KRONOS

Causal Domain-Invariant Learning for Cross-Domain Generalization

KRONOS is a research repository for causal/domain-invariant representation learning. It combines a custom KRONOS architecture, baseline implementations, multi-domain experiment runners, H-divergence analysis, ablations, and optional Google Cloud integration helpers.

Project Metadata

Field	Value
Author	George David Tsitlauri
Affiliation	Dept. of Informatics & Telecommunications, University of Thessaly, Greece
Contact	gdtsitlauri@gmail.com
Year	2026

Evidence Status

Item	Current status
Core KRONOS model implementation	Present
Baseline comparison pipeline	Present
Multi-domain result artifacts	Present
H-divergence analysis	Present
Optional GCP integration helpers	Present
Uniform empirical dominance across domains	Not supported by current committed summaries

Research Positioning

The strongest credible claim supported by the committed artifacts is:

KRONOS is a serious causal/domain-generalization research implementation with competitive multi-domain performance and stronger representation alignment on the committed H-divergence analysis, but the current results do not show uniform dominance over all baselines on aggregate predictive metrics.

Current Result Snapshot

Source: results/universal/universal_results_summary.csv

Model	Mean F1	Mean AUC
KRONOS	0.7826 ± 0.2186	0.8595 ± 0.1859
GroupDRO	0.7845 ± 0.2065	0.8621 ± 0.1824
ERM	0.7837 ± 0.2065	0.8619 ± 0.1822
DANN	0.7818 ± 0.2064	0.8572 ± 0.1872
IRM	0.7781 ± 0.2038	0.8570 ± 0.1799

These aggregate summaries show KRONOS as competitive rather than uniformly best.

Representation Alignment

Source: results/hdivergence/hdivergence_results_summary.csv

Model	Mean H-divergence
KRONOS	1.6281 ± 0.1834
IRM	1.9916 ± 0.0038
DANN	1.9919 ± 0.0033
ERM	1.9932 ± 0.0008
GroupDRO	1.9932 ± 0.0008

This is the clearest empirical strength in the current repository: KRONOS appears to learn materially more domain-aligned representations in the committed H-divergence analysis.

Flagship-Safe Positioning

KRONOS is close to flagship level when presented with the correct emphasis:

• not as a universal winner on every predictive benchmark,
• but as a strong causal/domain-generalization implementation with unusually good committed representation-alignment evidence,
• plus real experiment runners, baselines, ablations, and cloud-facing engineering integrations.

That framing is both stronger and more defensible than a generic "best-across-all-domains" story.

Methodology Notes

• The repository mixes real and synthetic domains depending on the experiment family.
• Results are best interpreted as a domain-generalization research study, not a universal proof of causal recovery.
• The theoretical sections should be read as motivating framework claims unless explicitly backed by committed experiments.

Optional Cloud Integrations

The repository includes experiments/gcp_integration.py, which exposes optional:

• BigQuery querying
• GCS artifact upload
• Cloud Functions deployment helpers

These are real integration paths in code and are reflected in the dependency notes below. They are supportive engineering integrations, not the core scientific claim of KRONOS.

Repository Layout

src/kronos/
baselines/
configs/
experiments/
  run_universal_tuned.py
  run_hdivergence.py
  run_ablation.py
  gcp_integration.py
results/
paper/
  kronos_paper.tex
requirements.txt

Reproducibility

Install:

pip install -r requirements.txt

Main runs:

python experiments/run_universal_tuned.py
python experiments/run_hdivergence.py
python experiments/run_ablation.py
python experiments/analyze.py

Limitations

• Current summaries do not justify a claim of universal empirical dominance.
• The repository mixes ambitious theory with a still-evolving empirical story.
• Some domains are synthetic proxies rather than operational deployments.