Research · QONTOS

RESEARCH AT A GLANCE

Seven contributors. One threshold. One programme.

Every research target on QONTOS-1 ties to a named contributor in the error budget, a numerical gate of the QONTOS roadmap, or an open mathematical question in the field. The five numbers below frame the v5.4 risk posture without implying fault tolerance at G4.

ERROR CONTRIBUTORS

7 budgeted explicitly · per-cycle

ABOVE THRESHOLD

0.75 pp 1.75 % Q1 budget · ~1.00 % threshold

DECODER BUDGET

≤ 5.5 µs stabiliser → Pauli frame

Q1 CODE DISTANCE

d = 5 first logical qubit · validation only

RSA-2048 CONTEXT

external benchmarked against Gidney-style estimates

RESEARCH PROGRAMME · THREE AREAS

Quantum error correction. Real-time control. Compilation.

QONTOS funds research in the three areas where progress most directly retires risk on the QONTOS-1 hardware. Each area has named gates, named owners, and a public deliverable.

AREA A · QEC STRATEGY

Codes that scale on heavy-hex.

The rotated surface code is the workhorse; the research charter is to find codes and decoders that move closer to optimal for QONTOS-1's heavy-hex lattice and limited connectivity, including bivariate-bicycle, colour, and qLDPC families.

Rotated surface code, d ∈ {5, 7, 11, 15} bench sequences
Bivariate bicycle / GKP candidates · simulated
Lattice surgery across module boundaries
Open: high-rate codes on bounded-degree lattices

AREA B · REAL-TIME CONTROL

Decoders that fit a 5.5 µs budget.

The decoder must close the loop before the next stabiliser cycle. The research charter is to find decoders that are both accurate enough and fast enough on FPGA/ASIC fabric, and to compile their detector graphs once per calibration epoch.

Sparse-blossom MWPM · pipelined FPGA
Union-find · constant-latency · default at G2
ML-assisted decoder · explored at d ≥ 11
Open: streaming decoders with provable latency

AREA C · COMPILATION

Algorithms that fit the budget.

Even a million physical qubits cannot run Shor naively. The research charter is to compress logical resource counts: better factories, better arithmetic, better algorithmic primitives. The output is open compilers and reproducible resource estimates.

Magic-state factory cadence · (15, 1, 3) → (15, 1, 9)
Modular exponentiation · Windowed arithmetic
Resource estimator under qontos-resource
Open: T-count lower bounds on algorithmic primitives

MATHEMATICAL FOUNDATIONS · § A.1

Five equations the rest of the page rides on.

QONTOS-1 research is grounded in five equations: the surface-code threshold, the logical-error scaling, the magic-state output infidelity, the cross-module Bell-pair fidelity, and the resource-overhead estimator. Every other figure in this paper resolves to one of them.

EQ.1 · THRESHOLD

p_th ≈ 1.0 % · circuit-level depolarising

The rotated surface code admits a fault-tolerant threshold of approximately 1 % per gate under a depolarising-noise model. The number is decoder-dependent; sparse-blossom matches it to within 0.05 pp. Crossing the threshold is necessary but not sufficient: useful operation requires well below.

Anchors: § 8.3 · § 9.4

EQ.2 · LOGICAL ERROR

ε_L ≈ a · (p_phys / p_th)^((d+1)/2)

Well below threshold, logical error per cycle is exponentially suppressed in the code distance. a ≈ 0.03 is a code-family prefactor; d is the rounded-up odd distance. Halving p_phys roughly halves d for the same target ε_L.

Anchors: § 8.3 · Fig. 2

EQ.3 · MAGIC STATE

p_out ≈ 35 · p_in³

The (15, 1, 3) distillation protocol turns fifteen inputs of infidelity p_in into one output of infidelity p_out. The cubic scaling is what makes a hierarchical ladder of factories productive at all.

Anchors: § 8.6

EQ.4 · BELL-PAIR FIDELITY

F_Bell = (η · F_t² · F_d) · g(τ)

The Bell-pair fidelity across the photonic seam is the product of heralding efficiency η, transducer fidelity F_t squared, detector fidelity F_d, and a time-jitter penalty g(τ). QONTOS-1 targets F_Bell ≥ 0.85 at the G3 gate.

Anchors: § 6.4 · § 9.6

EQ.5 · RESOURCE OVERHEAD

N_phys ≈ 2 d² · N_log + N_factory

Total physical qubits is approximately the per-logical surface-code patch overhead (2 d²) times the logical-qubit count, plus the factory footprint. The factory term dominates for short algorithms; the patch term dominates for long ones.

