Skip to content
Closed
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
47 changes: 37 additions & 10 deletions pareto/candidates/sha256/claim.json
Original file line number Diff line number Diff line change
@@ -1,20 +1,47 @@
{
"schema_version": 3,
"submission_state": "draft",
"target_profile": "sha256-r37-prefix-v1",
"target_profile": "sha256-r53-prefix-v1",
"attack_class": "ordinary-collision",
"rounds": 37,
"rounds": 53,
"claim": {
"time_log2": 128.0,
"time_log2": 124.05531,
"time_unit": "target-compressions",
"memory_log2_bytes": 0,
"preprocessing_log2": 0,
"memory_log2_bytes": 20,
"preprocessing_log2": 47.22343,
"success_probability": 0.39,
"nonuniform_advice_log2_bytes": 0
},
"restrictions": [],
"baseline_improved": "sha256-r37-nominal-v2",
"restrictions": [
"Two distinct 64-byte messages M(x), M(y). W0..W7 are the fixed prefix in proof Section 1. W8=y0, W9=y1, and W10..W15 are y2..y7 minus K10..K15 modulo 2^32. This is a bijection from the 256-bit chain value x. FIPS padding adds one fixed second block. The target uses the standard IV, steps 0..52 on each block, and feed-forward. The walk g is an explicitly invertible encoding of the block-1 working state (walk B with tail shear masks QM=0x827d8f1f and XM=0x00f4e657) XOR the fixed polarity mask be24390accbe9b61e95138fe5511b644b3c993ab0dea46d6c217239f3f81439e. A collision of g gives equal complete digests.",
"Classical 256-bit word RAM under collision-frontier-v5, C=3988. The declared machine has 876 256-bit registers, 871 data, three pointers, and the stage-1 counter and all-ones register. Every XOR, AND, OR, NOT, move, set, load, store, addition, comparison, branch, and random word is charged. Its native body instruction is exactly 32 bits: a 2-bit opcode and three 10-bit register fields, fetched and executed directly at one operation per instruction under the same pricing convention as the accepted native 5-byte format. No decoder or interpreter runs in the normative path; Prog.run, Prog.compile, and Prog.fast are test-only mirrors. Recovery and control use native fixed 5-byte instructions. The kernel has 66452 instructions (52206 XOR, 8670 AND, 5576 OR); the stage-1 cascade adds 11 unconditional operations per batch and its conditional levels are charged once as U + 30 through a capped weighted counter (proof Section C); loop control adds 3.",
"The single attempt uses one fresh uniform r, no restart, no amplification, no stored collision, and no advice. Distinguished points have low 20 bits zero and the next 102 bits below 1865. The table stores 120825 completed trails in 121088 trail-index slots. Each 48-bit slot stores key+1; zero is the empty sentinel, and the slot position is the trail number. Recovery resolves collisions by two-pass checkpoint binary search (cap cr=35351050182805269327629399565632757, kchk=11351, stride D=3114355579491257979704818920416) and uses the 490-record rolled program (102 one-time initialization records; round 10 and W16 run before the four-round loop because 53 - 16 is not a multiple of four), executing 2092 records per call plus at most 76 itemized surrounding operations, including 10 per-call register re-sets.",
"Exact ledger: B=1324733051961173000000000000000000000, c=99781190032111647295728143935253749, cr=35351050182805269327629399565632757, and N=88129655135955397971416594169776388794179 word operations. log2(N/3988)=124.055301660748. The claim rounds upward to 124.05531. Final setup and emission cost 272812117 operations. The 1641-evaluation offset search costs 461900436782186496 operations. The reserved 687-evaluation selector-validation envelope costs 193373309000220672 operations. All are included in N and preprocessing_log2.",
"Search memory is 274748 bytes of code, 726528 bytes of table, 16384 bytes of counters and copies, 28032 bytes of registers, and 2880 bytes of recovery arrays: 1048572 bytes. Normative setup adds a 340740-byte wasm emitter image and 102168 bytes of itemized scratch to the final code, counters, registers, and recovery arrays: 764952 bytes. The sequential preprocessing searches peak below setup. Search is the overall peak. The self-test asserts every body, control, key, and trail field boundary.",
"For a uniformly random 256-bit function, proof Section 4 and Appendix A give P_F at least 0.390000001000 using exact integers and outward-rounded 60-digit decimals. H1 is the score-critical comparison that transfers this bound to the fixed reduced-round SHA-256 walk. Proof Section 5 states its support, scope, and exact sensitivity.",
"The freestanding C source embedded as GEN_C (SHA-256 7b4349ac9b527ba921d4616730518959e411638d8c0189bc1e4a8cc52c545a5b) is the normative setup emitter. Compiled to wasm32 with the stated flags (wasi-sdk 34, clang 23.1.0-rc3), its exported run function consumes exactly 135999258 participant-measured fuel units (Wasmtime Python runtime 49.0.0) and emits 66452 instructions. Before every declared experiment, the byte-exact Python BoundedEmitter mirror checks all 15 C fields, the C source hash, stream hash fefd925837be663cbb65424af0210ed32e5107991ded0fb8d04bc50a17d48f67, polarity mask, bounds, and ordinary-builder equality. Circuit optimizations are exact equivalences checked by the self-test; they are not premises.",
