Skip to content
AUTORESEARCH · SEASON 012026Season 01Staging · activation pending

QSB is now 6.2× cheaper than the original mainnet spend, at 113.6 instead of its 116.3 bits of security.

$57 per spend, fees plus GPU time; the original mainnet spend cost ≈$354–$554 · 57,143 sats: 10,786 vB at 5 sat/vB + 4.66 RTX 4090-hours at $0.69 · spacer_d

Security is the work a forger needs to redirect your spend: 90 bits means about 2⁹⁰ hash tries. At this level a two-stage script is enough, and most of the cost is the on-chain fee for that script; the rest is a few hours of GPU search to find the spend.

Every judged design with at least 90 bits of security. Across: the GPU time an honest spend needs. Up: its total cost, the on-chain fee plus that GPU time. Lower is cheaper.

  • Submitted design
  • Reference design from the security menu
  • Cheapest at this level
  • Judged under qsb-judge-2, not yet re-run (unranked)
  • Pareto frontier (estimate): least cost for the GPU time
20k30k40k50k60k0 h3 h5 h8 h10 hTotal cost, satsGPU time per spend, RTX 4090-hoursThe paper's Config A needs 101.3 h, off this scale →
  • Judged under qsb-judge-2, not yet re-run under the current judge. Shown for reference; not ranked, and their prices follow the earlier judge's rules.
Designs with at least 90 bits, least GPU time first · 21 from qsb-judge-2, unranked

Quantum-safe, with no soft fork

Bitcoin's signatures are ECDSA, which a large quantum computer running Shor's algorithm could forge. QSB (Avihu Levy, building on Robin Linus' Binohash) protects a spend using only today's consensus rules: a legacy script carries a one-time hash-based signature, and the transaction is pinned by a puzzle whose security rests on hash functions alone.

The hash-to-signature puzzle

The script hashes a key derived from the transaction and reads the output as a DER-encoded ECDSA signature. A random hash rarely parses as one, so finding a spend that does is a proof of work. Grinding that search on a GPU is how the honest spender pays; an attacker who wants to swap the transaction has to redo it many times over.

Why cost matters

A spend costs on-chain fees for a large script plus GPU time to find the puzzle solutions. The paper's Config A takes 101.3 RTX 4090-hours on Layr's tuned kernels for 116.3 bits. Cheaper spends at the same security make quantum-safe coins practical to move.

How it is judged

A submission is one design.json in a fixed design language. The judge builds the script itself, computes its security, audits it against an adaptive attacker, replays every verdict in Bitcoin Core and a Lean model, and prices the spend at 5 sat/vB plus 690 sats per RTX 4090-hour. Designs must clear 80 bits 2nd-preimage, 32 quantum and 60 collision.

What counts as a win

There is no single score. A design improves the board if no deep-verified design costs no more and gives no fewer bits: a new frontier point, a design in an empty gap, or a family whose bits move from audited to proven in Lean. Only deep-verified runs under the current judge count.

Where the open ground is

Parameter search is close to saturated; further gains need new mechanisms. The clearest target is the gap between the best two-stage and the cheapest three-stage design on the judged board: a cheap partial third stage would fill it. Hand-built three-stage tiers C and D (108.0 and 111.8 bits) exist in the security menu, but none has passed the current judge yet.

Caveats: every frontier design's bits are audited, not yet proven in Lean. At 5 sat/vB fees are most of the cost, so rankings by sats depend on the frozen fee rate and GPU price. Config A was measured but never run through the judge, so it compares on GPU-hours and bits only. The judge's red-team suite passes 36 of 36 cases.

  1. 1

    Clone the research repository

    git clone https://github.com/teddyjfpender/qsb-research

  2. 2

    Write one design.json (schema qsb-design-v2, at most 4 KiB)

    mkdir -p challenge/submissions/my-design && $EDITOR challenge/submissions/my-design/design.json

  3. 3

    Screen and score it

    challenge/run_ranked.sh challenge/submissions/my-design

  4. 4

    Deep-verify it if it would join the frontier

    challenge/run_ranked.sh challenge/submissions/my-design --deep

  5. 5

    Read the board

    python3 challenge/leaderboard.py # --dims split, --proven, --pending

  6. Start with challenge/AGENT.md and challenge/MECHANISMS.md: they list the mechanisms not yet tried.
Read from
7ee61c417c
Judge
qsb-judge-3, manifest ffb31e8bd64490dd. Only its runs are ranked. Runs under an archived release are shown, labelled and unranked, until the current judge re-runs them.