Anchors: § 8.7

EQ.6 · DECODER LATENCY

τ_dec ≤ τ_T1 · ln(1 / p_target)

The decoder must close before drift overruns the syndrome. The right-hand side is the coherence-limited deadline. QONTOS-1 sets τ_dec at 5.5 µs, comfortably under the T₁-derived deadline at all targeted operating points.

Anchors: § 8.4 · Fig. 4

ERROR BUDGET · § 8.2

Seven contributors. One threshold. A 0.75 pp engineering gap.

The effective per-cycle physical error rate is the sum of seven named contributors. The QONTOS-1 target sums to 1.75 %, exceeding the rotated-surface-code threshold of approximately 1 %. Closing that gap is a portfolio, not a single fix.

FIG.1 · PER-CYCLE ERROR BUDGET · TARGET vs THRESHOLD

0.00 % 0.25 % 0.50 % 0.75 % 1.00 % 1.25 % 1.50 % 1.75 %

Single-qubit gate

0.10 %

Two-qubit gate

0.50 %

Readout

0.30 %

Idle / decoherence

0.20 %

Calibration drift

0.15 %

Decoder latency

0.10 %

Cross-module Bell-pair

0.40 %

Total · target

1.75 % · sum of contributors

Total exceeds the threshold by 0.75 percentage points. Three contributors carry most of that excess: the two-qubit gate (0.50 %), the cross-module Bell pair (0.40 %), and readout (0.30 %). They are the named owners of the largest research bets on QONTOS-1.

BIGGEST BET · TWO-QUBIT GATE

0.50 % → 0.10 % through tunable couplers and pulse-level optimisation.

Pulse re-shaping (DRAG, GST-informed) and tunable-coupler control retire approximately 0.30 percentage points. The remaining headroom moves with device improvements at the QONTOS-2 generation.

SECOND BET · INTERCONNECT

0.40 % → 0.05 % through raw-fidelity, added-noise, and purification improvements.

The heralded Bell pair is the structural cost of being modular. The path to retiring it is photonic: better transducers, better detectors, and per-link feedforward.

THIRD BET · READOUT

0.30 % → 0.10 % through TWPA upgrade and discriminator changes.

Higher SNR readout plus a learned classifier lifts assignment fidelity from 0.97 to > 0.995 within the present cryogenic envelope.

ROTATED SURFACE CODE · § 8.3

Exponential suppression as the lever.

Well below threshold, logical error per cycle is a steep function of how far below threshold the physical error rate sits. The exponent is set by the code distance.

ε_L ≈ a · (p_phys / p_th)^((d+1)/2) a ≈ 0.03 · p_th ≈ 1 % · d rounded up to the next odd integer

FIG.2 · LOGICAL ERROR PER CYCLE vs CODE DISTANCE · THREE OPERATING POINTS

QONTOS-1 is not claimed below threshold. The chart is retained as a research planning map: later scenarios only become engineering scope after measured reductions in physical error rate, decoder latency, and purified interconnect overhead.

LOGICAL QUBIT · ANATOMY

One logical qubit, twenty-five data qubits, twenty-four stabilisers.

QONTOS-1 stores a single logical qubit in a distance-five rotated surface-code patch. Each face is a stabiliser: X-type on alternating tiles, Z-type on the others. A stabiliser cycle measures all twenty-four in parallel; the syndrome stream feeds the decoder.

FIG.3 · ROTATED SURFACE CODE · d = 5 · X- AND Z-PLAQUETTES

STABILISER CYCLE

X and Z measured every 1 µs.

Each cycle measures all twenty-four stabilisers in parallel via the ancilla qubits. The result is a 24-bit syndrome, time-stamped, hashed, and emitted to the decoder.

DETECTOR GRAPH

One detector per syndrome change.

The decoder consumes detectors, not raw syndromes. A detector fires when consecutive stabilisers disagree. The detector graph is rebuilt once per calibration epoch.

LOGICAL OPERATORS

One row · one column.

The logical X is the product of X on one full row; the logical Z is the product of Z on one full column. Both are length-d strings; both commute with every stabiliser of the same type.

G3 ACCEPTANCE

ε_L < 10⁻⁶ on a sustained 24-hour run.

The acceptance criterion for the G3 gate is not the logical error rate at one cycle: it is the rate sustained against drift and cosmic-ray events across a calendar day.

SYNDROME EXTRACTION · § 8.4

One stabiliser, six time-steps, one classical bit.

