New technical note AGE-PN-001: reproducible benchmark for WDM stability See evidence
Method

The bank is one system. Not a collection of isolated channels.

AGE explicitly represents shared dynamics, local perturbations and serial observation. The gain comes from structure, not from applying a more elaborate filter to each channel in isolation.

State model

Common mode + local modes + incomplete observation.

For an N-channel bank, detuning is decomposed into a shared contribution and one device-specific contribution per channel. Parameters are not assumed universal: they are identified or bounded for the platform under study.

i

The model is a testable hypothesis. If residuals show the common/local structure is insufficient, it must be extended or rejected.

state_model.pyconceptual
state  xₖ = [ cₖ, ℓ₁ₖ, …, ℓₙₖ ]ᵀ

dynamics
  cₖ₊₁  = a_c cₖ  + w_c
  ℓᵢₖ₊₁ = aᵢ ℓᵢₖ + wᵢ

detuning
  Δλᵢₖ = cₖ + ℓᵢₖ

measurement
  yₖ = H(channelₖ) xₖ + vₖ

# every probe updates the common state
Cross-observability

The probed channel changes; the common mode remains observable.

Under round-robin probing, each channel may go many periods without its own measurement. Yet any measurement can update the shared component of every channel.

t₀CH 01

Probe

The monitor observes one channel and returns detuning or an equivalent signal.

t₁Status

Separate

The innovation is allocated between common and local components under the model.

t₂all

Update

Every channel posterior changes through the shared common state.

t₃CH 02

Predict

The state is propagated to the next probe with explicit uncertainty.

Validation design

Identify and score on different time periods.

Improvement in the same window used to fit parameters is not evidence of predictive performance. The recommended route therefore separates identification and evaluation chronologically.

  • Schema and units fixed before fitting
  • Closed identification window
  • Prediction produced for held-out period
  • Primary metrics and failure rules set in advance
Timelineno shuffling
Block A · identification0–60 %
Gap / guard bandoptional
Block B · held-out evaluation60–100 %
Integration architecture

AGE sits between telemetry and actuators.

The evaluation uses exported data. A later integration adds monitor and control adapters without assuming every platform exposes the same interfaces.

01

Telemetry

Detuning, wavelength, power, temperature, channel state and timestamps.

02

Identification

Effective parameters, covariance structure, noise and residual tests.

03

Estimation

Joint state, uncertainty and between-observation prediction.

04

Actuation

Corrections constrained by power, latency, range and system safety.

!

The actuation layer is platform-specific and requires experimental validation. The current open package demonstrates offline estimation in simulation.

Inspect the method

Evaluate the methodology before sharing data.

Download the technical note and reproducibility package. Both distinguish the nominal scenario from future measured validation.