econ-radar at Five Months: Auditing a Self-Publishing Newsletter
I audited the economy newsletter that has published itself every day since June. In October alone, five blog pages never went live while the deploy script re...
What I build, how far I check it, and what the checking turns up. Bioinformatics, model validation, and AI agents.
I audited the economy newsletter that has published itself every day since June. In October alone, five blog pages never went live while the deploy script re...
Every time I wrote a review agent, I assumed review was now covered. Counting from session logs, only 16 of 97 agent definitions had ever been called. Based ...
I published a trend report on generative AI in biotech and drug discovery. One question runs through all of it: how far can you actually trust the output?
I kept starting and stopping with economic news — not from lack of interest, but from the sheer volume of it. So I stopped trying to do it by hand and built ...
The detector flagged all twenty-four planted sentences at 0.00, and then flagged the true sentences at 0.00 too. All four talks at Villain Night 2026 were st...
The number of clusters that changes with a single seed, T cells with the same phenotype that see different antigens, proteins that appear in a different plac...
Tumor DNA at 0.02% in blood, transcripts at fewer than 10 molecules per cell, senescent cells that are rare in tissue and lack a common marker. The talks on ...
I read two open-source form builders to ground the design of a small MVP. Then I had the spec reviewed by someone other than its author, and four blocking de...
I reconciled all 120 numbers in my manuscript against committed results. For 27 of them I couldn’t trace where the value came from.
Can an LLM stand in for an expert judge? I tested it on public data, missed the target, and found the reason wasn’t the model.
Our initial target metric came back at Spearman 0.04 — essentially a failure. We put that on the slides anyway.
The skeleton took a day. The interesting part came after: 20 days where the frictions and failures of running it kept reshaping the design.
The CCR cloud sandbox blocks outbound HTTPS to Telegram. It took me a while to figure out why, and the fix turned out to be simpler than I expected.
The agent doesn’t know what it wrote yesterday. Without being told, it’ll run the same drug through deep analysis again today.
I built it so nothing would go out without my approval. Two days later I dropped that, and it runs better.
Months ago I gave up on this blog after the environment setup defeated me. This week I sat down with an AI coding agent and we found that almost nothing was ...