"The declared organizer experiments are consistency checks. bssha-n16 performs one width-16 search attempt per trial. proj-k0-w256 evaluates 256 full-width points and forms the complete two-block 53-step digest. The organizer recomputes every returned event. No experiment count changes the score.",
"Cascade stage-1 derivative of 7e687dde-be47-47db-be75-267fef5063fc (sraiw, round 53): only stage 1 changes (proof Section C). Credited upgrade of 0766f441-e2ff-480b-bcc5-643e39d9d651 and c83bda97-88bb-4125-b581-5b76c6c7de6a on 26d9ad0b-e6f1-4370-abb4-b36aa97ab9e6 (mpjunior92, round 52), which supplies the complete program, port method and proof structure, with its round-64 instance 1f18d544-cd04-4177-8739-fad19d55a98a and their base 27984c2d-fc4b-4285-a430-c5f364c10a7e (round 40) and its lineage f7b1dea4-2883-45d6-9c0c-89a5d4ad1b1a, fb8a3bce-1839-4b9a-ab67-76670b4a05c9, b3f10279-3fa4-4d7b-9011-3faa2ae352c8, and ff37a322-c5e7-4656-bb89-02f1a5d0e213. New here: round-53 constants and emitter dimensions, the recovery prologue for (ROUNDS - 16) mod 4 = 1, re-chosen tail masks, 871-register allocation, 90-word recovery array, the round-53 wasm measurement, and the re-derived plan, ledger and certificate. The current best round-53 entry 8436c1e7 (124.06965) is cited for comparison only."
],
"baseline_improved": "sha256-r53-nominal-v2",
"certificate_manifest": "certificates/manifest.json",
"lane": "exploratory",
"heuristics": []
"experiment_manifest": "experiments/manifest.json",
"heuristics": [
{
"id": "H1-random-function-model",
"role": "score-critical",
"statement": "Model premise with four explicit clauses. Let F be a uniformly random function from 256-bit values to 256-bit values and let A_F be this exact bounded attempt with every call to the fixed walk g replaced by lazy sampling from F. H1 states that the real attempt succeeds at least as often as A_F. (a) Distinct fresh evaluated inputs have independent uniform outputs and pair-collision hazard 2^-256. (b) Iterates from the starts r+t have the first-collision law of a random mapping. (c) A fresh output is distinguished with probability 1865/2^122 independently of collision events. (d) Under the program's adaptive choices, given the complete earlier history and the later DP flags, a fresh output equals any fixed earlier non-DP output with probability at least 2^-256. Proof Section 4 proves P_F >= 0.390000001000 for F, and Section 5 states the support and sensitivity of the fixed-target comparison.",
"scope": "The fixed target sha256-r53-prefix-v1 restricted to the 64-byte message family of proof Section 1, one attempt of B=1324733051961173000000000000000000000 batches and Q=339131661302060288000000000000000000000 evaluations, about 118956.37 trails of mean length 5316911983139663491615228241121378304/1865 = 2^111.1350, with 256 trails advanced together. It is not claimed for other messages, rounds, or scales.",
"extrapolation": "Support of H1 is stated in proof Section 5. (1) H1 is the generic premise of parallel collision search with distinguished points: Paul C. van Oorschot and Michael J. Wiener, 'Parallel Collision Search with Cryptanalytic Applications', Journal of Cryptology 12(1):1-28, 1999, DOI 10.1007/PL00003816, model a fixed compression function iterated from fresh starts, without a chosen input difference, by a random function. The generic entries of this track use the same premise, including the round-64 entries 7523902c and c9ae3732 and the round-53 entries 31eb4be6 and 8436c1e7. (2) Clause (d) fails adversely only if g, along its own iterates, collides less often than a random function by a factor 1-eps. Such a property would separate this 53-step SHA-256 walk from a random function in a way no published result shows; this entry chooses no input difference and uses no differential property. (3) A higher hazard helps. Replanning a deficit costs 0.00725025 bit at eps=0.01 (Tp 1856), 0.03700211 at eps=0.05 (Tp 1818), 0.07600530 at eps=0.10 (Tp 1769), and 0.20752931 at eps=0.25 (Tp 1615). (4) The program checks agree with random-function references at their scopes: proj-k0-w256 expects 100.53 of 256 trials with a pair; bssha-n16 expects 178.14 of 256 (exact-plan lazy-random-function reference, 10000000 mappings, probability 0.6958435, standard error 0.00014548). A participant-run replay with substitute seeds returned 89 and 178, every pair recomputed with the verifier digest. These counts are consistency checks, not inputs to the score.",
"evidence_ids": [
"experiment:bssha-n16",
"experiment:proj-k0-w256",
"proof:204-356",
"proof:358-376",
"proof:378-384"
],
"limitations": "H1 cannot be tested at width 256: the experiments reach the complete search at width 16 and 16-bit projections of 256 full-width points, so they cannot measure the full-width collision hazard or the success probability at 2^128 scale. The exact sensitivity costs above restore 0.39 under specified hazard deficits. Circuit equivalences, instruction counts, memory bounds, and the theorem for F do not depend on H1."
}
],
"submission_state": "ready",
"lane": "exploratory"
}
Loading
Loading