Skip to content

Commit 4b5a9cb

Browse files
docs: cite the red-push proof that Render follows release on this host
Item 13's notes say the first promoted run cannot distinguish autoDeploy-from-main from autoDeploy-from-release: both branches hold the same sha, so the wire looks identical either way. The discriminator is a RED push, and this host produced one by accident. Run 33337712632: ee75b5b reached main at 21:55:01Z and the matrix went red at 21:57:01Z on ci/lint, so the deploy job SKIPPED and `release` never moved — it went bdd8e77 -> 625c91c and never held ee75b5b. Those are this session's own readings. The ops seat's sampler, which this session did NOT run, shows the wire holding bdd8e77 across the whole window 21:54:34-22:01:29. On a host whose promote->wire time is about 90 seconds, a Render watching `main` would have built the red sha inside that window. It did not. Both halves are needed and the fence says which is whose: the run evidence is mine, the in-window wire sampling is the seat's. Recorded because `deploy: release-branch` can now be cited as measured rather than consistent-with, and because the next fork to adopt item 13 can reproduce it deliberately instead of waiting for a lint failure to hand it over. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01EjfiNPKsgLwAxtN3BexPSJ
1 parent c462973 commit 4b5a9cb

1 file changed

Lines changed: 19 additions & 0 deletions

File tree

‎DIVERGENCES.md‎

Lines changed: 19 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -187,6 +187,25 @@ re-measure when you change what this host serves:
187187
ahead of it is an uncertified push pending, never drift
188188
and never a hand deploy. ABSENT reads as `main`.
189189

190+
MEASURED on this host, not merely consistent-with, and by
191+
the negative control rather than the happy path. Item 13's
192+
notes say a first promoted run cannot tell
193+
autoDeploy-from-main from autoDeploy-from-release, because
194+
both branches then hold the same sha. A RED push separates
195+
them, and this host had one. Run 33337712632: `ee75b5b`
196+
reached `main` at 21:55:01Z and the matrix went red at
197+
21:57:01Z on `ci / lint`, so the `deploy` job SKIPPED and
198+
`release` never moved — it went `bdd8e77` → `625c91c` and
199+
never held `ee75b5b` (this session's own readings). The ops
200+
seat's sampler, which this session did not run, shows the
201+
wire holding `bdd8e77` across the whole window
202+
21:54:34–22:01:29. On a host whose promote→wire time is
203+
about 90 seconds, a Render watching `main` would have built
204+
the red sha inside that window. It did not. The two halves
205+
together are the proof; neither is sufficient alone.
206+
Corroborating, from the seat's timing of the next run:
207+
`release` 22:01:29Z → wire 22:02:56Z, promote-first.
208+
190209
`ai_bots` is CURRENT: round 3.4 (owner decision, 2026-08-30) retired
191210
this host's training wall in `run.py` (`block_ai_training=False`), the
192211
flip deployed in run 33337897170, and the values below are the wire's own

0 commit comments

Comments
 (0)