Every stabiliser is measured by a six-step quantum circuit: prepare ancilla, four CNOT or CZ couplings to the data qubits at the corners of the plaquette, measure the ancilla in the X or Z basis. The result is one bit. Twenty-four of them, every cycle, in parallel.

FIG.S · X-STABILISER EXTRACTION CIRCUIT · ROTATED SURFACE CODE

The Z-type stabiliser uses the same skeleton with H gates on the ancilla and CZ couplings; the syndrome bit is ⟨Z⟩ on the ancilla. All 24 stabilisers run in parallel on different ancillae; the timing diagram is per-stabiliser. The output bit feeds Stage P2 of the decoder pipeline.

REAL-TIME DECODER · § 8.4

Stabiliser measurement to Pauli frame, ≤ 5.5 µs end-to-end.

QONTOS-1 uses a minimum-weight perfect-matching decoder in the sparse-blossom variant on dedicated FPGA fabric. Six pipeline stages are mapped to FPGA, ASIC, or host CPU according to their latency budget. The whole loop closes before the next stabiliser cycle starts.

FIG.4 · DECODER PIPELINE · 6 STAGES · FPGA + ASIC FABRIC

P1

Measure stabiliser measurement 1.0 µs FPGA

P2

Buffer syndrome buffer 0.2 µs FPGA

P3

Parse detector graph build 0.5 µs FPGA

P4

Match sparse-blossom MWPM 3.0 µs ASIC / FPGA

P5

Apply Pauli frame update 0.5 µs FPGA

P6

Verify soft check + provenance 0.3 µs host CPU

TOTAL · END-TO-END ≤ 5.5 µs matched to QONTOS-1 feed-forward budget · per stabiliser cycle

The 5.5 µs budget assumes one decoder instance per logical qubit. At distances d ≥ 11 the matching graph grows quadratically and the decoder is parallelised across ASIC tiles, with a stitching protocol that adds ≈ 0.5 µs per tile boundary. Tile boundaries are part of the cal record.

DECODER BENCHMARK · § 9.5

Three decoders, one device, three trade-offs.

QONTOS-1 implements three decoders against the same detector graph format. Sparse-blossom MWPM ships as the production decoder. Union-find is the fallback when latency is paramount. A neural-net decoder is benchmarked as research preview at distances d ≥ 11.

FIG.7 · DECODER COMPARISON · LATENCY × LOGICAL-ERROR · D ∈ {5, 7, 11, 15}

Decoder	d=5	d=7	d=11	d=15	Memory	Substrate	Status
Sparse-blossom MWPMproduction default	0.4 µs / 1.0×	0.9 µs / 1.0×	2.1 µs / 1.0×	3.0 µs / 1.0×	0.8 MB	ASIC / FPGA	SHIPS · G2
Union-findlatency-bound fallback	0.2 µs / 1.6×	0.4 µs / 1.7×	0.9 µs / 1.9×	1.3 µs / 2.1×	0.3 MB	FPGA	FALLBACK
Neural-net decoderresearch preview · d ≥ 11	—	—	1.8 µs / 0.9×	2.4 µs / 0.85×	4.2 MB	ASIC + TPU	RESEARCH

Sparse-blossom · d=15

3.0 µs · matches best · 1.0× error

Union-find · d=15

1.3 µs · 2.1× error

Neural-net · d=15

2.4 µs · 0.85× error · research

0 µs123455.5

Latency is wall-clock on Xilinx Versal AI Core at d=15 (median over 10⁴ syndrome inputs from the QONTOS-1 digital twin). Error column is the multiplicative factor vs sparse-blossom on the same input; lower is better. The neural-net decoder is faster than sparse-blossom at d=15 but the training pipeline is not yet production-grade.

LATTICE SURGERY · § 8.5

Two modules. One logical qubit. A photonic seam.

A two-module QONTOS-1 carries one logical qubit by merging two physical patches across the photonic interconnect. The seam is a row of heralded Bell pairs supplied by the cryogenic transducer at 1.5 K and consumed inside the syndrome cycle.

FIG.5 · MERGED PATCH ACROSS MODULE A AND MODULE B · BELL-PAIR SEAM

During the merge phase the joint stabilisers measured at the seam absorb the Bell-pair entropy; the patches behave as a single distance-d code for the duration of the surgery. Heralding success is part of the stabiliser cycle: if a Bell pair fails to herald, the cycle re-tries within the 5.5 µs decoder budget.

MAGIC STATES · § 8.6 · RESOURCE ESTIMATE · § 8.7

From (15, 1, 3) factories to a 2,048-bit Shor budget.

