From prediction to optimization
The project evolved from asking which exact numbers might appear into studying how to preserve strong candidates across a whole ticket portfolio.
Research notes
Blog and research notes
A place for transparent notes about models, optimization, performance, failures and new experiments.
V7 correction learning
Research on a correction layer that learns ranking mistakes while keeping the base model stable.
V8 attack subset selection
Work on selecting realistic attack candidates from Top50 without collapsing into one numeric zone.
T8 portfolio search
Experiments with covering designs, maximum coverage, heuristic search, Monte Carlo mutations and pair synergy.
Production safety
Public workflows can use shared models while heavy training is exported to external workers.