Replacing manual approval of mailbox requests with proof of work plus local IP scoring. The design does not hinge on the algorithm but on one rule: no reputation score causes a rejection, it only decides how long you hash. Covers why a hand-written SHA-256 must be checked against block-boundary vectors, why difficulty is a UX parameter rather than a security one, and why external reputation APIs and country weighting were deliberately left out.
Building a JAX environment on a dual-GPU Windows box under WSL2 produced traps that all share one shape: what you can observe is not what is there. device_kind reports the same model for two different cards, so routing by name puts a 16 GB batch on the 8 GB one. Also a CPU-only jaxlib that installs happily and does nothing, background processes killed without a journal entry, a best-fit policy that starved the larger card, and stdin pipes that deadlock across the WSL boundary.
A router died under load, recovered on reboot, and left no evidence behind. The only way in was an observation that should have been impossible: 62 retransmissions on loopback, all concentrated in LAST_ACK and CLOSE_WAIT. Following it led to a firewall rule protecting the wrong address and dropping the teardown signals between two local processes. Includes the full failure chain, the fork that sends you after the wrong bug, and a twin-rule experiment showing why a ct condition silently never matches in a raw chain.
Eight problems found while writing an equivalent-circuit fitter for impedance spectra, none of which threw an exception. The most expensive was a missing minus sign in the Levenberg-Marquardt step: every trial walked uphill, all were rejected, and the function handed back the initial guess while the interface reported success. Also a parameter running to 1e154 along a flat valley, a guard that misclassified normal convergence, and a weighting choice worth a 250x difference in bias.
Three months, 12,491 impressions, 53 clicks. After bucketing Search Console data by position and ruling out truncated titles, missing descriptions and mobile demotion one at a time, one statistically significant split remained: in the 1-10 bucket, tool pages convert at 4.76% against 0.74% for articles, p=0.0012. Includes the failed attempt that mistook “nobody built it” for “somebody needs it”.
Adding third-party sign-in looks like a solved problem: take an OIDC library, exchange the code, verify the signature, find the user by email. Wiring up …
Wiring a self-hosted mailbox into an AI assistant: the protocol half is easy, making read-only structural rather than a promise is not. Traps hit along the way — a rejection letter whose boilerplate says has been accepted, a syntax check that passed while authentication became dead code, a shared failure latch that took every user down at once, and why a synthetic click cannot tell a dead button from a blocked popup.
I attached featured images to 19 posts. The admin looked right, the page source had no og:image — and the JSON-LD on the same page already pointed at the new cover. That pairing is the diagnosis: the two outputs travel different paths, one computed live and one read from a precomputed index table.
A tool that takes character images, cuts them out, separates layers and produces an SVG asset library — with nothing uploaded. The original motivation was a single-core origin that could not host server-side image processing; what it actually saved was storage, moderation, privacy obligations and concurrency. Covers the five-step implementation, the price of IndexedDB, replacing a compression library with ZIP stored mode, and the one unavoidable cost.
Trying to quantify whether my articles all look the same, I first measured 92% and wrote it into my conclusions. That number counted the theme template as article structure; corrected, it is 12%. This is about the measurement: defining “shape”, where I drew the boundary wrong, why one metric hides the real problem, and the accidental control that finally killed it.
