Skip to content

Notes: VRAM cliff, irreversible approval, guard placement - #10

Open
ShortArrow wants to merge 7 commits into
mainfrom
preview
Open

Notes: VRAM cliff, irreversible approval, guard placement#10
ShortArrow wants to merge 7 commits into
mainfrom
preview

Conversation

@ShortArrow

Copy link
Copy Markdown
Owner

Three notes drafted in English and mirrored in Japanese, plus the parity check now reporting which exceptions are still live.

  • docs(notes): what fits in 8 GB of VRAM, and why one model was not enough
  • docs(notes): the approval handshake is not a rule about push
  • docs(notes): put the check where the message is, not where the command is
  • docs(notes): 握手 was handshake translated, and it means nothing in Japanese
  • fix(ci): declare the Japanese-only prose repairs
  • fix(ci): report which parity exceptions are still live
  • fix(ci): repoint the parity exceptions to the rebased commit

Rebased onto main; fast-forward.

Measured on an RTX 3060 Ti: fully resident is 73-113 tok/s and any spill is
7-21, with nothing in between. Which side a model lands on is set by how its
KV cache grows rather than by its parameter count -- gemma3:4b holds 32k in
4.95 GB while granite3.3:8b reaches 13.89 GB -- and the models that hold a
long context are the ones ollama refuses to give tools to.

The notes index said every number came from one machine. This one did not,
so it says so and the index points at the note instead of promising.
An agent's destructive call is refused whatever was agreed in the
conversation; only a choice made through AskUserQuestion is read as
authorisation. That was written down as a push rule until a curl -X DELETE
against ollama came back with the same refusal, with no git involved. Asked
again through the same handshake, the identical command ran and freed
152 GB.

The note carries what has to be read back before asking, per family of
action, including the row that points the other way: a revert destroys
nothing, and listing a risk that is not there is how the next real one gets
waved through.
…d is

A check placed just before the shell call reads the command line, so it
catches an inline -m and nothing else. Measured: a message passed with -F,
an --amend --no-edit, a gh --body-file, a string assembled at run time and
a push of commits written elsewhere all went through. The five that had
actually reached this history arrived by pull, so that check could not have
stopped any of them.

commit-msg is handed the finished message as a file and pre-push can
enumerate the commits, which closes all five. What stays open is
--no-verify, and the fact that a message explaining the guard is
indistinguishable from a message carrying what it refuses.
…anese

The English was written first and carried across. 握手 in Japanese is two
people shaking hands; the protocol sense does not exist. Replaced with 手順.
断崖 went the same way -- a cliff is an English figure for a discontinuity,
not a Japanese one -- so the heading now says what happens instead of what
it resembles.

While there, the guard note had eight sentences in a row ending in the same
ます form. Two of them moved to 常体 inside the list, where Japanese
technical writing puts it anyway. Longest run is now two.
Three notes were fixed on the Japanese side alone. The English carries no
equivalent defect: handshake and cliff are ordinary English, and a run of
identical sentence endings is a Japanese phenomenon.
An exception names the commit that may stand alone, so it expires by itself
when the next commit touches that file. Expired lines then sit there doing
nothing, and reading the file cannot tell them from live ones -- three of
the six were already dead. The check now prints live, expired, or does not
resolve, which also surfaces the ones a history rewrite broke.

Three expired entries deleted. Two notes added to the header: what the
exception actually covers, which is everything that commit did to that file,
and that an expired line is meant to be removed.
Rebasing preview onto main rewrote a2f144d as 13a2ef3. The exceptions
name the commit by full hash, so all three went stale at once and a clean
clone would report the old hash as unresolvable.
@cloudflare-workers-and-pages

Copy link
Copy Markdown

Deploying with  Cloudflare Workers  Cloudflare Workers

The latest updates on your project. Learn more about integrating Git with Workers.

Status Name Latest Commit Preview URL Updated (UTC)
✅ Deployment successful!
View logs
dotfiles fe83897 Commit Preview URL

Branch Preview URL
Aug 16 2026, 10:39 PM

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant