Valiant Quantum — NISQ Reality: Drug Candidate Selection Today

⬡ R&D Submission Batch — identical to Demo A

Step 1 — Candidate Batch

Same Problem, Same Data

The 9 qualified mRNA candidates are identical to Demo A. The QUBO, the objective, the constraints — all the same. The only thing that changes is what happens when this circuit meets real quantum hardware today.

⚠ What's different in this demo

Demo A shows the intended future: a clean quantum selection of the optimal 3 candidates. This demo shows what actually happened when we submitted the same circuit to IQM Garnet on 2026-02-28. The results are real, unedited, and honest.

▦

9 Qualified Candidates (post-vault)

BNT-V005, V008, V012 filtered by RNAfold MFE score < 0.40

Qubit	Candidate	Antigen	Immu.	Stab.	Cost

⬡ Vault + QUBO — same as Demo A

Steps 2–3 — Vault & QUBO

IP Protection & Problem Encoding

These steps are identical to Demo A. The Vault anonymizes proprietary sequence data and enforces the RNAfold MFE filter. The QUBO encodes the selection objective. The output is a 9×9 matrix ready for circuit compilation.

🔒

Vault Output

All identical to Demo A

Candidates cleared9 / 12

Filtered (RNAfold MFE < 0.40)3

IP anonymization✓ complete

GxP audit trail✓ HMAC-signed

QUBO Parameters

Same objective, same matrix

Matrix size9 × 9

Immunogenicity weight (α)0.60

Cost penalty (β)0.25

Phase I slots (k)3

Penalty λ0.8202

⚠ Where the Problem Begins

Step 4 — Circuit Transpilation

The SWAP Overhead Problem

QAOA requires a two-qubit gate between every pair of candidates (36 RZZ gates for 9 candidates). IQM Garnet’s crystal-20 topology only has edges between physically adjacent qubits. Non-adjacent pairs need SWAP chains — and SWAPs cost gates, and gates cost fidelity.

LOGICAL CIRCUIT

DEPTH

9 qubits · 36 RZZ · 9 RZ · 9 RX

AFTER TRANSPILATION (IQM Garnet)

155

DEPTH

115 two-qubit gates · 8.2× depth increase

⚠ Why this matters

Each two-qubit gate on IQM Garnet has ~0.5% error rate. At 115 two-qubit gates, the probability that at least one gate fails is approximately 1 − 0.995¹¹⁵ ≈ 44%. The quantum state is corrupted on nearly half of all shots before measurement.

This isn’t a bug in Garm or in the QAOA algorithm. It’s the current state of NISQ hardware. Fault-tolerant quantum computing (post-2030 roadmap) will reduce gate error rates by 3–5 orders of magnitude through quantum error correction — at which point circuit depth becomes tractable.

⇄

Sirius vs. Garnet — Topology Comparison

Both runs submitted 2026-02-28 via IQM Resonance

Device	Topology	Qubits	Transpiled Depth	2Q Gates	Valid Shots	Job ID
IQM Sirius	star-16	16	183	144	23.6% (242/1024)	019ca35e…
IQM Garnet PRIMARY	crystal-20	20	155	115	20.4% (209/1024)	019ca361…

Note: Garnet’s crystal topology gave fewer 2Q gates (115 vs 144) but similar noise floor — both circuits are deep enough that error rates dominate.

⬡ Aleta Pricing — same formula, real job

Step 5 — Pricing & Contract

Transparent Cost — Noisy or Not

Garm’s Aleta pricing is the same whether the hardware result is clean or noisy. The enterprise pays for quantum time consumed, not for result quality — exactly as with any infrastructure service. The SLA covers delivery, not optimality.

€

Aleta Pricing (IQM Garnet)

Computed from actual job parameters

Qubits (n)9

Circuit depth (d)155 (transpiled)

Shots1,024

Backend multiplier1.15 (IQM Garnet)

Runtime factor (τ)—

Base cost—

Platform fee (22%)—

Vault surcharge€12.00

Total—

The Pricing Honesty

Why Aleta prices noise-affected jobs the same

Quantum time is consumed regardless of result quality. IQM Garnet ran 1,024 shots; the QPU operated correctly. The noise comes from physics, not from the platform.

This is identical to classical HPC: you pay for CPU-hours even when a simulation doesn’t converge. The difference is Garm makes the noise visible and auditable — the result package includes circuit depth, gate count, and expected error rate so the enterprise can assess result confidence.

As hardware improves, the same price buys a dramatically more reliable result. The pricing model is hardware-agnostic by design.

⬡ IQM Garnet — Resonance — REAL JOB

Step 6 — Hardware Execution

Executed on IQM Garnet — 2026-02-28

This is not a simulation. The circuit was submitted to IQM Resonance and executed on a real 20-qubit superconducting QPU in Finland. Results are taken directly from the API response.

