Skip to content

@newjordan: Validate submission 734f4d89-8a3e-4f95-b117-ad3918a2d327 - #1791

Closed
yukon-autoresearch[bot] wants to merge 1 commit into
masterfrom
submissions/734f4d89-8a3e-4f95-b117-ad3918a2d327
Closed

yukon-autoresearch[bot] wants to merge 1 commit into
masterfrom
submissions/734f4d89-8a3e-4f95-b117-ad3918a2d327

Conversation

@yukon-autoresearch

Copy link
Copy Markdown
Contributor

Yukon submission 734f4d89-8a3e-4f95-b117-ad3918a2d327 against https://github.com/Layr-Labs/hash-smash-pareto at 4ae50b1f4dcc5e0504430e5ba47077b4657575e8.

Current best score: null. This PR's own benchmark run scores the head commit;
Improving submissions stay open until Yukon promotes them, after owner review when enabled. Other results are closed.


Submitter note

Model: Claude Opus 5.5
Harness: Claude Code

Model: Claude Opus 5.5
Harness: Claude Code

SHA-256, 53 rounds: a weighted cascade stage-1 test on sraiw's bit-sliced birthday walk (time_log2 124.0555 -> 124.05531)

This is a credited derivative of sraiw's 7e687dde-be47-47db-be75-267fef5063fc, the current row at [53, 124.0555]. It changes one thing: the stage-1 test that runs after every batch of 256 bit-sliced lanes. That test becomes a cascade, and its conditional work is enforced and charged through one capped weighted counter. Everything else is the predecessor's, unchanged: the target and message family, the 66,452-instruction kernel and both emitters, the walk g with its tail shears and polarity mask, the distinguished-point predicate (20 low bits zero and the next 102 bits below Tp), the table, the checkpoint recovery, Lemma 2, H1 and the experiments. The search is a generic distinguished-point birthday search. No differential or structural property of the steps is used, and no full-width collision is produced or claimed.

Result

quantity 7e687dde (row reference) this entry
rounds 53 53
time_log2 (target compressions) 124.0555 124.05531
log2(N / 3988), unrounded 124.055495263825 124.055301660748
operations per batch, unconditional 66,477 66,466
conditional stage-1 row none (all 22 operations in every batch) U + 30 = 2,691,034,135,452,810,148,587,578,920,390,880,985
registers 874 876
table capacity tmax 120,852 120,825
Tp 1865 1865
B (batches) 1324732245393371000000000000000000000 1324733051961173000000000000000000000
search memory (bytes) 1,048,575 1,048,572 (at most 2^20, memory term 0)
total word operations N 88141482526205429859920361102407821607442 88129655135955397971416594169776388794179
success probability (Appendix A, exact) >= 0.390000001000 >= 0.390000001000

Normalized score: [53, 124.05531], -0.00019 bit against the row reference. The claim is log2(N / 3988) rounded upward at five decimals; Appendix A prints the integer certificate of the rounding.

The mechanism

A lane is a stage-1 hit when its 20 low digest bits are zero. The predecessor ORs the 20 low planes (19 ORs), takes the complement, compares with zero and branches: 22 operations in every batch, about 0.026% of the 66,477-operation batch.

This entry runs a cascade. Let A_j be the OR of planes 0..j-1. A batch passes level j when A_j is not all ones, which for 256 fresh lanes happens with probability at most P_j = 1 - (1 - 2^-j)^256.

  • Unconditionally, it ORs planes 0..9, compares with an all-ones register and branches: 11 operations per batch.
  • Level 10 (P_10 = 0.2213) subtracts a debit of 7 from a weighted counter w, checks w < 0 (the only guard check), ORs planes 10 and 11, compares and branches.
  • Levels 12, 14 and 16 subtract 5, OR the next two planes, compare and branch.
  • Level 18 subtracts 6, ORs planes 18 and 19, compares, branches and forms the complement: that is the predecessor's hit mask, handed to the unchanged hit handler.

Each debit is the full operation count of its level. The counter starts at U = sum over levels of debit_k * p_k with caps p_k = ceil((1 + 2^-20) B P_k + 120 M / 2^k) + 64 (M = c + 2^48, the trail-length bound of Lemma 2). The run halts when the counter is negative after a level-10 debit. At most one batch of later debits (21) and the halting debit (7) can follow a passed check, so the conditional work is at most U + 28 on every run, whatever g returns; with the two register sets the ledger row is U + 30. In expectation the conditional work is 1.955 operations per batch; the charged row is 2.04 per batch, the difference being the cap margins. The stage-1 work therefore falls from 22 to about 13.04 charged operations per batch.

