2026-08-11-public-flip-and-scrub

Public flip, runbook retirement, and the first scrub of the note dumps

Directory created 2026-08-11, after the round. The round was dispatched as short inline instructions during the branch-protection session rather than as a written prompt, so there is no document to keep verbatim. What follows is the maintainer's instructions as recorded, and nothing more — no reconstruction of intent.

The work in this round landed as PR #1, merged at b67ff8d.

The instructions, as given

Mid-session, following the branch-protection round (prompts/2026-08-11-branch-protection/), whose dispatch is kept verbatim and which staged protection behind a runbook for the maintainer to apply:

wait sorry just make it public and delete the runbook once the repo is set up

Then:

scrub the note dumps

Then, on the release gate that the flip left undischarged:

and i'm going to do the read through after it's public which is fine

What is not here

No further specification was given. The scope of "scrub" — what counts as personal, whose names, how to record a judgment call — was decided by the executing round and reported in SCRUB_REPORT.md, not set by the dispatch. The maintainer's standing instruction on personal material arrived with the round that followed, which is part of why a second scrub round was needed.

The DCO-gate fix in this pull request was not dispatched at all: the gate failed the first pull request through it, and the fix was made in the same branch.

Public flip, runbook retirement, first scrub — report

Reconstructed after the fact on 2026-08-11, from the pull-request body, the three commit messages, and the artifacts the round left in the tree. It is not the report the round wrote — the round wrote none, which is the gap this directory exists to close. Nothing here is asserted beyond what those sources support, and where the round's own reasoning is not recoverable, this says so.

Landed as PR #1, merged at b67ff8d. Three commits.

What was done

The repository went public and branch protection was applied, in the same sitting, at the maintainer's direction. Verified by read-back rather than by trusting the write: eight required checks, zero required approvals, code-owner reviews off, enforce-for-admins on, force-pushes and branch deletion blocked. Direct pushes to main are refused for everyone, maintainers included. The reasoning is in DECISIONS.md under 2026-08-11 — public, and branch protection live, which is contemporaneous.

**FLIP_RUNBOOK.md was retired.** It sequenced a flip that had happened, so under no negative ontologies it goes rather than becoming a historical note.

The deference note dump was scrubbed — two cuts, both candid assessments of a named colleague's abilities, marked [scrubbed] inline. The dose-response bundle needed none. Frozen digests were recomputed and frozen/MANIFEST.md annotated in the same change, which is the sanctioned way frozen content moves. Four judgment calls went the other way and were recorded with reasons so they could be reversed; two of them were in fact reversed by the round that followed.

The DCO gate was fixed. It counted GitHub's synthetic pull-request merge commit, which has no author who could sign it, so it failed the first pull request through it and would have failed every one after. The fix is --no-merges.

What the round said it could not do

From SCRUB_REPORT.md and the commit messages, both contemporaneous: the bundles were public before the scrub ran, and removal at HEAD is not erasure from history. A mechanical scan for emails, phone numbers, API keys and home paths came back clean across all 51 files, and that scan cannot see the two categories only a person can judge — personal-life passages, and candid remarks about named third parties. The release gate was undischarged at the flip, by the maintainer's explicit decision.

One error is recorded in the round's own commit message: an image was deleted on the strength of its filename before its contents were checked, and restored when it turned out to be a rendered-LaTeX theorem rather than an email screenshot.

Attribution

prompt author	the maintainer, inline; see PROMPT.md
executor	Claude Opus 5 (Anthropic)
dates	2026-08-11
this reconstruction	Claude Opus 5 (Anthropic), 2026-08-11, in the round recorded at prompts/2026-08-11-attribution-provenance-names/

What is not recoverable

The round's own account of its deviations and of what it left undone. Reports under this convention state what was not shown; a reconstruction from a merged pull request can only report what was. Anything the round decided and did not write down is gone, and this document does not guess at it.