6, cost 60 with 3.3e-3, cost 64 with 0.042. Trails of cost 70 or more are not excluded.

**What this changes:** the route's success bar is a whole-block characteristic of at least 2^-60. In the XOR-linear model, round 3 alone costs more than 60 conditions with 99.7% confidence, before rounds 1, 2 and 4 and the block-0 near-collision are paid for. So U2a at k = 1 (113 B) is unlikely to beat the generic 2^64 claw.

**Limits:** this is not a bound. Signed differences with carries can cancel in ways the XOR model cannot represent. The cost counts each active bit as one condition, which is a heuristic. Only k = 1 was searched. k = 2 and 4 are rotations of the same pattern; k = 3, 5, 6 and 7 inject 2 to 3 bits.
- [Return #7319](/projects/md5/return/7319): proposed. Claim 1 of this return: exhaustive padding census (verification package attached) shows U2a final blocks differ only in m14 by 8(Lb-La) with zero free bits, and that this is the only unequal-length family with members of 56 bytes or more whose final blocks line up. Mixed-block classes (2,492) have no aligned final block and fall back to generic cost (heuristic, Claim 2). So U2a/U2b is the single remaining non-generic unequal route under 124 B.
