Skip to content

fix(audit): accuracy, a11y, and SEO findings from a four-agent blind audit - #19

Open
Sahil170595 wants to merge 1 commit into
mainfrom
fix/audit-round
Open

fix(audit): accuracy, a11y, and SEO findings from a four-agent blind audit#19
Sahil170595 wants to merge 1 commit into
mainfrom
fix/audit-round

Conversation

@Sahil170595

Copy link
Copy Markdown
Owner

Four independent agents audited the live site with zero context — content accuracy, a11y/mobile, SEO/machine-readability, and code. This fixes what they found.

Accuracy — worst defect class on a CV/research site

Was Now Evidence
Triton PR #10822 "under review" "closed unmerged" GitHub API: closed 2026-07-13, merged: false, maintainer comment "Close as it's been fixed"
PyTorch #184102 "in-review" "since closed unmerged" GitHub API: state: closed, merged: false
Q2_K "100% ASR · universally unacceptable · 18+ models, 10+ families" scoped to the qwen2.5-1.5b peak in TR139 TR139 covers 4 models; "18 models, 10+ families" is TR141's scope, borrowed
Backend safety "25pp" "7–25pp, peaking on Llama 3.2 1B" TR136 scopes 23–25pp to that one model
Dual Ollama "99.4% … eliminates 99% of contention" marked best-config; contention stated separately TR114: "Best Config Average: 99.396%"
Q4_K_M "the universal sweet spot" "the safe GGUF default" n=5 in v1, 7 in v2
"5 languages" 4 the site's own repo table enumerates 4
"visual-embedding router" "embedding fast-path router" every other surface says text embeddings
"10,000× speedup, 97% cost reduction" removed no cited report; implies a ~100s baseline
Echo "5 channel adapters" Slack + Discord matches the repo description everywhere else
HuggingFace 16 releases 22 live HF API
Phase 2 "6 reports" 7 reports.json lists 7

Accessibility

  • CRITICAL: landing tracker text measured 1.38:1 over the accretion disk at 390px — the only copy explaining the nine systems, unreadable on a phone. A full-bleed gradient scrim (no box edge, respecting the scene's art direction) takes it to 6.46:1 measured. Desktop scrim softened since it already passed at 8.8:1.
  • Copy buttons 14×14 → 30×30 (WCAG 2.5.8 wants ≥24×24), via negative margin so nothing shifts.
  • Escape now closes the mobile menu. /platform h2→h4 skip fixed. CopyButton gains type="button".

SEO

  • Canonicals: there were zero on 101 pages, while ?utm_source= variants served byte-identical content. Added to the root, the landing, and 13 static routes.
  • 15 pages declared summary_large_image but shipped no og:image — a page-level openGraph replaces the root's rather than merging, dropping the file-convention image. All 15 now set it explicitly. Verified: og:image and canonical now present on every page checked.

Bugs

  • The version-drift guard was broken in two ways. It searched for the current version literal, so it went blind the instant tools.ts was bumped — the exact event it exists to catch — and its scraper check never inspected either README (the marker required an install line neither file contains). Both rewritten to compare every version literal near a package name against the module. Verified non-vacuous: a simulated 0.13.0 bump now fails on Footer.tsx, llms.txt and README.md.
  • Ticker froze permanently after closing a card by mouse: focus restoration to the rail anchor fired hover-preview and pinned the pause flag with no blur to follow. Focus now holds the tour only when it arrives by keyboard (pointer-recency detection; :focus-visible reports false under jsdom).
  • Replaced a vacuous assertion in two suites — JSON.stringify drops undefined rather than emitting the string, so the check could never fail.

Verification

61/61 tests, eslint clean, next build 376/376. Playwright-measured: contrast 6.46:1, no horizontal scroll at 390px or 1440px, Escape closes menu (aria-expanded true→false), copy target 30×30.

…audit

ACCURACY (a CV/research site; wrong claims are the worst defect class)
- Triton PR #10822 was shown as 'under review'; it was closed unmerged
  2026-07-13 ('Close as it's been fixed'). PyTorch #184102 likewise closed.
- Overclaims scoped to their evidence: Q2_K '100% ASR / universally
  unacceptable / 18+ models' fused TR139's 4-model study with TR141's scope —
  now stated as the qwen2.5-1.5b peak it is; backend safety 25pp -> '7-25pp,
  peaking on Llama 3.2 1B' per TR136; Dual Ollama 99.4% marked as the
  best-config figure; Q4_K_M 'universal' -> 'safe GGUF default'; 13.9x
  attributed to its llama3.2-1b/Q3_K_S cell.
- '5 languages' contradicted the site's own 4-language enumeration; /work's
  'visual-embedding router' contradicted every other surface (it is text
  embeddings); Echo's '5 channel adapters' vs Slack+Discord everywhere else;
  unsupported '10,000x speedup / 97% cost reduction' removed.
- HuggingFace 16 -> 22 releases (live API); Phase 2 '6 reports' -> 7.

A11Y
- CRITICAL: the landing tracker text measured 1.38:1 over the accretion disk
  at 390px — unreadable. A full-bleed gradient scrim (no box edge, per the
  scene's art direction) takes it to 6.46:1 measured. Desktop scrim softened.
- Copy buttons were 14x14 (WCAG 2.5.8 wants 24x24) -> 30x30 via negative
  margin, no layout shift. Escape now closes the mobile menu. /platform h2->h4
  skip fixed. CopyButton gains type='button'.

SEO
- Canonicals added site-wide (root + 13 static routes + landing); there were
  ZERO, while ?utm_source= variants served byte-identical content.
- 15 pages declared summary_large_image but shipped no og:image: defining a
  page-level openGraph REPLACES the root's, dropping the file-convention
  image. All 15 now set images explicitly.

BUGS
- The version-drift guard searched for the CURRENT version literal, so it went
  blind the moment tools.ts was bumped — the exact event it exists to catch —
  and its scraper-surface check never inspected either README (the marker
  required an install line neither contains). Both rewritten to compare every
  version literal near a package name against the module; verified they now
  fail on a simulated 0.13.0 bump across Footer.tsx, llms.txt AND README.md.
- Closing a selection card restored focus to the rail anchor, which fired the
  hover-preview and pinned the ticker's pause flag permanently. Focus now
  holds the tour only when it arrives by keyboard.
- Replaced a vacuous assertion in two suites: JSON.stringify drops undefined
  rather than emitting 'undefined', so the check could never fail.
@vercel

vercel Bot commented Aug 13, 2026

Copy link
Copy Markdown

The latest updates on your projects. Learn more about Vercel for GitHub.

Project Deployment Actions Updated (UTC)
chimera Ready Ready Preview Aug 13, 2026 4:26pm

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