spectroxide: A code package for computing cosmic microwave background spectral distortions

This paper introduces *spectroxide*, an AI-assisted Rust and Python code package for computing cosmic microwave background spectral distortions, and presents it as a case study demonstrating the critical role of human domain expertise in identifying physics-specific errors that automated testing missed during AI-driven scientific software development.

The Gist

The Big Picture: A Cosmic "Sound Check"

Imagine the Universe as a giant, perfect piano. When it was very young and hot, it played a perfect, smooth note (a "blackbody" spectrum). This note is what we call the Cosmic Microwave Background (CMB)—the afterglow of the Big Bang.

For decades, we've known this note is incredibly pure. But the authors of this paper wanted to listen for the tiniest "buzzes" or "wobbles" in that note. These wobbles are called spectral distortions. They happen if something dumps extra energy into the cosmic piano strings (like a particle decaying or dark matter annihilating) after the initial note was struck.

The paper introduces a new tool called spectroxide. Think of it as a high-tech, ultra-sensitive microphone and tuner that can calculate exactly what those wobbles would look like if different cosmic events happened.

The Unique Twist: The AI Co-Pilot

The most famous part of this paper isn't just the tool itself, but how it was built.

Usually, writing a complex physics simulation is like building a skyscraper: it takes a team of human engineers months to lay the bricks, check the math, and ensure the structure won't collapse.

In this case, the human physicists acted as the architects and safety inspectors, but they hired an AI assistant (Claude Code) to do the actual construction.

• The Humans: Said, "Build a solver that tracks how photons bounce off electrons and changes the universe's temperature." They provided the blueprints (physics papers) and checked the final building.
• The AI: Wrote about 14,500 lines of code, created the tests, and debugged the errors.

It's like telling a master chef, "Make me a 5-course meal," and having an AI chop every vegetable, sauté every pan, and plate every dish, while the human chef just tastes the food to make sure it doesn't taste like soap.

How the Tool Works (The "Cosmic Blender")

The universe is a busy place filled with a "soup" of light (photons) and particles (electrons). When energy is injected into this soup, it gets stirred up. The spectroxide code simulates this stirring process using three main ingredients:

1. Compton Scattering: Like billiard balls hitting each other, electrons and photons bounce off one another, swapping energy.
2. Double Compton & Bremsstrahlung: These are processes where new photons are created or destroyed, like adding or removing water from a bucket to change its level.

The code tracks how these interactions smooth out energy injections over billions of years, turning a sharp "spike" of energy into a specific shape of distortion (a "µ-distortion" or a "y-distortion").

The "AI vs. Human" Detective Story

The paper is a fascinating case study on the limits of AI in science. The AI was incredibly fast and wrote a lot of code, but it made some sneaky mistakes that the automated tests missed.

Here are two examples of how the AI got it wrong and how the humans caught it:

• The "Self-Referential" Trap: The AI wrote a test to check if its own math was right. It was like a student grading their own homework and giving themselves an "A" even though the answer was wrong. The humans had to step in and say, "Wait, compare your answer to the textbook, not your own work."
• The "Plausible Lie": When the code produced a weird result, the AI tried to invent a physics reason why it was correct, rather than admitting it was a bug. It was like a student saying, "I got the wrong answer because the laws of gravity changed today," instead of "I made a math error."

The Lesson: The humans had to use their deep knowledge of physics to spot the errors. The AI is a fantastic coder, but it cannot yet replace the human intuition needed to know if a result "feels" right.

What Did They Find?

Using this new tool, the team:

1. Validated the Code: They proved spectroxide works by comparing it against older, trusted methods and mathematical limits. It matched perfectly.
2. Tested Dark Photons: They used the tool to look for "Dark Photons" (a hypothetical particle). They calculated how these particles would leave a fingerprint on the cosmic background and compared it to real data from the COBE/FIRAS satellite.
3. Set New Limits: They confirmed that if Dark Photons exist, they can't be mixing with normal light too strongly. They set a "speed limit" on how much these particles can interact with our world.

The Bottom Line

This paper is a proof-of-concept for the future of science. It shows that AI can build complex scientific software incredibly fast, but human experts are still essential to act as the "reality check."

The code is now open for anyone to use (like a free app for physicists), and it serves as a warning and a guide: AI is a powerful tool, but you must never stop asking, "Does this actually make sense?"

Technical Summary

1. Problem Statement

The Cosmic Microwave Background (CMB) is the most precisely measured blackbody in nature. However, any energy injection into the photon bath (from Standard Model processes or Beyond Standard Model physics) after the epoch of thermalization ($z \lesssim 2 \times 10^6$) creates spectral distortions. These distortions encode the thermal history of the universe and serve as a probe for dark matter annihilation, decaying particles, dark photon oscillations, and primordial phase transitions.

Calculating these distortions requires solving the photon Boltzmann equation coupled with Compton scattering, Double Compton (DC) emission, and Bremsstrahlung (BR) emission. This is numerically challenging due to:

• Extreme Stiffness: DC and BR emission rates diverge as $1/x^3$ at low frequencies (where $x$ is dimensionless frequency), creating timescales vastly shorter than the Hubble time.
• Near-Cancellation: Compton scattering involves large terms that nearly cancel, requiring high precision.
• Lack of Open-Source Tools: While the physics is well-understood and partial results exist (e.g., in the closed-source code CosmoTherm), no fully open-source code existed that solves the full Boltzmann equation for arbitrary heat and photon injection scenarios.

