My independent studies compare existing methods, examine their mistakes and document what the results can support. Each project below links to its code, methods or written findings.
Phishing feature robustness
Independent studyHow well does a phishing detector work when some of the information it relies on is missing? I compare models and study the errors caused by losing website or URL features.
Methods and limitations
Grouped historical-data tests and a paired screenshot/URL study examine feature reliability. In the paired study, URL ranking exceeded compact image and equal-fusion baselines. Dataset and collection limits remain; this does not establish performance on current attacks.
Read the study and evidence
Citation-support checkpoints
Independent studyWhen an AI answer includes citations, do those sources actually support it? I compare ways of flagging answers for review and examine simple factors, such as answer length, that can affect the results.
Methods and limitations
Uses ALCE sentence labels and 900 RAGTruth answers from 150 question groups. Length and generator controls complicate claims about the tested evidence-reading scores. The manuscript is a working draft, not peer reviewed.
Read the study and evidence
Early transient prioritisation
Independent studyCan the first few observations help identify which astronomical events deserve closer attention? I compare models that rank simulated events when follow-up resources are limited.
Methods and limitations
A RAPID-inspired GRU and forest model use matched three-day PLAsTiCC observations. In one independent simulated test, both missed the two targets at the 5% follow-up budget. No performance on observed astronomical events is claimed.
Read the study and evidence
These are independent studies using public datasets. Manuscripts are working drafts, not accepted publications. Results from historical or simulated data do not establish performance in real-world use.