The fix went in as seven small changes, each checked on the bench before the next:

• Readings are scheduled on the UTC clock grid, so all nodes sample in the same minute and nothing drifts.
• An alarm that comes due while a node is awake is handled before it sleeps, instead of being cleared.
• A node never releases power over a pending RTC alarm. On the V3 board a set alarm can hold the power switch on, which used to leave the node awake and doing nothing.
• A node that loses its clock listens for the hub for 4 s of every 45 s, then recovers its time from the next window. Before, it stayed awake with its radio off until the battery was flat.
• A node removed from a hub stays on the radio for its idle period instead of cutting power at once.
• A queued time update is aged by however long it waited before it is applied.
• Daily sync mode is removed, so the interval is the only schedule.

Picking the parts meant checking each passive against what it does in the signal path. Class II ceramic capacitors (X5R, X7R) are piezoelectric, so on a high-impedance bias node, wind or rain vibration on the enclosure can show up as signal. The signal-path capacitors are therefore C0G, and the bias and DC-blocking capacitors are tantalum. The mic-bias filter moved from "tantalum or polymer" to a tantalum part, because polymer capacitors leak more and that leakage would add noise to the bias. Where no DC flows, as in the ADC input series resistors, thick film does the job. Thin film stays on the resistors that carry DC or set the gain.

The check also caught one of my own mistakes. The preamp mid-rail capacitor was listed as a ceramic, which is exactly what the reasoning above rules out, and it is now tantalum.

The power and LTE sections get the same check next.

With the original timing, both nodes joined at every delay up to 30 s and neither did from 35 s. That is where the timing model puts the edge. Nodes stop listening 25 s after the slot, so a hub whose boot takes longer than its wake-up lead plus 25 s misses them completely.

Waking the hub 45 s before the slot instead of 10 s, and closing the window as soon as every node has checked in, moved the edge to 70 s, again where the model predicts. Both nodes joined at every delay up to 65 s, except one step at 30 s where a node-side bug lost the window in both passes. With both nodes present a window now takes about four seconds. A missed window costs time rather than data: after four empty windows each node arrived with ten queued readings, and the hub took all twenty in under seven seconds.

This was two nodes in one room, with a bench hub that has no modem, so it says nothing yet about fleet size or radio contention. A later single-pass re-run put the edge in the same place. The node-side bug is fixed in the next note, but I haven't repeated the sweep since.

The reed-switch input works functionally, but testing confirmed that the current GPIO choice is susceptible to interference from Wi-Fi and ADC activity. That will be corrected in the next revision.

A few tests remain, including ESP-NOW communication, RTC backup retention and longer wake-cycle testing.

The goal is to replace theoretical node-capacity estimates with measured behaviour before scaling the system further.

A final independent review identified several small but important changes before fabrication, including an incorrect SDI-12 resistor value and improvements to reverse-battery protection.

Those changes are being incorporated before prototype production.

The current architecture uses a dedicated audio ADC with configurable input paths, a low-power solar system derived from FieldMesh, and a compact LTE modem selected for coverage across Europe, Australia and New Zealand.

The audio, power and communications sections are now defined, with the MCU and processing architecture next.

I’m also tightening the written case study against the actual repositories so that public descriptions distinguish clearly between what has been built, what has been tested and what is still planned.

The review process has also become more rigorous, with important findings checked against primary-source documentation before they are accepted.

Recent work focused on short-term wind observations, better handling of manual forecast requests and improving data integrity when older sessions are edited.

The current board has been cleared for a small five-board prototype run, with a structured bench bring-up plan ready for when the boards arrive.

A read-only demo project is also available for demonstrations and conference use.