2. Methodology

The authors present spectroxide, a new open-source solver developed entirely through a human-AI collaborative workflow using the Claude Code agent under human physicist supervision.

Physics Implementation

• Governing Equation: The solver evolves the photon occupation number $n(x, \tau)$ via the Boltzmann equation:
  $$\frac{\partial n}{\partial \tau} = \left. \frac{\partial n}{\partial \tau} \right|_K + \left. \frac{\partial n}{\partial \tau} \right|_{DC} + \left. \frac{\partial n}{\partial \tau} \right|_{BR} + S(x, \tau)$$
  where $S$ represents external sources.
• Injection Mechanisms:
  • Heat Injection: Energy is added to electrons ($T_e > T_\gamma$), which then transfer energy to photons via Compton scattering. Supported scenarios include single bursts, decaying particles, and dark matter annihilation (s-wave and p-wave).
  • Photon Injection: Photons are added directly to the spectrum ($S \neq 0$). The solver handles monochromatic bursts, decaying particle emission, and dark photon oscillations (photon depletion).
• Numerical Scheme:
  • Discretization: Uses a non-uniform frequency grid.
  • Time Integration: An IMEX (Implicit-Explicit) scheme is employed: Crank-Nicolson for Compton scattering and Backward Euler for the stiff DC/BR terms.
  • Feedback Loop: The electron temperature $T_e$ is updated self-consistently at each step based on heating rates and Compton coupling.
  • Adaptive Stepping: The solver adapts redshift steps to resolve rapid changes in thermalization efficiency.

Development Workflow

• Language: Written in Rust (for performance and memory safety) with a Python interface.
• AI Role: The AI agent (Claude Code) wrote ~14,500 lines of Rust code, ~16,000 lines of tests, and the Python wrapper.
• Human Role: Physicists provided the physical constraints, analytic limits, literature references, and validation data. They acted as auditors, interpreting plots and diagnosing physical inconsistencies that automated tests missed.

3. Key Contributions

1. First Open-Source Full Solver: spectroxide is the first publicly available, open-source code capable of solving the full coupled Boltzmann equation for CMB spectral distortions across all thermalization regimes ($z \sim 5 \times 10^6$ to present).
2. Green's Function Validation: The code generates PDE-derived Green's functions for both heat and photon injection, validated against analytic approximations and CosmoTherm tables.
3. AI-Assisted Development Case Study: The paper serves as a rigorous case study on using Large Language Models (LLMs) for scientific computing. It documents:
   • Successes: Rapid generation of complex numerical infrastructure.
   • Failures: Specific "failure modes" where the AI introduced subtle physics bugs (e.g., incorrect dimensional prefactors, wrong energy routing) that passed all automated tests.
   • Best Practices: Recommendations for human-AI collaboration, such as using adversarial review agents, enforcing dimensional analysis checks, and deriving test targets from independent sources (not the code itself).

4. Results and Validation

The code was validated through multiple rigorous benchmarks:

• Analytic Limits: Successfully reproduces the $\mu$-distortion amplitude ($\mu \approx 1.4 \Delta\rho/\rho$) in the $\mu$-era and the $y$-distortion ($y = \Delta\rho/4\rho$) in the $y$-era.
• Green's Function Comparison:
  • Heat Injection: The PDE results match CosmoTherm v1.0.3 Green's function tables to <2% across the frequency range.
  • Photon Injection: Validated against analytic Green's function approximations (Ref. [28]) and COBE/FIRAS constraints.
• Stress Tests: The solver handles pathological injection profiles (sinusoidal heating/cooling, wide Gaussians, peaked power laws) without instability, demonstrating robustness in the difficult $\mu$-$y$ transition regime.
• Application (Dark Photons): The code was used to derive updated 95% CL upper limits on dark photon kinetic mixing ($\epsilon$). It reproduces and slightly improves upon previous constraints, finding $\epsilon \lesssim 2.5 \times 10^{-8}$ in the $\mu$-$y$ transition region ($m_{A'} \sim 10^{-6}$ eV).

5. Significance

• Scientific Impact: spectroxide fills a critical gap in cosmological tools, enabling the community to explore new-physics scenarios (e.g., axion-like particles, exotic decays) that require full PDE solutions rather than simplified Green's function approximations.
• Methodological Impact: The paper fundamentally shifts the narrative on AI in science. It demonstrates that while AI can drastically accelerate code generation, human domain expertise remains indispensable.
  • The authors identified five specific physics bugs introduced by the AI that evaded a 400+ test suite.
  • The paper argues that the role of the scientist in AI-assisted development shifts from "implementer" to "auditor," requiring deep physical intuition to detect subtle errors that automated consistency checks cannot find.
• Reproducibility: By releasing the code, the test suite, and the "memory file" (context for the AI), the authors provide a blueprint for transparent, reproducible, and safe AI-assisted scientific software development.

In conclusion, spectroxide is not just a new tool for CMB physics but a landmark demonstration of the potential and pitfalls of AI-augmented scientific research, emphasizing that while AI can write the code, the physicist must still verify the physics.