· 31 August – 6 September 2026
Making the paper say only what the analysis licenses
A week of walking claims back. The title and the release boundary disagreed, the safety language was stronger than the evidence, and a recovery path only worked on the machine it was written on. Also built the perception bridge Research 3 would later depend on.
This week
No new experiments. The work was reconciling what the publication surfaces said against what the frozen analysis actually supports, which turned out to be a longer list than expected.
What changed
The title and the release boundary disagreed. The title implied a scope the release candidate did not cover. One of them had to move, and it was the title, because the boundary was fixed at freeze and moving it afterwards is how a study quietly becomes a different study.
Safety claims were clarified. Language that read as "this makes navigation safer" was narrowed to what the hypotheses licensed, which is that calibration improved and the safety hypotheses did not pass. The gap between those two sentences is the entire result.
Recovery portability. The recovery path had absolute assumptions baked in and worked only where it was written. A reproduction step that depends on the author's filesystem is not a reproduction step.
Building for a study that had not started
The landmark perception bridge for Research 3 was defined and tested in the ROS overlay this week — built inside Research 1's repository because that is where the perception stack and the verified platform already were.
That decision looks efficient now and it carries an obvious risk: a component shared across studies means a defect propagates across studies too. It is worth recording that Research 3's grounding stage descends from here, because when Research 3 later hit a hard limit in its scoring pipeline, the provenance of that pipeline was part of the diagnosis.
Failures
Nothing broke. The uncomfortable observation is how much of this week was correcting overstatement I had written myself a fortnight earlier, when the result was fresh and the temptation to describe it generously was strongest. Freezing the analysis protects the numbers. It does not protect the prose, and the prose is what most readers actually read.
Open questions
Whether an independent reader will agree the claims are now correctly bounded, or find more of the same.
Next
Assemble the reviewer packet and send it out.