· 14–20 September 2026
A transform race, two host restarts, and evidence that will not be re-run
Research 3's collection schedule finished, but only after two instrumentation snapshots failed on a timestamp race. Two completed attempts lost their bytes to a host restart and were written off rather than repeated, which is a smaller loss than the alternative.
This week
The 400-attempt development expansion schedule completed: 399 emitted, zero nondetections, one infrastructure failure that used its single permitted retry. The 160-attempt validation schedule emitted all 160. Both ran under instrumentation snapshot v7.
Getting to v7 was the week's actual work.
Experiments
Snapshots v4 and v6 both failed, and they failed the same way: an exact-stamp transform race. The capture path asked for a transform at a precise timestamp before the provider was guaranteed to have published it, so whether a frame captured correctly depended on scheduling rather than on anything about the scene.
v7 gates arming on a provider transform-readiness file instead of assuming readiness. The fix is unglamorous and it is the difference between a schedule that completes and one that produces frames nobody can trust.
An earlier diagnostic pass had a related problem: the v1 assets were placed using
the localisation camera_depth_frame transform, which sits about 0.06 m behind
the frame they should have used. Six centimetres is small enough to look like
noise in an aggregate and large enough to matter for a reference-point judgement
measured to 0.35 m.
Failures
The host restarted twice on the morning of 14 September, at 07:17 and 07:38 UTC. Two attempts had completed just before the second restart but had not flushed their retained bytes, so the evidence is gone.
They are recorded as evidence_unrecoverable in both the collection report and
the kit accounting, and they were not re-run. Re-running them would have
produced two attempts drawn under different conditions from the rest of the
schedule and then filed as though they were part of it. A schedule with two
declared holes is more honest than one that is quietly complete.
Results
Nothing that bears on the research question. This was an infrastructure week, and the emissions it produced are inputs to a review that had not yet happened.
Open questions
Whether the review of these emissions will yield enough negative outcomes to fit the per-class models the analysis plan requires. The floor is five incorrect outcomes per class, and nothing so far indicates the detector is making that many mistakes.
Next
Hand review of every scoreable emission, then fit against the preregistered gate.