New technical note AGE-PN-001: reproducible benchmark for WDM stability See evidence
TM-05 · Lab simulator

Two-node chiral cascade

Node A emits a temporally shaped photon, the chiral waveguide carries it with loss, and node B absorbs it. Everything is integrated in one evolution, with the published parameters of ETH Zurich's 30 m link.

— chiral · Markovian · 1 excitation

Populations of the four modes

|e_A|² emitter storage, |b_A|² emitter cavity, |b_B|² receiver cavity, |e_B|² receiver storage.

Emitted photon vs target mode

|b_out|² = κ_A|b_A|² vs |f_E|². If they overlap, the emission control is exact.

Coupling controls

g_A(t) and g_B(t) in units of Γ. Outside the body (tail mask) they are zero.

Realizability: radicand ρ(1−F) − |f|²

Must be positive where the control acts. Below zero the control does not exist and is set to zero: this is the manuscript's condition.

Matrix entry M[i][j] is the transfer probability when A emits mode i and B runs the absorption control for mode j. Selectivity Σ is the mean diagonal divided by the mean off-diagonal. It uses the simulator's current parameters.

Uncompensated (δω_c = 0)

Compensated (δω_c = δω_B − δω_A)

How much does Σ depend on the noise realisation?

Σ ≈ 3578 is a single realisation of the shaping noise (seed 2026). Here the compensated matrix is repeated over many seeds, with the same noise model, to show the full distribution. The Python reference is 40 seeds computed with the original code.

Distribution of log₁₀ Σ (compensated, 5% noise)

About 110 full integrations; takes a few seconds.

Transfer vs relative detuning

Opposite detunings ±δ/2, mode 0, no loss. Dots: benchmark 09 Python reference.

Error in the δω_A estimate

At ±1.75 MHz the receiver compensates with an estimated value. Benchmark 09 feeds the exact value; in operation benchmark 07 would supply it.

The tail sets the 0.9996 figure

Loss 1 − P vs truncated tail probability, per mode. Unit slope means the limit is the truncation, not the physics.

Receiver width κ_B and realizability

Loss 1 − P and the fraction of the grid where the receiver control does not exist, mode 0. Below ρ_B = 1 the control vanishes in the tail, but the loss grows slowly.

Guardians

The six benchmark-09 guardians with the same thresholds, plus two tests added by this bench.

Cross-validation against the Python code

This browser recomputes every number in the benchmark-09 JSON and compares it with the Python values (LSODA, rtol 10⁻⁹). The seed-2026 noise is the exact NumPy realisation, embedded.

Auditor's findings

Eight experiments. In each one, choose your prediction first. Then the bench runs the simulation and shows you the measured result and why it happens.

01 · Common frame

You detune both nodes by +1.75 MHz at once, without compensation. What happens to P?

Your prediction:
02 · Relative detuning

Now the emitter sits at +1.75 MHz and the receiver at −1.75 MHz (3.5 MHz relative). What is P?

Your prediction:
03 · Compensation sign

You compensate with δω_c = δω_A − δω_B, the wrong sign. What happens?

Your prediction:
04 · Link loss

With the matched link you add a loss p = 0.17. What is P?

Your prediction:
05 · Narrow receiver

You lower κ_B/2π to 22 MHz, so ρ_B = 0.92 < 1, the minimum for mode 0. What happens?

Your prediction:
06 · The 0.9996 figure

You change the truncated tail probability from 10⁻⁴ to 10⁻³. What happens to matched transfer?

Your prediction:
07 · Error estimating δω_A

At ±1.75 MHz you compensate with a δω_A estimate that is off by 0.5 MHz. What is P?

Your prediction:
08 · Is Σ ≈ 3578 stable?

With 5% shaping noise and compensation, how much does Σ vary across noise realisations?

Your prediction:

Integrated equations

ė_A = −i g_A b_A − i δ_A e_A ḃ_A = −i g_A* e_A − (ρ_A/2) b_A − i δ_A b_A ḃ_B = −i g_B* e_B − (ρ_B/2) b_B − i (δ_B − δ_c) b_B + √((1 − p) ρ_A ρ_B) b_A ė_B = −i g_B b_B − i (δ_B − δ_c) e_B e_A(−∞) = 1 ; P = |e_B(+∞)|² b_out = √ρ_A b_A

Dimensionless time τ = Γt, ρ = κ/Γ. Detunings act on both amplitudes of each node, so a common detuning is an exact frame change.

Modes and controls

f₀ ∝ sech(τ/2) f₁ ∝ τ sech(τ/2) f₂ ∝ (τ² − π²/3) sech(τ/2) emission: g(τ) = (f' + ρf/2) / √(ρ(1−F) − f²) absorption: g_abs(τ) = g_em[f(−τ)](−τ) F(τ) = ∫ |f|² ; body: tail ≤ F ≤ 1 − tail realizable ⇔ ρ(1−F) − f² > 0

The mode family is sech(τ/2) times a polynomial in τ, exactly orthogonal. It was reconstructed and verified against the benchmark-09 JSON with differences ≤ 3.3×10⁻⁸.

Parameters and their origin

SymbolValueMeaningOrigin
Γ/2π 24 MHz Transfer rate of the experiment ETH · arXiv:2604.12947
κ_A/2π 26.7 MHz Emitter node width ETH · arXiv:2604.12947
κ_B/2π 30.7 MHz Receiver node width ETH · arXiv:2604.12947
p 0.17 Link loss in the composite scenario AGE · benchmark 09
δω_A, δω_B ±1.75 MHz Opposite detunings of the composite scenario AGE · benchmark 09
noise 5% · τ_c = 5 ns Multiplicative noise on the emitter control AGE · composite_channel
tail_prob 10⁻⁴ Control tail truncation AGE · numerical
grid τ ∈ [−40, 40] · N = 2501 Time grid; RK4 with 4 substeps numerical

Outside the model (open)

  • Bidirectional back-reflection and imperfect circulators: these need the SLH feedback-reduction rule.
  • Propagation delay: the Markov approximation is applied; the 30 m link delay is folded into time alignment.
  • Multi-photon sectors and quantum jumps: only no-jump single-excitation amplitudes; loss is amplitude attenuation.
  • Compensation is fed the true δω_A. In operation benchmark 07 would have to estimate it.
  • Noise affects only the emitter control; the receiver control is ideal.
  • It validates single-excitation transfer between two nodes. It does not validate any multi-qubit computing architecture.

Export and seal this run

JSON with the current parameters and results, and its SHA-256 fingerprint. Use it as a pre-registration: publish the fingerprint before comparing with a measurement.