Non-Clifford gates ride on magic states. QONTOS-1 baselines the (15, 1, 3) protocol; the QONTOS roadmap raises the distillation rate until a 2,048-bit Shor instance fits in a one-week run. The numbers below show how far the present design sits from that endpoint.

FACTORY · (15, 1, 3)

15 noisy inputs, one clean output.

One high-fidelity magic state from fifteen lower-fidelity inputs, with output infidelity scaling as 35 · p_in³. At p_in = 10⁻³ the output infidelity falls below 4 × 10⁻⁸, sufficient for algorithms whose T-gate count stays under 10⁷.

Footprint: ≈ 200 physical qubits @ d=5
Cadence: 11 code cycles per round
Rate · base: 8–12 magic states / second
Output infidelity: < 4 × 10⁻⁸ at p_in = 10⁻³

RESOURCE · 2,048-BIT SHOR

RSA-2048 remains a benchmark, not a claim.

Recent resource estimates for 2,048-bit RSA factoring require millions to tens of millions of physical qubits depending on physical error rate, code distance, decoder assumptions, and factory design. QONTOS-5 treats this as an external benchmark and resource-closure exercise, not as a committed deliverable.

T gates: ≈ 1.5 × 10¹⁰
Logical qubits: ≈ 4,000 @ d ≥ 27
Overhead: 2d² ≈ 1,500× per logical
Total qubits: external-benchmark dependent

FIG.6 · MAGIC-STATE FACTORY CADENCE · 4 PARALLEL FACTORIES · 60 µs WINDOW

0 µs102030405060

Factory F1d=5 · level 1

distil 1 distil 2 distil 3 distil 4 distil 5

Factory F2d=5 · level 1

distil distil distil distil distil

Factory F3d=5 · level 1

distil distil distil distil distil

Factory F4d=5 · level 1

cal distil distil distil distil

Consumerlogical qubit · T pulls

T T T T T T T T T T T

Four factories run staggered so the consumer logical qubit never waits. The dark "cal" tile is a calibration recapture; the consumer pulls a T magic state on each algorithmic call. QONTOS-1 targets four factories per consumer at the d=5 base.

MAGIC-STATE LADDER · § 8.6.2

One ladder, three levels, three orders of magnitude.

A single-level (15, 1, 3) factory reaches p_out ≈ 4 × 10⁻⁸ at p_in = 10⁻³. Hierarchical ladders cube the residual at each level, at a cost of fifteen lower-level factories per upper-level one. Three levels reach p_out ≈ 10⁻²⁴, sufficient for any application QONTOS can target on the family arc.

FIG.9 · MAGIC-STATE LADDER · LEVEL 0 → LEVEL 3

QONTOS-1 carries Level-1 factories only; Level-2 enters service at QONTOS-3 once the per-module footprint can absorb 3,000 qubits without starving the data patch. Level-3 is reserved for the QONTOS-5 datacentre footprint.

T-COUNT · COMPRESSION TARGETS

Six algorithmic primitives. Four orders of magnitude of headroom.

Magic-state cost dominates the wall-clock of any practical FTQC algorithm. The research charter is to lower the T-count of each primitive without compromising the algorithmic guarantee. The six primitives below are the ones the QONTOS resource estimator targets.

FIG.T · T-COUNT BENCHMARK · NAIVE → QONTOS RESEARCH TARGET · LOG SCALE

Modular multiplication · 2,048-bit core of Shor's algorithm

naïve · 1.5 × 10¹⁰ windowed · 4.8 × 10⁷ target · 2 × 10⁵

Quantum phase estimation · 32-bit chemistry, finance

naïve · 1.6 × 10⁸ qubitised · 6.4 × 10⁶ target · 8 × 10⁴

VQE · H₂ · 4-orbital variational baseline

naïve · 4.0 × 10⁵ opt · 6 × 10³ target · 2 × 10²

QSP / signal processing kernel · width 64

naïve · 1.4 × 10⁷ block-encoded · 9 × 10⁴ target · 5 × 10³

Grover · 32-bit search oracle dominated

naïve · 8 × 10⁷ amortised · 1.4 × 10⁶ target · 6 × 10⁴

HHL · linear system 128×128 well-conditioned

naïve · 3 × 10⁷ qubitised · 1.1 × 10⁶ target · 9 × 10⁴

10⁰10²10⁴10⁶10⁸10¹⁰

naïve compilation state-of-the-art QONTOS research target

All numbers are T-gate counts after standard Clifford+T decomposition. Lower is better. The QONTOS targets are independently published in the open qontos-resource repository and are auditable per algorithm.

