hp /blog.
the blog — reasoning in public

The reasoning,
written down.

Long-form essays where the discipline gets argued in public — each with its figures drawn out, its numbers traceable to committed evidence, and its code and data one link away. The blog is its own surface, styled apart from the rest of the site's appearance — its figures only restyle diagrams and tables already committed in each project's docs.

essay · towards ai · jul 2026 · 9 min read →

When You Can't Measure What Matters: A Closed-Loop Harness for Probability-of-Default Estimation

Lending data only labels the loans a past policy approved — so the declined cohort, exactly where a new PD model must be right, is the one it can't be graded on. Plant the truth in a synthetic causal world, hide it the way approval policies do, escalate selection severity until correction fails, and report the operating frontier as a distribution, not a point.

selective labelscausal attributiondeclined-cohort calibrationoperating frontier
published in Towards AI code & data /cldd figures · tables · citations inside