When the Physics Says No: A Fidelity-Threshold Story
We spent real Braket credits and got a REJECT.
We're publishing it because it's the single most important thing you can know about our verifier: it discriminates on physics, not on money.
Here's what happened.
The stronger claim: Svetlichny(3)
The Mermin(3) test we walked through in our first post is a Bell inequality that violates for both biseparable and genuinely tripartite entangled states. It proves quantum-vs-classical separation, but it doesn't distinguish "the three qubits are entangled pairwise" from "the three qubits are entangled as a genuine trio."
Svetlichny's inequality does. It's an 8-term multipartite Bell inequality with classical bound 4 and quantum ceiling 4√2 ≈ 5.657, and crucially, only genuinely tripartite entangled states can violate it. Passing Svetlichny is the strongest single-experiment evidence you can produce that your three qubits are entangled as a unified quantum object, not as two pairs plus a spectator.
We wanted that in the anchor set. It's the paper-worthy witness on our stack.
Attempt one
Same code path as before, same block-binding, same anchor format. New witness spec: 8 terms with the Svetlichny coefficient signs, tilted measurement angles chosen (via numerical optimization) to saturate the 4√2 quantum ceiling on an ideal GHZ state.
1024 shots on Rigetti Cepheus-1-108Q, 8 tasks (one per group), about $1.30 in Braket credits.
Result: |Sv| = 3.310. REJECT. Below the classical bound of 4 by 0.69.
Attempt two, thirty minutes later
Real QPUs drift. Calibration state changes across queue slots, and 5 to 15% over a day is normal. So we ran it again, 2048 shots this time to tighten the statistical noise, same challenge-binding, same code, submitted about 30 minutes after attempt one.
Result: |Sv| = 3.469. REJECT. Closer to 4 but still below. Cepheus's fidelity had drifted from about 58.5% to 61.3% between the two runs.
Both honestly REJECTED. Both cryptographically committed. Both live on qblex.com/verify right now, on the same buttons as the passing runs, just carrying a rose-colored "REJECTED" badge instead of an emerald one.
The physics
Bell-family Tsirelson bounds all follow the same ratio: quantum ceiling = √2 × classical bound. This means the hardware fidelity threshold to violate any Bell inequality is exactly:
F_threshold = classical_bound / quantum_ceiling = 1/√2 ≈ 70.7%
Below that fidelity, no amount of shots will move the mean above the bound. Statistics tighten error bars; they don't relocate central values. Only better hardware does.
Cepheus at the tilted-Svetlichny circuit lands at 58 to 61% fidelity. That's about 10 percentage points below the threshold. The REJECTs aren't a bug; they're what the physics predicts to the third decimal place: expected |Sv| ≈ 4√2 × 0.60 = 3.39, measured |Sv| = 3.31 and 3.47. Prediction and measurement match.
For comparison, we ran CHSH on the same Cepheus hardware and it passed at 2.119 (our last post). Cepheus's 2-qubit fidelity for the shorter CHSH circuit is about 75%, above the same 70.7% threshold. Different circuit, different fidelity, different verdict, and both verdicts arrive from the same verifier code.
Why we're publishing the failures
Because a verifier that only says yes proves nothing.
Every quantum computing announcement in the last five years has been a highlight reel. Curated best-of results, published PASSes only, silent failures. The reader has no way to know whether a claimed violation is a real physics signal or a cherry-picked lucky draw. The reader has no evidence the verifier can say no.
Ours can. It just did. Twice. And the mathematical structure of why it said no is transparent: fidelity below 70.7%, therefore Bell violation impossible, therefore REJECT. You can rederive that number from Tsirelson's bound on the back of a napkin.
Publishing the REJECTs is what makes the PASSes credible. The verify page has nine anchor buttons right now. Five PASS. Four REJECT. Click any of them, the code path is identical, the verdict is a function of the physics.
What's next
The finale is coming. Same code. Same operator. Same Svetlichny witness. Different QPU.
You already know what direction that's going. What you don't know yet is how big the number is.
Don't trust us. Run the math.
QBLEX technologies are patent pending (U.S. App. Nos. 19/711,486; 19/702,980; 19/707,649 — priority June 2025). "Patent pending" does not indicate a granted patent.