VERIFICATION · BENCHMARKING · § 9

Every reported number, anchored to a protocol.

No QONTOS-1 number sits in the whitepaper without a measurement protocol behind it. Each protocol pins the gate set, the sample size, the bias model, and the reproducibility envelope. Calibration epochs are explicit; cross-epoch comparisons are reported separately.

SINGLE-QUBIT · RB

Clifford randomised benchmarking.

Sequences of length m ∈ {1, 2, 4, …, 512} sampled from the single-qubit Clifford group, each inverted by its computed final Clifford. Interleaved-RB attributes error to specific gates (X, X/2, Y, Y/2). Calibration re-baseline triggers when the drift envelope exceeds 2 × 10⁻⁴ from the previous epoch.

Sample: 128 circuits · 1,000 shots
Bias model: SPAM-corrected
Target: F_1q > 0.9995

TWO-QUBIT · INTERLEAVED RB · XEB · MIRROR

Three complementary methods.

Interleaved-RB for per-gate CZ error, cross-entropy benchmarking on random circuits of depth 10 to 30 for circuit-level fidelity, and mirror benchmarking for SPAM-robust single-number estimates. The QONTOS-1 two-qubit target is the worst-case value across all three methods on any coupler in the lattice.

Sample: 256 circuits · 5,000 shots
Bias model: worst-case · all couplers
Target: 2Q error ≤ 5×10⁻³

SYSTEM · STABILISER CYCLE · BELL-PAIR TOMOGRAPHY

Code as the benchmark.

Repeated stabiliser-cycle experiments measure logical error rate as a function of code distance and number of cycles, directly validating ε_L = a · (p / p_th)^((d+1)/2). Cross-module Bell-pair tomography exercises the interconnect at system scale.

Sample: 10⁵ cycles · per d
Bias model: device + decoder
Target: F_Bell ≥ 0.85 · sustained 1 h

REPRODUCIBILITY · PROVENANCE

Numbers come with their lineage.

Every reported value is bound to a provenance record that captures the calibration epoch, the AWG and digitiser firmware versions, the decoder weight matrix, the compiler version, the device temperature time series, and a hash of the raw measurement data. Re-running the same protocol against the same calibration epoch reproduces the result to within statistical uncertainty. Re-running against a different epoch is treated as a distinct measurement and reported separately.

cal epoch · ID + timestamp firmware · AWG + digitiser SHA decoder · weight matrix hash compiler · qontos-sdk version thermal · 100 mK time series raw data · SHA-256 manifest

CALIBRATION CADENCE · § 9.8

One device, three calibration loops, one daily epoch.

QONTOS-1 runs three concurrent calibration loops at different cadences: pulse-level retuning every 10 minutes, gate-level sequence recharacterisation every 4 hours, and a full device epoch every 24 hours. The epoch ID is part of the proof manifest of every job.

FIG.8 · CALIBRATION GANTT · 24-HOUR ROLLING WINDOW

00 h040812162024

L3 · DAILY EPOCH full device · 45 min · 1× per 24 h

EPOCH

L2 · GATE RECHAR 2-qubit RB · 12 min · 6× per 24 h

RB RB RB RB RB RB

L1 · PULSE RETUNE DRAG · 90 s · 144× per 24 h

QUEUE · JOBS scheduled around calibration · published

Pulse-level retuning runs every 10 minutes against control qubits flagged by the drift envelope; gate recharacterisation runs every 4 hours regardless; the daily epoch is the only event that suspends user traffic. Cross-epoch comparisons of benchmark results are always reported with epoch IDs, never aggregated silently.

OPEN PROBLEMS · 2026 H2

Six questions QONTOS pays to answer.

Each open problem below has a defined acceptance criterion, a public dataset or simulator, and a named owner inside the research programme. Each retires risk on a specific QONTOS-1 gate. Submissions are tracked under qontos-research.

OPEN · OP-01

Streaming decoders with provable latency.

Decode infinite syndrome streams with bounded worst-case latency. Closes the gap between sparse-blossom (fast in average) and matching theory (slow in worst case).

retires risk on · G3 decoder budget

OPEN · OP-02

High-rate codes on bounded-degree lattices.

QONTOS-1 has degree-3 connectivity. Surface code is rate-1/d² there. Find an asymptotically higher-rate qLDPC code that remains decodable in 5 µs.

retires risk on · QONTOS-3 footprint

OPEN · OP-03

Magic-state distillation with γ < 1.

