01 · Certified
2.00×
[[192,8,24]], distance proven exactly by a MIP solver — every coset solved. The solver even corrected the fast-decoder estimate from 26 down to 24.
A useful quantum computer is mostly error correction. Sutra hunts for the codes that do it better — and a reproducible search found ones that beat IBM's Gross code on the field's figure of merit, one of them certified by an exact solver. The same engine takes aim at quantum computing's other unsolved, billion-dollar problems.
The certified result — [[192,8,24]], distance proven exactly. A real bivariate-bicycle code on its torus, drawn from the published polynomials.
Logical qubits k, distance d, physical qubits n. It rewards encoding more logical qubits and tolerating more errors, while penalising qubit overhead — the single number that captures whether a code is a good deal.
Every logical qubit you can actually compute with is built from hundreds or thousands of noisy physical qubits, held together by a quantum error-correcting code. The code's quality is the exchange rate between the two — and it dominates the cost of fault-tolerant quantum computing. Sutra treats finding better codes as a search problem: we built a validated evaluator, swept 60,000 candidates under a real hardware-connectivity constraint, and found bivariate-bicycle codes that beat IBM's Gross code on the standard figure of merit, k·d²/n — one of them certified by an exact solver. The same machinery points at the field's other open problems in fault tolerance.
Logical qubits (k), error tolerance (distance d), physical-qubit cost (n). k·d²/n is the field-standard single number for whether a code is a good deal. IBM's Gross [[144,12,12]] code scores 12. Our codes score 22–53 — 1.8× to 4.4× better — under a heavy-hex connectivity constraint Gross does not satisfy.
Bivariate-bicycle codes are defined by two polynomials over a torus. We search that space — random at scale, then evolutionary — scoring each candidate with an evaluator first validated to reproduce all five of IBM's published BB codes exactly. 14 of the top 50 beat Gross after rigorous re-checking.
Fast decoders give an upper bound on a code's distance; they can over-state it. For our lead code we ran an exact MIP solver over every logical coset — and it corrected the estimate from 26 down to 24. We label every distance as exact or bounded, and we publish a code that failed verification to make the point.
A higher k·d²/n buys you less margin elsewhere: our highest-scoring codes currently have a circuit-level threshold below today's IBM gate-error rates, and the hardware demo is designed but not yet run. We report that as the central result, not a footnote — it is what the next phase is for.
[[192,8,24]], distance proven exactly by a MIP solver — every coset solved. The solver even corrected the fast-decoder estimate from 26 down to 24.
[[196,18,24]] — the highest figure of merit found, and just 10.9 physical qubits per logical vs Gross's 12. Distance is a rigorous upper bound, not yet exact.
The honest part: the highest-scoring codes currently have a circuit-level threshold below today's hardware. The tradeoff is the finding — and the next phase's target.
The same search-and-certify machinery generalises. These are the adjacent open problems we are built to attack — scoped honestly as the venture's expanding frontier, not as solved work.
Real-time, hardware-aware decoding and hook-error-avoiding syndrome schedules — what turns a good code into a usable one.
Exact distance via MIP/SAT, generalised so that a code's protection is proven, not merely estimated by a fast decoder.
Codes with asymmetric distance (d_X ≠ d_Z) matched to a chip's real, biased noise — squeezing more protection from the same qubits.
Codes designed to sit on a specific processor's connectivity graph from the start, not retrofitted — cutting routing overhead.
Magic-state and lattice-surgery layouts on high-rate qLDPC codes — the path from quantum memory to quantum computation.
Generalised and multivariate qLDPC families, searched with the same validated, reproducible pipeline.
In quantum error correction the gap between a claimed distance and a certified one is where careers go to die. So we keep a ledger, and label every line exactly.
Quantum computing's bottleneck is not the number of physical qubits; it is how many of them each logical qubit costs. That overhead is set by the error-correcting code, the decoder, and how well both fit the hardware. Sutra's bet is that this whole stack is searchable and co-designable — that better codes, better decoders, and hardware-tailored fault tolerance can be discovered systematically rather than hand-crafted one paper at a time. We start where we already have a verified, reproducible result — better codes — and extend outward into the adjacent million-dollar problems of fault tolerance.
This is the stage where backing matters most — before the revenue, while the result is hardened from “strong and reproducible” into “undeniable.” The decisive next steps are verification compute and a hardware demo.
Sponsorship & partnership — not equity. There is no product yet, and we won't pretend otherwise.
Every code, its lattice, its figure of merit, and exactly how its distance was verified — open and reproducible.