Repository navigation
docs: regenerate release graphics with the 19 Oct run framing - #394
Merged
Merged
Conversation
…parison Add two source-of-truth cards to /graphics/ so the public square graphics are regenerated from the site's shadcn components instead of the stale 11 Oct dev-sweep framing (62.5% 10/16, ~384s vs ~383s parity, 0.69x): - release card: v0.43.0 pitch with the 19 Oct run's figures (30/30 PASS both arms, 0.78x, 9/10) and their provenance limits beside them. - progression: one bar per mini-vs-pi run, rounds 1-6 plus the 19 Oct closure-on line, each keeping its corpus, model and judge. Round 5's single-run 0.69x is drawn hollow because round 6 did not reproduce it, and the closure-on line is not pooled with the earlier rounds.
…ming Both square graphics carried the 11 Oct dev-sweep framing. Regenerate them from the new /graphics/ cards (1080x1080, LANCZOS from a 2x cu capture): - update.png: v0.43.0 release card (bots + board + WhatsApp + browser), headline figures 30/30, 0.78x, 9/10 with the run's limits in the note. - benchmark.png: every mini-vs-pi run as one bar, each labelled with the corpus, model and judge it came from; the 0.69x is hollow. Reference both in the README's 'Graphics with separate provenance' section, which keeps provenance beside every result.
The `as const` literal gave each row a distinct type, so `hollow` and `closure` were only known on Round 5 and the last row respectively, and reading them on every row was a type error (2 errors in svelte-check). Annotate the array as `readonly Run[]` with the two flags optional; the rendered markup is byte-identical.
Owner
Author
PR review — iteración 1 (manual)
Critical Issues
Important Blockers
Medium Blockers
Non-Blocker Nits (no bloquean)
Verificación
Resultado: 0 críticos + 0 important + 0 medium → el loop queda limpio en la |
Owner
Author
|
All blockers resolved across 1 iteration (0 critical + 0 important + 0 medium; 2 non-blocker nits documented above, none actioned — one is flagged for a possible follow-up on |
…dead CSS Review delta on PR #394 (figures and linear scale unchanged): - Round 6 (0.98x) was labelled "a tie" but drawn in the legend's green "faster" colour. Colour now comes from an explicit per-run `outcome` field ("faster"/"slower"/"tie") set from the documented result — not a numeric threshold, which would have reclassified Round 3's real 0.95x win as a tie. Round 6 uses a neutral slate; the legend gained an "a tie" entry. Round 3 stays green; the 19 Oct closure line stays green. - benchmark.png re-rendered: Round 6 bar + its legend dot repainted to the neutral slate; scale, figures and every other bar untouched. - Dropped the dead global overrides for .run-row/.run-note/.run-track/ .features that sat outside the @media block and were silently shrinking the desktop legend; they now live inside @media(max-width:700px) where they belong. Folded the duplicate .metric .versus declaration. - README: marked docs/img/time-per-task.png (11 Oct dev sweep) obsolete, kept for context — it has no references left in the repo.
The exported asset diverged from the page: its legend still showed only the two old categories (missing the 'a tie' entry the page gained in 60b4321), and Round 6's bar carried a faint green antialias border left by the earlier in-place colour patch. Re-exported from the real render: build the site, capture the #progression card through the base-path server so the stylesheet resolves, scale 992x992 -> 1080x1080 with LANCZOS. The card's CSS now paints everything, so the legend has the same three categories as the page and Round 6 is neutral grey everywhere, border included. No figures and no scale touched; no benchmark or data regenerated -- only the asset render.
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
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.
Summary
The two public square graphics that ship in the repo —
docs/img/update.png(the release card) anddocs/img/benchmark.png(the benchmark chart) — were still generated from the 11 October dev-sweep framing: 62.5% (10/16), wall-clock parity of ~384s vs ~383s, and a single-run 0.69x speed ratio plotted against the public leaderboard. That framing was superseded by the 19 October closure-on run, which has already been reconciled in the v0.43.0 release body and the evidence notes indocs/pi-vs-mini-benchmark.md. Anyone landing on the README or grabbing these images for a release post was seeing numbers the project no longer stands behind, with none of the provenance that the rest of the docs now carry.This PR regenerates both graphics from source-of-truth cards in
site/src/routes/graphics/+page.svelte(the same shadcnCardcomponents the other square graphics already use), so the images are reproducible instead of hand-made artifacts. The new framing puts the 19 October run's figures front and centre — 30/30 PASS on both arms, 0.78x of pi's time as the sum of per-task medians, mini at least as fast on 9 of 10 tasks — with each figure's limits stated beside it. The benchmark graphic plots one bar per mini-vs-pi run (rounds 1–6 plus the 19 October closure-on line), and every bar keeps its own corpus, model and judge next to it. Round 5's single-run 0.69x is drawn hollow because round 6's three fresh repetitions (1.13x, 0.98x, 0.93x) did not reproduce it, and the closure-on line is explicitly not pooled with the earlier rounds because it is a different setup.No speed claim is invented here: every number is a figure already verified in the release body or the evidence notes, and the graphics state their provenance rather than collapsing distinct runs into one headline.
What Happened
Not a bug fix — this is a docs/presentation regeneration. (Kept for context: the stale images were carried over from the 11 October dev sweep and were never regenerated when the 19 October closure-on run became the headline result.)
Changes Table
Landing / site (graphic source of truth)
site/src/routes/graphics/+page.svelterunsarray so the optionalhollow/closureflags check cleanly; update the section note and the SEO description/modified date.site/src/lib/routes.ts/graphics/sitemapmodifieddate to 2026-10-21.Docs (regenerated graphics + references)
docs/img/update.pngdocs/img/benchmark.pngREADME.mdReview delta (commit 60b4321)
A bounded adjustment after review — no figures and no scale changed.
legend's green "faster" colour. Bar colour now comes from an explicit per-run
outcomefield (faster/slower/tie) set from the documented result, never anumeric threshold — a threshold would have silently reclassified Round 3's real
0.95x win as a tie. Round 6 uses a neutral slate and the legend gained an "a tie"
entry. Round 3 stays green; the 19 Oct closure line stays green.
benchmark.pngwas re-rendered with Round 6's bar and its legend dot repainted to the neutral
slate; scale, figures and every other bar are untouched.
.run-row/.run-note/.run-track/.featuresthat sat outside the@mediablock and were silently shrinking thedesktop legend; they now live inside
@media(max-width:700px)where they belong.Folded the duplicate
.metric .versusdeclaration.time-per-task.png. Marked obsolete in the README, kept for context — verifiedit has no references left in the repo. Not deleted, not re-measured.
Verification
bun run check(svelte-check) insite/— 0 errors, 0 warningsbun run buildinsite/(prerenders every route, base path loopback) — succeeded;build/graphics/index.htmlrenders all five card headings and the provenance strings byte-for-byte as before the type annotation/graphics/renders both new cards with correct bars (1.00x marker line, hollow round 5, green/grey bars) and the provenance note; no console errors; the home page still shows the closure-check section with 0.78x, 112.9 s vs 145.6 s, 30/30, 9/10Note:
docs/img/time-per-task.pngalso carries the old dev-sweep framing but is referenced by nothing in the repo and by no release body, so it was left untouched; flagging it in case it should be removed or regenerated separately.