Repository navigation
Close the deploy gap: a workflow, and the audit's one verified number on the panel - #655
Merged
Merged
Conversation
A51 concluded that no test in the source repo could see the output repo go stale. That was wrong by one assumption: a test that FETCHES THE LIVE SITE can. This is that test. It compares the entry bundle t27.ai serves against the one dist/index.html loads. Vite hashes by content, so equal names mean equal bundles. Three paths, each verified rather than assumed: match exit 0, 'the live site is this build' mismatch prints both hashes and points at A51 unreachable skips -- offline is not a defect in the change It never fails the job. A merge is not a deploy, and gating one on the other inverts the order of operations. It is here so the gap shows in the log instead of being found a week later in the site's <head>. What it cannot tell you: which direction, or whether the difference matters. Whitespace moves the hash. Проверка, сравнивающая бандл на t27.ai с собранным. Не валит сборку -- слияние это не деплой; она делает зазор видимым в логе. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
…rified row A45 asked what should replace 578.8x. The obvious answer -- the audit's own "500x (FPGA projected)" -- does not survive checking: 500x tied to energy appears exactly once in three repositories, in the audit line proposing it. It is the auditor's rounding, not a measurement. Swapping 578.8x for 500x would exchange one unsourced number for another while looking like a fix, so neither figure is changed here. What did change is a duplicate. The headline cards read: 578.8x Energy Eff 100x SU(3) Speed 99.8% Cost Reduction 80.2% Code Density and 1 - 1/578.8 = 99.83%, so card 0 and card 2 were the same unsourced number twice (A46). The audit's table has exactly one row marked VERIFIED with a citation: | Memory Compression | 20x | 20x | BitNet b1.58 (arXiv:2402.17764) | VERIFIED | So the duplicate becomes "20x / Memory, verified", in all five locales. One unsourced claim removed, one sourced claim added, and the panel now carries a number a reader can check against a published paper. 578.8x still has no source. A45 stays open, and the answer to it is a measurement someone has to take, not a wording anyone can choose. Дубликат неподтверждённого числа заменён на единственную строку аудита с пометкой VERIFIED и ссылкой. Само 578.8x не менялось: предложенное аудитом 500x обеспечено не лучше. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
A51: four PRs merged, every check green, and the live site changed nothing for a week, because this app deploys from a second repository that a human copies dist/ into. check:deployed now reports that gap; this closes it. Three decisions worth stating. Manual, not on-push-to-main. Publishing on every merge would mean any merge broadcasts, and this project's rule is that content is approved before it goes out. workflow_dispatch with a required 'reason' input, which lands in the commit message. It runs the checks against the tree being published, not just the PR that produced it. A deploy that skips them is how a broken build reaches the domain. It deletes bundles the new index.html cannot reach, computed by following every ./-relative import transitively. Skipping that is how 74 orphans accumulated once, keeping withdrawn claims greppable for weeks after the components carrying them were fixed. It needs a secret named GHIO_TOKEN -- a fine-grained PAT with Contents: write on gHashTag/ghashtag.github.io and nothing else. A workflow token cannot reach a second repository. Until that secret exists this does not run, and check:deployed keeps reporting the gap on every PR, which is the honest failure mode. Workflow деплоя, запускаемый вручную: публикация на каждый merge означала бы вещание без согласования. Требует секрет GHIO_TOKEN. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Пять файлов apps/website/messages/*.json разрешены в пользу main: карточка "99.8% / Cost Reduction", которую PR заменял на "20x / Memory, verified", на main уже удалена целиком (#743, чистка мёртвых ключей и отключённых секций), как и блок su3reality с заявлениями о кремнии. Правка PR стала неактуальной — затирать удаление нельзя. package.json: обе стороны добавляли скрипты, объединены (check:deployed + check:aria + typecheck:update), JSON валиден. Полезный остаток PR: deploy-site.yml, website-checks.yml, deployed-check.mjs.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Follows #648, #649, #651, #653.
A45 does not close the way it looked like it would
The obvious answer was the audit's own
500× (FPGA projected). It does not survive checking:500×tied to energy appears exactly once in three repositories — in the audit line proposing it. It is a rounding, not a measurement. Swapping 578.8× for it would exchange one unsourced number for another while looking like a fix, so neither figure is changed.What did change is a duplicate
1 − 1/578.8 = 99.83%, so two of the four headline cards were the same unsourced number twice:578.8× / Energy Eff99.8% / Cost Reduction20× / Memory, verifiedThe audit's table has exactly one row marked VERIFIED with a citation — Memory Compression, 20×, BitNet b1.58 (arXiv:2402.17764). One unsourced claim removed, one sourced claim added, and the panel now carries a number a reader can check against a published paper.
578.8× still has no source. A45 stays open, and closing it needs a measurement someone takes, not a wording anyone chooses.
And the deploy gap gets a workflow
A51: four PRs merged, every check green, the live site unchanged for a week — because the SPA deploys from a second repository a human copies
dist/into.deploy-site.ymlis manual on purpose: publishing on every merge would mean any merge broadcasts, and this project's rule is that content is approved before it goes out. It takes a requiredreason, runs the checks against the tree being published, and deletes bundles the newindex.htmlcannot reach — skipping that is how 74 orphans once kept withdrawn claims greppable for weeks.It needs a secret,
GHIO_TOKEN— a fine-grained PAT withContents: writeongHashTag/ghashtag.github.ioand nothing else. Until it exists the workflow does not run andcheck:deployedkeeps reporting the gap on every PR, which is the honest failure mode.🤖 Generated with Claude Code