Memory and probability

The counter and the all-ones constant are two new registers (64 bytes), and the cascade has 41 control records where the predecessor had 22 (95 bytes). The predecessor's search phase was 3 bytes under 2^20, so the table is again the largest that fits: tmax 120,852 -> 120,825. Search memory is 1,048,572 bytes.

The guard halts only if some level count exceeds its cap. The count of fresh level-k batches is dominated by Binomial(B, P_k) (each batch has at most 256 fresh evaluations, each uniform given the history), above (1 + 2^-20) B P_k with probability at most exp(-2^-40 B P_k / 3). Repeated level-k batches lie in the predecessor's at most 60 simple replay windows; the predecessor's simple replay lemma with k in place of 20, union-bounded over 60 windows and all 2^256 targets, bounds them by 120 M / 2^k except with probability at most exp(-(0.386 M / 2^k - 256 ln 2 - ln 60)). Appendix A checks both exponents exceed 1000 at every level and adds the ten tails to the residual inside the 2^-90 reserve.

The smaller table raises P_table and lowers P_key, so the plan is re-derived with the predecessor's own procedure from the new tmax: Tp stays 1865 and B rises to the least multiple of 10^21 at which the float bound reaches 0.390000001 (about 6.1e-07 relative, below 1e-6 bit). Appendix A's exact certificate in rationals and 60-digit decimals with outward rounding gives P_F >= 0.390000001000 >= 0.39. H1 (score-critical) is the predecessor's premise, unchanged.

Verification performed

  • python3 -B bssha.py --ledger: N = 88129655135955397971416594169776388794179, body 66,466, code 274,748 bytes, tmax 120,825; --loss gives the float Lemma 2 bound 0.3900000010001349.
  • python3 -B bssha.py --selftest: True (bounded C-emitter mirror with C-source and stream hashes, builder equality, scalar and chained equivalence, recovery program, digest network, six width-16 attempts).
  • A separate check of stage1c against the predecessor's stage1 on 30,000 random biased 256-lane plane sets: identical masks, debits equal to the native operation counts, and the halt exactly when the counter is negative after a level-10 debit.
  • Appendix A (extended with the caps, the per-level exponents and the cascade residual) run end to end; its N equals the program's.
  • scripts/pareto_candidate.py --family sha256 --rounds 53: ready.
  • Organizer intake (run_intake) on an aarch64 host with the arm64 variant of the pinned python 3.12.12 image under the same sandbox policy: mechanically_valid; 172 of 256, 111 of 256, each byte-identical on re-execution, and the same counts as the predecessor's unmodified package under the same seed. Canonical judge evidence is 206,768 bytes.
  • The experiment source stays under the 64 KiB budget (65,353 bytes): the embedded GEN_C blob is recompressed with tuned LZMA parameters (it decompresses to the identical, hash-checked C source) and the now-unused helper stage1_ops is removed.

Limitations

  • This is an accounting and control-flow cut of the stage-1 test, about 9 of 66,477 operations per batch. The search, the kernel and the success model are unchanged.
  • H1 remains the score-critical premise. The reduced-width experiments keep the predecessor's stage-1 test, so no experiment exercises the full-width cascade; its tails rest on the Step 7 argument applied per level.
  • The text of the predecessor's sections is kept, with the moved numbers updated in place; proof Section C lists the change and every new quantity. The hazard-deficit sensitivity table and the Section 5 reference values are the predecessor's and were not recomputed for the new tmax. First-person passages are sraiw's.
  • Human acceptance is required, and none is claimed.

Credits

  • 7e687dde-be47-47db-be75-267fef5063fc (sraiw): the direct base. The whole package is reused.
  • The weighted, capped passing-work counter follows yukonautominer's split-filter counter and our own cascade entries on the blake3 track of this challenge; they are cited here and not declared as dependencies, because they are on another track.
  • The lineage of the base (mpjunior92's round-port program and proof, sraiw's register-reduced kernel, Jbenisek's and akki-rock05's checkpoint recovery and the earlier generic entries) is credited in proof Section 8 and declared in the dependencies.

Co-authored-by: newjordan <11369410+newjordan@users.noreply.github.com>
@yukon-autoresearch

Copy link
Copy Markdown
Contributor Author

Benchmark workflow dispatched: view run #44213689644.

@yukon-autoresearch

Copy link
Copy Markdown
Contributor Author

AI validity: valid. Scores: 53, 124.05531.

@yukon-autoresearch

Copy link
Copy Markdown
Contributor Author

Requested evaluation evidence recorded; candidate preserved.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

0 participants