Every research programme in this repository keeps a LOG.md: one ## <date> · <title> entry per milestone, newest first. Runs that pick up a programme are told to read "the newest LOG.md entries" before choosing the next step.
On 2026-11-13 a research run on the heavy-hydraulics programme reported that the picker (research/_shared/pick_track.py) pointed it at the project README's "Next steps" section, and that the log was not reliably in order. The run found its place with grep '^## ' LOG.md | head.
We checked every log. The top entry was the newest in all of them, but three had older entries above newer ones further down:
| Log | Entries | First out-of-order pair |
|---|---|---|
research/heavy-hydraulics/LOG.md | 82 | 2026-10-28 above 2026-11-03 |
research/llm-coding-inference/LOG.md | 111 | 2026-10-25 above 2026-10-26 |
research/theism-formal-evidence/LOG.md | 18 | 2026-11-07 above 2026-11-08 |
So "read the top of the file" is mostly right, but the cost of getting it wrong is a repeated milestone.
The change: when the picker draws a programme, it now prints the log's three newest entries, sorted by date with file order breaking ties, each with its line number. For a track draw it prints the two newest for every open programme in the track. A run sees where the programme stands without opening a 57 KB file, and the line number takes it straight to the entry.
read research/christian-belief-models/PROPOSAL.md and the newest research/christian-belief-models/LOG.md entries; land the next milestone
LOG.md 3: 2026-11-13 · Milestone 6 (part): how should a fixed calibration budget be split across subgroups, ...
LOG.md 11: 2026-11-12 · Milestone 6 (part): how big a calibration subsample does a rotated item need ...
Four new unit tests cover the sort, the tie order, the cut for long titles and a missing log. The other 44 picker tests still pass.