[shekharprateek]

Description

Two fixes to the front page, plus the chart rebuild they forced.

The video

The asset was never broken — it still 302s to a valid mp4. The markup around it was. GitHub rewrites any video embed into its own <details> player, so the wrapper the README added could only get in the way. Checked against the /markdown API rather than by eye:

Form fed to GitHub	What came back
bare asset URL	<details> player
<video src=...>	<details> player
plain markdown link	<details> player

So the chrome is GitHub's either way, and the wrapper was doing three things, all bad:

• <p align="center"> never applied. A block element cannot nest in a paragraph, so GitHub closed the <p> before the player and left a stray empty <p></p> behind it. The video was never centred.
• The fallback was discarded. The rendered <video> body came back empty — the "Your browser cannot play this video inline" text and the download link were dropped outright. Dead markup that read like a working fallback to anyone maintaining the file.
• width="820" was dropped for GitHub's own max-height:640px.

Five lines become one: the bare asset URL on its own line, the form GitHub documents. Same player, no broken wrapper, no dead fallback. The centred italic caption below it is untouched and still centres. markdownlint will flag MD034/no-bare-urls on that line — that is the intended form, and the file already carries MD033 warnings by design.

The Step 1 chart

Regenerated docs/images/cost-quality-abstracted.png from docs/diagrams/cost-quality-abstracted.html at 1600×1104 with a 2× device scale factor, which reproduces the committed 3200×2208 geometry exactly. The diagram also gained a data point that had not been rendered yet, and the rebuild was the moment to settle the naming:

• Renumbered the closed-source labels to run over the full measured set: gen 3→gen 5, gen 4→gen 6, gen 2→gen 4, mid-tier gen 3→gen 5. The chart was the last artifact still calling closed-frontier-gen5 "gen 3", so one model now carries one name across the chart, vend/swe-router/models.json, allowed-models.txt and both swe-router evaluation docs.
• Corrected the counts to twenty-three measured and eleven omitted. That number had lagged the data twice (nineteen, then twenty-one, against 23 rows in models.json).
• Repointed the footnote at vend/swe-router/models.json instead of stating the numbering is local to the chart, which stopped being true.

One more README correction

The caption under the chart told readers to regenerate this diagram "with the chart scripts in benchmarks/scripts/". It is hand-maintained HTML, and those scripts plot from run metrics. The caption now carries the actual render recipe and sends the plot scripts to their real use.

Type of Change

• Bug fix
• New feature / enhancement
• Documentation update
• Refactor (no functional change)

Checklist

• I have read CONTRIBUTING.md
• Shell scripts pass bash -n syntax check
• Python files pass python3 -m py_compile
• I have updated documentation if needed
• I have not included credentials, keys, or sensitive data

Verification

• Every relative link in README.md resolves.
• "Seven files" in Step 2 checks out against SKILL_FILES in vend/swe-router/install.sh.
• Chart values cross-checked against models.json: $11.95/82.83, $4.44/81.16, $4.67/76.97, $5.32/74.69, $4.18/66.32.
• cd benchmarks && uv run python -m unittest discover -s tests → 614 tests, one failure (test_judge_common.test_lowercase_preferred_on_collision), which also fails on a stashed tree: a case-insensitive-filesystem issue unrelated to this change. pre-commit passes.

One claim left alone: the Models-45 from 11 providers badge. litellm-mantle.yaml carries roughly 39, but the badge may be counting the repo's whole reach, so it needs an owner's call rather than a guess.