I build testable ideas across cosmology and machine learning—from new mechanisms for cosmic acceleration to evidence systems that make model claims harder to fool.
I’m a student researcher working in machine learning and theoretical physics, focused on ideas that can be turned into evidence.
My current work includes Configurational Dispersal Cosmology (CDC), developing configurational dispersal as a possible mechanism behind cosmic acceleration, and Marginal Baseline Evaluation (MBE), ML research on generalization metrics, meaningful controls, and more reliable ways to evaluate model behavior.
01 About
I am Aparajeet Shadangi, a student researcher working across cosmology and machine learning. I turn questions into equations, software, controlled experiments, and public evidence—and document failure boundaries as carefully as successes.
CDC tests a geometric route to cosmic acceleration. MBE asks whether machine-learning metrics survive ordinary baselines. Through Phason Labs ↗, I turn this method into open tools and independent evidence audits.
The original hypothesis and mathematical framework now advance through a non-fluid Geo K1 implementation and observational tests.
The machine-learning work is released as inspectable packages, protocols, and reproducibility code through Phason Labs.
Versioned PEA cases that reconstruct machine-learning claims, preserve failure modes, and publish the evidence.
02 Research & Writing
A developing framework that asks whether cosmic acceleration can emerge from configurational dispersal. The work moves from mathematical formulation to numerical consistency and observational tests.
Research code & current direction ↗ Cosmology / evidenceA reproducible baseline analysis of the S8 and growth-rate question, documenting what a curated fσ8 compilation does—and does not—support under transparent assumptions.
Read the preprint ↗ Machine learning / research directionMBE tests whether a proposed metric contributes information beyond ordinary baselines. The current benchmark asks a related practical question: which training signals remain trustworthy for a specific task and intervention?
Research repository ↗ Writing / wider inquiryLong-form writing that connects the formal research program to broader questions about explanation, uncertainty, and how ideas become knowledge.
Read the essays ↗03 Phason Labs
Phason Labs ↗ is the founder-led AI research lab growing out of my work on model internals. Its first commercial wedge is Phason Evidence Audit: independent claim testing for ML teams and maintainers, backed by public protocols, seven completed audits, and a merged upstream fix.
Why “phason”? In a quasicrystal, a phason is a coordinated internal rearrangement. It is a useful analogy for the question behind the lab: what changes inside an intelligent system when its behaviour changes?
Tests whether a machine-learning metric contributes information beyond ordinary baselines.
Repository ↗ P–02 / 7 public auditsPhason Evidence AuditsVersioned reconstructions and stress tests of machine-learning claims, including failure boundaries and executable evidence.
View the audit collection ↗ P–03 / Open sourceTrainToolsPyTorch diagnostics for gradient noise, model plasticity, and early-stopping decisions.
Repository ↗ Upstream / MergedLLMs from ScratchPEA-006 identified cached-generation failures and delivered 11 focused regression tests with the accepted fix.
View merged PR #1048 ↗The current CDC implementation is focused on the non-fluid geometric direction, including the Geo K1 formulation. Rather than treating CDC only as an effective fluid or parameterized background model, this work develops the geometric structure needed for consistency checks and eventual observational comparison.
CDC II develops the original hypothesis into a self-contained mathematical theory: an effective spacetime action, binding-dependent kinetic suppression, background evolution equations, a dimensionless e-fold formulation, and perturbation-level structure suitable for Boltzmann integration. The next development step is to move from reference checks into full observational inference.
04 Timeline
05 Approach
The standard approach in cosmology is to work within accepted frameworks until evidence demands otherwise. That is the right default. But the persistent uncertainties around dark energy — its unknown physical nature, the cosmological constant problem, and the growing statistical tensions in observational data — create genuine scientific room for alternative models, provided they are rigorously developed and falsifiable.
CDC is built to that standard. The goal is not to claim it is correct, but to develop it to the point where observational data can answer that question. A framework that cannot make falsifiable predictions is not science, regardless of its internal structure. CDC makes concrete predictions for H(z), CMB distance observables, growth, and the matter power spectrum that can be compared directly against observation.
The same philosophy guides my software projects. A training diagnostic should surface a real failure mode while a run is still alive. A scientific ML model should report uncertainty and domain gaps, not only headline scores. A browser tool should preserve the messy details users actually need: timing, visits, options, answer keys, reports.
I try to be precise about what has been established and what has not. Structural consistency is not observational support. Synthetic generalization is not field validation. Preliminary numerical agreement is not a result. Those distinctions matter, and I take them seriously in how I describe the work.
06 Contact