Cross-project retrospectives, technique write-ups, and field notes.
When a dashboard, a SQL query and a BI model each define enrollment their own way, they will disagree, and nobody can say which one is right. Define each metric once, as data, and compile it to every tool.
The clinical data platform on this site now has an AWS warehouse of record. It runs on free quotas, no stored key can reach it, and no page request ever touches it.
Every form CRF Studio exported failed CDISC's own ODM schema, and nothing on the site had noticed. Here is what turned up when I checked the exports against the standards' own tools.
This site used to ship whenever I remembered to press a button. Now it ships once a week, and a release only counts after the live site has passed its checks.
I ran CDISC's own rules engine over sixteen deliberately messy versions of a study. Its output was unchanged for twelve of them, and that is the right answer to a different question.
For twelve days every release of this site needed a manual override past a failing check. The fix was to delete the check, and the reason is worth keeping.
My CRF Studio advertised compliance the code never implemented. Taking the claim down took a day. Checking what was left underneath found far more.
Thirty-eight pull requests reached main in one day from four parallel agent lanes. What made it work was a plain rule: only one of them merges at a time.
Each arcade game here is being split into one headless session and three views. The first move touched 57 files and, by design, none of the gameplay.
A mutation run over the study data derivations found weak spots that coverage missed, and one gap that was not in the tests at all but in the test data.
My first job was a drive-thru window, on and off from 2014 to 2017. I turned one shift of it into a small first-person game. These are the notes behind it.
An inspector asks you to prove what your pipeline did to one record on one date. Most pipelines can't answer.
A test suite samples a function's behaviour. A proof quantifies over it. Knowing which one a problem deserves is most of the engineering.
In one week this project found four controls that reported success while being wrong. Every one was caught the same way, and none by the tooling built to catch it.
Treating a typeface as a brand decision rather than a reading system is a defensible choice, right up until you watch someone you love fail to read your page.
Most repositories that claim to be agent-ready have documented their conventions and wired none of them. An agent reads the document and believes it.
Why regular expressions break down on clinical edit checks that an auditor will question, and how compiling each rule to an Abstract Syntax Tree gives it one defined meaning.
Fixed-timestep physics that runs the same on every screen, sub-pixel rasterization, and a browser game loop that allocates nothing.
Modelling frontend UI state as a statechart, so a high-stakes workflow has no impossible states and no unreachable ones.
How to stream CDISC ODM XML into submission-style SDTM datasets in bounded memory, with the same output on every run.