The (15, 1, 3) factory pays γ ≈ 1.3 in shots. Find a distillation routine whose shot overhead γ drops below unity at the same target output infidelity.

retires risk on · 2,048-bit Shor budget

OPEN · OP-04

Photonic Bell pairs with measured noise closure.

QONTOS-1 now accepts η ≥ 0.1% as the base transduction target, with raw Bell-pair fidelity ≥ 0.85 and false-herald fraction below 1% during the G3 run. Higher fidelity logical operations require purification overhead to be budgeted explicitly.

retires risk on · G3 Bell-pair fidelity

OPEN · OP-05

Cross-module lattice surgery without re-init.

Today's merge protocols re-initialise the seam each cycle. Find a surgery routine that keeps state coherence across cycles, eliminating the per-cycle 0.2 µs reset.

retires risk on · sustained logical operation

OPEN · OP-06

T-count lower bounds for VQE / QPE.

The community estimates T-counts via Trotter steps and oracles. Find lower bounds that are tight enough to predict QONTOS-1 wall-clock to within 2× on chemistry primitives.

retires risk on · application targeting

PUBLICATIONS · ARCHIVE

Six papers. Open source. Citation-tracked.

The publication record below is the in-house output of the QONTOS research programme to date. Each entry links to the arXiv preprint and to the reproducibility bundle in qontos-research. Citation counts refresh weekly from OpenAlex.

PREPRINT · arXiv:2605.04217

Heralded Bell-pair distribution across a millikelvin module boundary.

Tamilselvan R., Khoury A., Mehrabi N., Ortiz-Suslow D., et al. · 2026 · Phys. Rev. Applied (in review)

Demonstrates how the v5.4 transducer envelope is closed: base η ≥ 0.1%, aggressive η ≥ 0.5%, research η ≥ 1%, with pump thermal load and added-noise acceptance criteria.

citations · 12code · qontos-research/bell-1fields · QED · photonics

PREPRINT · arXiv:2604.08901

Sparse-blossom MWPM on FPGA: hitting a 3 µs match window at d=15.

Park J., Mehrabi N., Tamilselvan R., et al. · 2026 · Nature Electronics (under review)

Pipelined sparse-blossom decoder synthesised on Xilinx Versal AI Core. Worst-case match latency stays under 3 µs for distances up to d = 15. Open RTL.

citations · 31code · qontos-research/decoder-fpgafields · QEC · FPGA

PUBLISHED · PRX Quantum · 2026

Error budgets for two-module quantum computers: a structured methodology.

Tamilselvan R., Ortiz-Suslow D., et al. · 2026 · PRX Quantum 7, 020315

Formalises the seven-contributor budget used in the QONTOS-1 whitepaper. Provides a portable methodology for any modular architecture with a photonic interconnect.

citations · 47code · qontos-research/budget-1fields · QEC · methodology

PUBLISHED · Quantum · 2026

Magic-state factory cadence under concurrent decoder load.

Mehrabi N., Park J., et al. · 2026 · Quantum 10, 1283

Closed-form expression for the steady-state T-pull rate when N factories share a decoder. Resolves a long-standing ambiguity in roadmap papers.

citations · 21code · qontos-research/cadencefields · resource estimation

PREPRINT · arXiv:2603.11077

A reproducibility manifest for quantum benchmarks.

Khoury A., Tamilselvan R., et al. · 2026 · Nature Computational Science (in review)

Defines the provenance record described in this paper as a portable artifact. Adopted by three peer groups for cross-vendor benchmark comparison.

citations · 9code · qontos-research/manifestfields · methodology · reproducibility

PUBLISHED · Phys. Rev. A · 2025

Tantalum-on-silicon transmons: T₁ statistics from chiplet modules.

Parquet · proof-bundled

LICENCE

Apache-2.0

SHA

v5.4 · c189..40

Digital-twin re-runs of every dataset above. Use to disambiguate hardware-specific effects from simulator predictions, and to evaluate cross-system decoders.

REPRODUCIBILITY · SEVEN PRINCIPLES

If it cannot be reproduced, it does not count.

Every published QONTOS research result honours seven principles. The principles are not aspirational; they are gating criteria. Reviewers reject results that violate them. The list is short and binding.

01

Every number names its protocol.

No number appears without the protocol that produced it, the sample size, the bias model, and the calibration epoch under which it was measured. Aggregate numbers report the unbiased estimator and the 95 % confidence band.
02

Every figure ships with its data and its code.

Each figure in this paper has an entry in qontos-research/figures with the raw data file, the plot script, and the SHA-256 of both. Independent reviewers reproduce the figure from the bundle.
03

Calibration epochs are never merged.

