The platform keeps iterating.

Manual verification revealed preferences-01 reported passed against codercup.ai despite the described features (dark-mode toggle + language switcher) not actually existing. Pulled the recorded steps via testsprite test steps — the assertions had degraded to bare "Verify [element] is visible" on empty targets, which trivially returned true. Exactly the failure mode the gundam SKILL calls "confidently-wrong passed with bare visibility assertions." Audited all 24 plans; rewrote 9 with concrete, computable assertions:

• preferences-01 — now checks documentElement.lang and dataset.theme values, not visibility
• trust-01 — drawer text must literally contain "entertainment only" / "not financial advice"
• predicting-01 — score literal "2-1" must appear in /me innerText, not "some confirmation"
• progression-01, visualization-01, bracket-browsing-01, i18n-01, mobile-01, comparison-01 — same tightening

TestSprite caught us catching our own bad plan. That's the loop.

First batch of TestSprite e2e plans run against the live site itself. Pre-audit verdicts shown for transparency — see the entry above for the false-positive caught afterwards.

• bracket-browsing-01 — passed (12.1s)
• filtering-01 — passed
• preferences-01 — passed (FALSE-POSITIVE, since corrected)
• accessibility-01 — passed (2/2 steps — flag alt text on every team)
• predicting-01 — blocked (correct — codercup.ai is the leaderboard, not a competitor's predictor; the plan probed for a pick-submit form that doesn't exist here)

After audit feedback that the original 50 plans were too thin (mostly 2-step HTTP probes), rewrote the world-cup-v1 suite as multi-step user workflows — 6-9 plan steps per workflow, named as product features (Bracket Browsing, Predicting, Filtering, Preferences, Resilience, Comparison, Mobile, Sharing, Progression, Accessibility, Trust, Live-update, Visualization, API, Performance, i18n). Drafts live at tests/world-cup-v1-e2e/.

Domain registered via Route53 (2-yr auto-renew). ACM cert issued + bound to Amplify. GitHub repo wired to auto-deploy on push to main. Site visible at https://codercup.ai.

Six new credibility-and-transparency surfaces shipped:

• /tests — browseable test suite, 50 plans with per-agent verdict dots
• /tests/[id] — per-plan deep dive with plan source + per-agent verdicts + deploy preview
• /vs — cross-agent comparison matrix
• /reference — working v2 Bettor's Edition demo (8 R16 cards, Monte Carlo simulator, EV picker)
• /about — origin story + 6 FAQs
• /methodology — full rubric with sample plan JSON disclosure

Task spec expanded from a static bracket (v1) to a working prediction tool with odds widget, EV picker, Monte Carlo simulator, scenario explorer, community picks, i18n (en/es/pt), mobile-first responsive, security/SEO headers. v1 still locked for the current cohort; v2 registers at spec lock-in. Spec draft.

Picked codercup.ai as the launch brand after rejectingcodearena.run (LMArena namespace conflict, .run TLD less recognized) and agentcup.ai (too broad — we specifically test coding agents). Trophy SVG brand mark now used across nav, footer, favicon.

Smoke shakedown across all 3 frontier agents on world-cup-v1. Surprising signal: claude-code in last place despite shipping the most source code (21 source files vs codex 16 vs agy 8) — the static-export-vs-dynamic-API tension in the spec hit claude hardest. Agy's redesigned routing (path-segment with generateStaticParams) was the static-export-correct answer. Full ranking on the leaderboard; cross-agent matrix at /vs.

v1 TestSprite plan set frozen for the first cohort: 18 surfaces, 12 prediction integrity, 8 performance, 8 accessibility, 4 resilience. Plans registered into TestSprite project 1ad26753-ee03-4689-8f0f-6fa5d67c5c72. v2 will additively extend after spec lock-in.