✓

Job Record

IQM Resonance API response

{
  "job_id":      "019ca361-6339-7e32-a69f-bd10c611058b",
  "backend":     "IQMBackend (garnet)",
  "qubits":      20,
  "shots":       1024,
  "status":      "COMPLETED",
  "circuit_depth_logical":     19,
  "circuit_depth_transpiled":  155,       // 8.2× increase — SWAP overhead
  "two_qubit_gates_logical":   36,
  "two_qubit_gates_transpiled": 115,      // 3.2× increase
  "valid_selections_k3":       209,       // shots with exactly 3 bits set (20.4%)
  "distinct_valid_bitstrings":  75,        // vs C(9,3)=84 possible — near-uniform
  "best_bitstring":            "001110000",
  "best_bitstring_count":      10,        // 1.0% — just above noise floor
  "endpoint":    "https://resonance.meetiqm.com/",
  "date":        "2026-02-28"
}

⟐

Execution Pipeline

What Garm + Arvak did

QUBO → QAOA circuit (Qiskit, 9 qubits, depth 19)

IQMProvider connected to resonance.meetiqm.com/garnet

Transpile: optimization_level=3, depth 19→155, gates 36→115

POST /jobs — submitted to IQM Resonance queue

1,024 shots executed on IQM Garnet 20-qubit QPU

Results retrieved: 446 unique bitstrings observed

∼

Expected vs. Observed

Ideal QAOA vs. NISQ reality

Valid selections (k=3) expected>50%

Valid selections observed20.4%

Top bitstring prob. expected>10%

Top bitstring prob. observed1.0%

Distinct bitstrings expectedfew dominant

Distinct bitstrings observed446 of 512

⬡ Real Measurement Data — IQM Garnet — 1024 shots

Step 7 — Results

What the QPU Returned

446 distinct bitstrings observed across 1,024 shots. A uniform distribution over all 512 nine-bit states would give 2 shots each. The observed distribution barely departs from uniform — the signal is buried in noise.

446

Distinct Bitstrings

of 512 possible (87%)

20.4%

Valid Selections

209 shots with exactly k=3 bits

1.0%

Best Outcome

10 shots — barely above noise

▨

Top 15 Measurement Outcomes

Real counts from Job 019ca361 — ✓ marks valid k=3 selections

Uniform noise baseline: ~2 shots per bitstring. Max observed: 13 shots (1.3%). Compare to Demo A simulation: top outcome ~14%.

Quantum Selection (best valid bitstring)

|001110000⟩ — 10 shots (1.0%)

Rank	Candidate	Antigen	Immu.
#1	BNT-V006	PIK3CA-E545K	0.83
#2	BNT-V007	NRAS-Q61K	0.70
#3	BNT-V009	CDKN2A-R58X	0.79

✓

Classical Optimum (brute force)

Best of all C(9,3) = 84 combinations

Rank	Candidate	Antigen	Immu.
#1	BNT-V001	KRAS-G12D	0.88
#2	BNT-V004	EGFR-L858R	0.91
#3	BNT-V011	ALK-F1174L	0.92

The one signal in the noise

The quantum and classical selections don’t match — expected at this noise level. But both selections independently include high-immunogenicity candidates (quantum avg immu: 0.77, classical avg immu: 0.90). The hardware is finding something — the noise hasn’t completely erased the objective signal. This is consistent with 1-layer QAOA on NISQ: partial correlation with the optimum, not convergence to it.

⬡ Garm Routing Intelligence & Hardware Roadmap

Step 8 — What’s Next

From NISQ to Advantage

The honest picture: gate-based QAOA on 9 qubits is noise-dominated today. Garm knows this — and routes accordingly. Here is what the platform does with this information, and what the hardware roadmap looks like.

⬥ GARM ROUTING RECOMMENDATION FOR THIS JOB

↳ Route to D-Wave, not IQM Garnet

For binary candidate selection (QUBO, k=3 from n=9), D-Wave Advantage annealing is the correct hardware today. D-Wave maps the QUBO directly to physical qubits — no circuit compilation, no SWAP overhead, no gate depth. At 9 variables the circuit-depth problem disappears entirely.

Garm’s AI classifier (Nathan ONNX satellite model) scores this problem as OPTIMIZATION / ANNEALING-preferred when the problem is a pure QUBO with no parametric structure. The IQM routing in this demo was intentional — to show the NISQ reality for a live investor meeting. In production, Garm would have routed to D-Wave automatically.

TODAY — NISQ

D-Wave Advantage: QUBO → native annealing, no circuit depth, near-optimal selections on problems up to 5,000 variables

IQM / IBM (gate): Small circuits (≤5 qubits, depth ≤20) — Bell states, small VQE. Not yet useful for combinatorial optimization at this scale.

2026–2028 — EARLY FAULT-TOLERANT

Error-corrected logical qubits emerge. Circuit depth becomes tractable. 20–50 logical qubits.

QAOA at 9–20 candidates becomes reliable. Molecular simulation (VQE for drug-target binding) becomes feasible on small molecules.

2030+ — QUANTUM ADVANTAGE

Hundreds of logical qubits. QAOA at 500+ candidates. Full molecular simulation of drug-target interactions.

Demo A becomes real: the quantum selection converges to the classical optimum, and eventually surpasses it on problems classical computers cannot solve.

⬥ WHY GARM’S VALUE INCREASES AS HARDWARE IMPROVES

Every hardware generation — from today’s NISQ to fault-tolerant — makes the routing problem more valuable to solve, not less. As more backends become viable for more problem types, the combinatorial space of “which hardware for which problem” expands.

Garm is not betting on one hardware winner. It is betting on the infrastructure layer that sits above all of them — the enterprise API, the pricing engine, the IP vault, the SLA framework — all of which are hardware-agnostic and needed regardless of which QPU technology prevails.

The company that funded BioNTech before mRNA was proven understood: platform bets pay when the underlying technology matures. The quantum infrastructure layer is that bet.