Cross-epoch comparisons are reported as distinct rows, with epoch IDs. Aggregating across epochs is forbidden unless the protocol explicitly allows it and the drift envelope is reported alongside.
04

Negative results are published.

Failed protocols and below-target benchmarks are published with the same provenance as the positive ones. The cosmic-ray dataset is the canonical example: it is a catalogue of catastrophic events.
05

External reviewers run on the same data.

Time allocation includes a reviewer quota: independent labs run the protocol against the same calibration epoch and publish their own analysis. Conflicts are documented in the steering committee minutes.
06

Compiler and decoder versions are pinned.

The compiler and decoder versions used to produce any reported number are pinned in the manifest. Upgrading the compiler resets the benchmark series; the old series is preserved under its commit SHA.
07

The roadmap commits to publishing what does not work.

The five-gate roadmap reports gates whether or not they are met. A missed gate is reported with the same diligence as a met one; the engineering response is part of the publication.

COLLABORATIONS · ACADEMIC + INDUSTRY

Eight institutions. One device. Open notebooks.

QONTOS does not own QEC research. The institutions below run protocols against QONTOS-1 telemetry, contribute open code, or co-author results. Time on the device is allocated through the steering committee on a published quota.

RESIDENCY · QEC

Zhyra Quantum Research Institute · Abu Dhabi

Host institution. Owns the decoder pipeline reference implementation and the error budget methodology.

PROTOCOL · DEVICES

NYU Abu Dhabi · Center for Quantum Information

Co-authored the tantalum-on-silicon transmon study. Runs T₁ statistics on QONTOS-1 chiplets monthly.

PROTOCOL · DECODER

TUM · School of Computation

Streaming-decoder open problem. Provides reference RTL for the sparse-blossom variant on Xilinx Versal.

PROTOCOL · PHOTONICS

Stanford · Ginzton Lab

Electro-optic transducer characterisation. Owns the cryogenic loss budget that feeds the heralding efficiency target.

PROTOCOL · QEC

ETH Zurich · Quantum Device Lab

Bivariate-bicycle / qLDPC code evaluation on heavy-hex. Joint publication track for Area A.

PROTOCOL · ALGORITHMS

Tsinghua · IIIS

Resource-estimator co-development. Independent reproduction of the 2,048-bit Shor budget under different assumption sets.

INDUSTRY · CONTROL

Quantum Machines · OPX 1000

Real-time loop integration. Joint publication on FPGA-resident MWPM at d=15.

INDUSTRY · DETECTORS

Single Quantum · SNSPD

Detector calibration for the photonic seam. Owns the < 100 ns timing-jitter spec.

EXTERNAL REVIEW BOARD · SCIENTIFIC ADVISORY

Six advisors. Three continents. No financial interest.

The QONTOS scientific advisory board reviews the research charter quarterly and the gate criteria annually. Advisors hold no equity, take no consulting fee, and publish their independent assessment alongside the steering-committee minutes.

JG

Dr Julia Goldberg

Quantum error correction · ETH Zurich

Specialty: bivariate-bicycle codes, decoder design for non-uniform lattices. Chairs the Area-A review.

RM

Prof Ravi Menon

Resource estimation · Tsinghua IIIS

Specialty: algorithmic compilation, T-count optimisation. Reviews the resource-estimator releases.

FK

Dr Frieda Kohl

Cryogenic photonics · Stanford Ginzton

Specialty: electro-optic transducers at millikelvin. Reviews the photonic-interconnect roadmap.

YH

Prof Yusuf Hassan

Superconducting devices · NYUAD

Specialty: tantalum-on-silicon transmons, T₁ statistics. Reviews the device-physics targets.

CW

Dr Catherine Whitfield

FPGA decoders · TUM

Specialty: streaming-decoder theory. Chairs the open-problem panel on bounded-latency decoding.

SR

Prof Sergio Ramírez

Open-science methodology · UC Berkeley

Specialty: reproducibility manifests, dataset governance. Independent reviewer of the seven principles.

GRANTS PROGRAMME · 2026 H2

Three tracks. Public proposals. Public outcomes.

QONTOS funds external research aligned to the six open problems. Proposals are submitted through the steering committee; awarded grants come with device time, a research engineer pair-up, and a publication commitment.

TRACK A · DECODER

Streaming-decoder seed grant.

Bounded-latency decoders with formal worst-case guarantees. Targets OP-01 in the open-problems list.

