<?xml version="1.0" encoding="utf-8"?>
<rss version="2.0"><channel><title>crtdrops · Notes</title><link>https://crtdrops.mx/notas/</link><description>Technical notes with measured data</description><language>en</language><item><title>The lab&#x27;s ofrenda: what didn&#x27;t survive October</title><link>https://crtdrops.mx/notas/the-lab-ofrenda.html</link><guid>https://crtdrops.mx/notas/the-lab-ofrenda.html</guid><pubDate>Fri, 30 Oct 2026 09:00:00 -0600</pubDate><description>Six of the lab&#x27;s departed in October 2026, and a Raspberry Pi that came back from a power outage. Each with its certificate: date, technical cause of death, what it leaves us, and its calaverita.</description></item><item><title>Ghosts in the frame: my 2001 robot-vision tricks against two local vision models</title><link>https://crtdrops.mx/notas/ghosts-in-the-frame.html</link><guid>https://crtdrops.mx/notas/ghosts-in-the-frame.html</guid><pubDate>Wed, 28 Oct 2026 09:00:00 -0600</pubDate><description>A disk in haunted frames with analog TV&#x27;s echo, noise and bars. The 2001 eye, in C, against Qwen2.5-VL 3B and Gemma 3 4B on an N100: 0.1 px in 37.5 ms against 1.3 px in 146 s where Qwen gave a box, and Gemma answered the center.</description></item><item><title>The bytes that are not yours</title><link>https://crtdrops.mx/notas/the-bytes-that-are-not-yours.html</link><guid>https://crtdrops.mx/notas/the-bytes-that-are-not-yours.html</guid><pubDate>Fri, 23 Oct 2026 09:00:00 -0600</pubDate><description>A 60-byte report can cost 209 or 4,653 bytes on the network, depending on how the device connects. Where a fleet&#x27;s data goes, why shrinking the report barely helps and what does change the math.</description></item><item><title>Night shift: a local Gemma writes my 7 a.m. lab report, and I measured where it makes things up</title><link>https://crtdrops.mx/notas/night-shift.html</link><guid>https://crtdrops.mx/notas/night-shift.html</guid><pubDate>Sat, 24 Oct 2026 09:00:00 -0600</pubDate><description>An N100 and a Raspberry Pi 5 watch each other overnight; fixed rules detect and Gemma 3 4B, offline, writes. With injected faults, detection got all of them; the model invented on a quiet night and explained against the evidence.</description></item><item><title>The query that took too long</title><link>https://crtdrops.mx/notas/the-query-that-took-too-long.html</link><guid>https://crtdrops.mx/notas/the-query-that-took-too-long.html</guid><pubDate>Sun, 18 Oct 2026 09:00:00 -0600</pubDate><description>An eight-million-row table, a 504 and a plan that estimated few rows and did a lot of work. How it was read with EXPLAIN and EXPLAIN ANALYZE, what an index on operation and date changed, and what was not measured.</description></item><item><title>An open socket is not a connected device</title><link>https://crtdrops.mx/notas/an-open-socket-is-not-a-connected-device.html</link><guid>https://crtdrops.mx/notas/an-open-socket-is-not-a-connected-device.html</guid><pubDate>Sat, 17 Oct 2026 09:00:00 -0600</pubDate><description>A TCP connection can die without the server knowing. How long each mechanism takes to find out, what happens to a command written to a dead socket, and the reconnection after a restart.</description></item><item><title>Four versions of the same top-up</title><link>https://crtdrops.mx/notas/four-versions-of-the-same-top-up.html</link><guid>https://crtdrops.mx/notas/four-versions-of-the-same-top-up.html</guid><pubDate>Sat, 17 Oct 2026 09:00:00 -0600</pubDate><description>What was requested, delivered, recorded and charged have to be the same operation. What reconciling costs without a shared identifier, and why a match does not authorize moving money.