Technical reports written from my own products. Each states its question, its method, its numbers and their limits, and ships with the scripts and data to rerun it. None of them has been peer reviewed; they are written to the standard that would be.
From WayClubTechnical report · Nov 2026
119×shorter waits behind slow requests after one lock fix
Moving an audit-log lock to commit time cut a customer's median wait behind a slow request 119-fold, in local tests on one laptop.
From BupplesTechnical report · Nov 2026
97%of float-based ledgers ended up not adding up
Integer cents kept my expense app's ledger exact over about 4 million synthetic bills; rounded floating-point shares unbalanced 97% of ledgers.
From AdelanteTechnical report · Nov 2026
8 bugsaround clock changes that the existing tests missed
My app's quote rule runs in Dart, Swift and Kotlin. Running all three side by side found eight disagreements its string-matching tests missed.
From PhotoshootTechnical report · Nov 2026
4×lower frame rate once face tracking switches on
Timed all 53 effects in my browser photobooth on one Mac: shaders fit the frame; main-thread face tracking capped face effects near 30 fps.
+6Technical report · Nov 2026
67%of serious bugs were found by adversarial review
183 defects from eight AI-agent-built products: 21 sat under a passing test or claim, and adversarial review found 24 of the 36 high-severity ones.
From CareerOS and AdelanteTechnical report · Nov 2026
75%of reworded violations still got past the checks
Rule-based publication gates blocked all 22 hand-written bad drafts of career text but only 4 of 16 rephrased ones. Every draft was template-built; no language model was run.
+5Technical report · Nov 2026
71%of design corrections came after a rule existed
In 242 design corrections given to AI coding agents, 71% came after a rule for that kind of tell was already written, and correction rates were no lower once a rule existed.
+5Technical report · Nov 2026
58%of my catches came from using or viewing the app
In a coded sample of my messages to AI coding agents, 58% of problems I caught came from using or looking at the app.