Budget: $ 180 k · seed
Device time: 120 h · per quarter
Deliverable: open-source RTL + paper
Duration: 12 months

TRACK B · CODE

High-rate qLDPC fellowship.

Codes that improve on rate-1/d² on bounded-degree lattices, with a decoding strategy that fits the 5 µs budget. Targets OP-02.

Budget: $ 400 k · 2-year
Device time: 240 h · per quarter
Deliverable: code + decoder + paper
Duration: 24 months

TRACK C · ALGORITHM

T-count lower bounds.

Tight lower bounds on the T-count of chemistry, ML, and optimisation primitives. Targets OP-06.

Budget: $ 120 k · seed
Device time: 40 h · per quarter
Deliverable: open repository + paper
Duration: 12 months

Applications open quarterly. Submit through grants@qontos.xyz with a 4-page proposal, a track choice, and proof of independence. Reviewed by the external scientific advisory board within thirty days. Awards announced publicly.

CONFERENCES · TALKS · COVERAGE

Twelve venues. Eighteen months. One record.

The QONTOS research output is presented at peer-reviewed venues and tracked publicly. The list below is the rolling eighteen-month talk record. Slides and recordings are linked from the qontos-research repository when speaker rights allow.

2026 · KEYNOTE

"Two-module quantum computers: an engineering survey."

Q-CTRL Summit · Sydney · May 2026 · Tamilselvan R.

keynotearchitecture

2026 · INVITED

"Sparse-blossom MWPM on FPGA at d=15."

QIP 2026 · Cambridge MA · January 2026 · Park J.

inviteddecoder

2026 · CONTRIBUTED

"Error budgets across the photonic seam."

APS March Meeting · Las Vegas · March 2026 · Mehrabi N.

contributedQEC

2025 · INVITED

"Tantalum-on-silicon T₁ statistics on modular chiplets."

ICAP · Lisbon · September 2025 · Park J.

inviteddevices

2025 · CONTRIBUTED

"A reproducibility manifest for quantum benchmarks."

AQIS · Tokyo · August 2025 · Khoury A.

contributedmethodology

2025 · INVITED

"Lattice surgery across a photonic interconnect."

Q2B · Santa Clara · December 2025 · Tamilselvan R.

invitedarchitecture

PRESS · COVERAGE

Nature · "A new quantum architecture takes the modular bet." · Apr 2026 IEEE Spectrum · "Inside QONTOS-1." · May 2026 The Quantum Insider · "QEC research, in the open." · Mar 2026 The Register · "Open by default." · Jun 2026

RESEARCH ROADMAP · G1 → G5

Five scientific milestones, mapped to the QONTOS family arc.

The research roadmap rides the same gate-keyed arc as the hardware roadmap. Each gate is a scientific milestone, not just an engineering one: a publishable result, a public dataset, and a reproducible benchmark.

G1

Architecture freeze. Error budget published.

The seven-contributor budget and the protocol catalogues are released in the v5.4 whitepaper. Reference implementations of the decoder pipeline and factory estimates are open under qontos-research.

paper · v5.4 whitepaper

G2

First module acceptance. d=5 telemetry.

Stabiliser-cycle telemetry from a 500-qubit module establishes p_phys, recovers ε_L vs d for small code distances, and validates the d=5 decoder budget on real silicon.

paper · "Architecture-validation telemetry"

G3

Two modules. Bell-pair validation.

Cross-module Bell-pair tomography verifies raw fidelity, false-herald fraction, added-noise budget, and purification assumptions before any useful logical operation is claimed.

paper · "Two-module Bell-pair validation"

G4

First logical qubit. Above-threshold honesty retained.

The G4 acceptance criterion is a distance-5 logical-qubit validation experiment. ε_L < p_phys is not required until QONTOS-2.

paper · "First logical-qubit validation"

Research

Module fleet · scaled QEC architecture study.

Later module-fleet concepts remain research vision until decoder scaling, purification overhead, protected distillation, and facility cryogenics have credible blueprints.

paper · "Fault-tolerant utility"

Research programme

Open research. Open code. Open dataset.

QONTOS research is openable in the literal sense: every protocol has a published methodology; every plot has a reproducibility bundle; every open problem has a public acceptance criterion. The device exists to ground that research in measurements that survive replication. Visiting researchers can apply for time through the steering committee.

The programme is hosted at the Zhyra Quantum Research Institute in Abu Dhabi. Reference code lives under github.com/qontos/qontos-research. Steering-committee minutes, time-allocation decisions, and quarterly research reports are published on the site.

github.com/qontos-research research@qontos.xyz

Architecture validation first. Fault tolerance stays gated.