</description></item><item><title>The average that never happened</title><link>https://crtdrops.mx/notas/the-average-that-never-happened.html</link><guid>https://crtdrops.mx/notas/the-average-that-never-happened.html</guid><pubDate>Sat, 17 Oct 2026 09:00:00 -0600</pubDate><description>Together, the 25,898 inferences of a Raspberry Pi 5 average 347 ms, and none took that long. Why latency should be asked for per condition, in percentiles and without discarding timeouts.</description></item><item><title>The right position at the wrong time</title><link>https://crtdrops.mx/notas/the-right-position-at-the-wrong-time.html</link><guid>https://crtdrops.mx/notas/the-right-position-at-the-wrong-time.html</guid><pubDate>Sat, 17 Oct 2026 09:00:00 -0600</pubDate><description>A device gets its signal back and sends its history while it keeps sending the present. If the platform cannot tell them apart, a correct position leads to a wrong decision.</description></item><item><title>The geofence computed on the fly</title><link>https://crtdrops.mx/notas/the-geofence-on-the-fly.html</link><guid>https://crtdrops.mx/notas/the-geofence-on-the-fly.html</guid><pubDate>Thu, 15 Oct 2026 09:00:00 -0600</pubDate><description>Positions that jump, arrive late or never arrive: geofences fail because of that, not the geometry. And computed on the fly, changing a value rewrites the history.</description></item><item><title>The throttling your system does not show you</title><link>https://crtdrops.mx/notas/the-throttling-you-do-not-see.html</link><guid>https://crtdrops.mx/notas/the-throttling-you-do-not-see.html</guid><pubDate>Wed, 07 Oct 2026 09:00:00 -0600</pubDate><description>Seventy minutes of YOLO26n non-stop on a Raspberry Pi 5 and an N100. The Pi throttles after ten minutes, and Linux keeps reporting 2,400 MHz.</description></item><item><title>The top-up nobody knows went through</title><link>https://crtdrops.mx/notas/the-top-up-nobody-knows.html</link><guid>https://crtdrops.mx/notas/the-top-up-nobody-knows.html</guid><pubDate>Mon, 05 Oct 2026 09:00:00 -0600</pubDate><description>The carrier does not answer and the system has already debited. With money, a timeout is not an error: it is a third state, and it has to be designed.</description></item><item><title>The parked truck that kept racking up kilometers</title><link>https://crtdrops.mx/notas/the-parked-truck.html</link><guid>https://crtdrops.mx/notas/the-parked-truck.html</guid><pubDate>Sun, 04 Oct 2026 09:00:00 -0600</pubDate><description>A GPS at rest jumps around, and adding up points never gave the distance. The filter everyone adds cleans up the dashboard, but hides real movement.</description></item><item><title>The ten-year battery that would not make five</title><link>https://crtdrops.mx/notas/the-ten-year-battery.html</link><guid>https://crtdrops.mx/notas/the-ten-year-battery.html</guid><pubDate>Sun, 04 Oct 2026 09:00:00 -0600</pubDate><description>A water meter designed for ten years on a 19 Ah cell. The spreadsheet said fourteen; measured in the lab, the GPRS version would not make five.</description></item><item><title>On the N100, OpenVINO was not faster. And I almost measured it wrong</title><link>https://crtdrops.mx/notas/openvino-was-not-faster.html</link><guid>https://crtdrops.mx/notas/openvino-was-not-faster.html</guid><pubDate>Sat, 03 Oct 2026 09:00:00 -0600</pubDate><description>I measured ONNX Runtime against OpenVINO on an Intel N100. With precision pinned to f32, OpenVINO was slower. Earlier, in a rehearsal, it had dropped to FP16 on its own, and validation did not notice.</description></item><item><title>Why your telemetry platform loses frames and nobody notices</title><link>https://crtdrops.mx/notas/a-read-is-not-a-frame.html</link><guid>https://crtdrops.mx/notas/a-read-is-not-a-frame.html</guid><pubDate>Fri, 02 Oct 2026 09:00:00 -0600</pubDate><description>Almost every backend receiving devices over TCP assumes each socket read brings exactly one frame. It is false by definition of the transport.</description></item></channel></rss>
