import waveml_jax
from waveml_jax.core.representation import WaveState



Error Rectification by Alignment (ERA) — Reference Implementation

Lightborne Intelligence

Truth > Consensus. Sovereignty > Control. Coherence > Speed.


Overview

WaveML is a reference implementation of Error Rectification by Alignment (ERA), a governance mechanism for wave-native learning systems.

ERA enforces physically meaningful invariants on internal model states — amplitude, energy, and phase — to ensure bounded trajectories and graceful degradation under noise, drift, and long-horizon propagation.

This repository provides a canonical, minimal, and correct implementation intended for research, evaluation, and extension.


Core Idea

Conventional machine learning systems attempt to achieve robustness indirectly through preprocessing, regularization, or loss shaping.

ERA takes a different approach:

Robustness is governed, not optimized.

ERA constrains how internal states may evolve, independently of task objectives, by enforcing invariant-preserving rectification at every computational step.


What This Repository Is

✅ A reference implementation of ERA
✅ Wave-native state representations (amplitude + phase)
✅ Deterministic, reproducible JAX code
✅ Synthetic and academic benchmarks
✅ Baseline comparisons for stability and robustness


What This Repository Is NOT

❌ A production-ready system
❌ A domain-calibrated solution
❌ A collection of tuned hyperparameters
❌ A deployment playbook

Domain-specific calibration, operating envelopes, and production integrations are intentionally out of scope.


Repository Structure
"""
WaveML/

├── waveml_jax/
│   ├── core/
│   │   ├── representation.py
│   │   ├── invariants.py
│   │   └── era_rectify.py
│   │
│   ├── models/
│   │   ├── waveseq.py
│   │   ├── wave_rf.py
│   │   └── baselines.py
│   │
│   └── benchmarks/
│       ├── delayed_copy.py
│       ├── noise_sweep.py
│       └── snr_sweep.py
│
├── README.md
├── pyproject.toml
└── LICENSE
"""


Installation

CPU (default)

pip install git+https://github.com/lightborneintelligence/WaveML.git@v1.0.0

CUDA (NVIDIA GPU)

Install JAX with CUDA support first, then WaveML:

pip install --upgrade "jax[cuda12]"
pip install git+https://github.com/lightborneintelligence/WaveML.git@v1.0.0

Ensure the CUDA version matches your system.
See the official JAX installation guide for supported configurations.


Quick Sanity Check

python - <<'PY'
import jax.numpy as jnp
from waveml_jax.core.representation import WaveState
from waveml_jax.core.invariants import InvariantBounds
from waveml_jax.core.era_rectify import era_rectify

state = WaveState(
    amplitude=jnp.array([0.5, 2.0]),
    phase=jnp.array([0.0, 10.0])
)
bounds = InvariantBounds(max_amplitude=1.0, max_energy=1.0)
print(era_rectify(state, bounds))
PY

This verifies invariant enforcement (energy bounding + phase wrapping).


Error Rectification by Alignment (ERA)

ERA enforces the following invariants at every step:
	1.	Amplitude non-negativity
	2.	Amplitude bounds (element-wise)
	3.	Total energy bounds (global)
	4.	Phase wrapping to ([-π, π])
	5.	Phase gating where amplitude vanishes

Rather than suppressing or filtering states, ERA realigns them to the nearest admissible configuration while preserving information content.


Design Principles
	•	Wave-native — Phase preserved end-to-end
	•	Governed dynamics — Invalid states cannot arise
	•	Differentiable — Fully compatible with gradient-based learning
	•	Architecture-agnostic — Applicable to recurrent, convolutional, and hybrid systems
	•	Reproducible — Fixed seeds and documented configurations (noting device-specific JAX/XLA nondeterminism)


Intended Use

WaveML is intended for:
	•	Research into robustness and stability
	•	Evaluation of wave-native representations
	•	Benchmarking against classical baselines
	•	Educational and exploratory use

It is not intended as a drop-in production system.


Publications

This repository accompanies the technical report:

Error Rectification by Alignment (ERA): A Governance Primitive for Wave-Native Learning Systems
Jesus Carrasco, Lightborne Intelligence (2026)
Zenodo: https://doi.org/10.5281/zenodo.18263860

Application-focused reports (Part II: sensing, RF, long-horizon sequence modeling) are forthcoming.


License
	•	Code: Apache License 2.0
	•	Documentation: CC-BY-4.0

See LICENSE for details.


Citation

If you use this work, please cite:

Carrasco, J. (2026).
Error Rectification by Alignment (ERA): A Governance Primitive for Wave-Native Learning Systems.
Zenodo. https://doi.org/10.5281/zenodo.18263860


#Disclaimer

This is a reference implementation provided for research and evaluation purposes.
No guarantees are made regarding fitness for production or safety-critical use.


© 2026 Lightborne Intelligence