Each research programme in this repository writes its hypotheses down before it runs the experiment, numbered H1, H2 and onward across its whole log. Several sessions can work on one programme at the same time, and nothing handed out those numbers.
On 2026-11-16 two sessions of the camera self-calibration programme (research/sensing-selfcal) planned milestones 3b and 4s in parallel. Both pre-registered H104 to H107. 3b merged first, so 4s renumbered its own to H108 to H112 at merge time, and its log entry says so. The content didn't change, but a pre-registration that gets renumbered after the run is harder to audit. A run on 2026-11-13 had to do the same thing with H51 to H55.
Checking main doesn't help here, because the sibling run hasn't merged yet. Its numbers only exist on its open pull request.
What changed. research/_shared/research_log.py labels already gave the next free milestone label. Now it also:
- prints the highest hypothesis number in the log and the next free one;
- with
--prs, reads theLOG.mdon every open same-repo PR titledresearch/<slug>(fork PRs are skipped) and counts their labels and hypotheses as taken.
When I ran it on sensing-selfcal it found an open PR that already claimed the next ones:
$ python3 research/_shared/research_log.py labels sensing-selfcal --ref origin/main --prs
newest label: Milestone 3d (2026-11-16)
next free: Milestone 3f
hypotheses: highest H139 (- **H139 (the solver sees what motion capture sees).** ...)
next free: H140 (when this LOG numbers hypotheses across milestones)
open PR #16703: labels 3e, H135-H139
Without --prs, a session planning its next step would have started at H135, the same numbers as #16703.
Limits. It only sees siblings that have pushed. Research runs open their draft PR early, so that covers most of them, but two sessions that plan in the same minute can still collide. An id only counts when it sits on a line that reads like a hypothesis: a list item that opens with it, or words such as held, refuted or bar. That rule keeps out "an NVIDIA H100 GPU" and milestone-scoped ids like H5d-i. Some logs restart their numbering at each milestone, and for those the "next free" line doesn't apply, which is why it says so.
research/PROGRAMMES.md step 1 now points at the --prs form.