Skip to content

0004. UIUCTag: run-anchored sample grid, lazy block open

Date: 2026-08-31

Five changes made to the UIUCTag data-collection plan while building its firmware stage (S3), cut verbatim from embedded/tags/UIUCTag/design/data-collection.md. The date is that of the commits adding UIUCTag data collection and anchoring the sample grid at the run start.

Decision

Five decisions were changed while building S3.

  1. Blocks open lazily, and the checkpoint is written after the sample it anchors. The plan opened a block on entry and at each rollover, before writing samples. Instead the first sample of a block writes its pressure and temperature first, then appends the checkpoint. A reset in between now costs one block of samples rather than leaving a checkpoint that points at data which was never stored, and every checkpoint describes a block holding at least one sample.
  2. No sample at RUNNING entry. The first sample lands on the first sample boundary after entry, so entry does not have to special-case a pressure conversion alongside accelerometer setup. At most one slot is lost at run start, and it is recorded as an ordinary gap.
  3. Hibernation is permitted at any sample boundary, not only a block boundary. A resumed run anchors a fresh block at its own first sample, so there is no reason to defer hibernation by up to two hours. The wake that enters hibernation still stores its sample first.
  4. Activity spanning a reset is left unwritten rather than reconstructed. lastwrite is cleared on entry, so the sample written before a reset keeps an erased activity word. Its accumulation is genuinely lost, and the log says so instead of recording a partial count as if it were complete.
  5. The sample grid is anchored at the run's first minute boundary, not at absolute epoch multiples. See the note at the end of Time mapping. It removed the host's normalization step entirely.

Alternatives considered

An earlier revision of this plan anchored the grid to absolute epoch multiples of 7200 instead, and had the host round a header epoch down to its window. That worked, but it cost more than it bought: the first block of a run wasted the slots before the run started, a header epoch and its slot 0 were different instants, and two blocks could share a normalized start after a restart inside one window. Anchoring at the run's first minute boundary removes all three, and reduces the host's time reconstruction to one addition.

Consequences

The resulting record layout, time mapping and firmware/host contract are in data-collection.md.