An explorable, bilingual celebration of Argentina's 2026 World Cup run — with the homework done on the gossip. A fan's honest look at the refereeing whispers, the reported FBI/AFA money story, and who actually wins, built as resume-quality work.

The arc: joy → honest examination → joy with a clear head. A celebration with footnotes, not an exposé. Every probability is labelled speculative; every allegation is phrased as an allegation; the match-fixing hypothesis (H1) and the financial-probe hypothesis (H2) are kept visually distinct throughout.

By Emmanuel Oga. Made with Claude. Opinion & analysis — not reporting, not affiliated with FIFA or AFA.

How this was built. The entire site — research verification, content, sims, chat, video, and deploy config — was produced by Claude (Opus 4.8) from a single detailed brief. That brief is committed verbatim as PROMPT.md so you can see exactly what was asked for and how the result maps to it. The ideas backlog for where it could go next is in IDEAS.md.

• Bilingual site (English /en/…, rioplatense Spanish /es/…) — 7 deep-linkable routes: The Story, Could It Be Done?, The Probabilities, Who Wins?, Ask the Research, the Blog (one post per re-run), and Sources & Fine Print.
• Three interactive simulations (vanilla TS + Canvas, mobile-friendly): a Conspiracy Simulator (Grimes 2016 math), a Bayesian Explorer (log-odds), and a seeded Monte Carlo bracket simulator.
• A text-first, bracket-aware masthead — the headline tracks the bracket state (final set / champion / eliminated). The three.js hero component still exists in the tree but no longer ships on the home page's first paint.
• Ask the Research — an LLM chat grounded only in the site's compiled research, running as a Cloudflare Pages Function; unanswerable questions are logged to feed the next research re-run.
• A video render target — pnpm render:video produces a 9:16 MP4 per locale from the same content layer.
• A living re-run loop — the analysis is meant to be re-run; each run is archived and its predictions graded once outcomes are known.

All prose lives in typed Astro content collections (src/content/*.json, schemas in src/content.config.ts) — bilingual, with graded fields (headline / short / full) so any output target can pick a length without a rewrite. Volatile numbers (bracket results, model probabilities, team strengths) live in src/data/*.ts with a single visible lastUpdated. Components, pages, the video, and the chat all render from this one layer — edit a fact once and it propagates everywhere on rebuild.

src/content/   canonical bilingual prose (sections, timeline, hypotheses, precedents, sources, glossary, updates)
src/data/      volatile numbers (bracket, models, team strengths, bayes config, meta.lastUpdated)
src/components/ layout, chrome, sims, page bodies (render FROM the layers above)
functions/api/ Cloudflare Pages Function for the chat (grounded in a compiled corpus)
scripts/       corpus builder, OG images, sanity checks, prediction scorer
video/         Canvas + ffmpeg render target
research/       dated archive of each analysis re-run (findings + machine-readable snapshot)

• Astro 7 + TypeScript (strict), pnpm, Vite. Toolchain pinned in mise.toml (Node 26, pnpm 10) — run mise install for a reproducible setup.
• three.js (lazy-loaded, hero only). Sims are plain TS + Canvas — no UI framework.
• Self-hosted fonts (Bricolage Grotesque / Instrument Sans / Space Grotesk) — no CDN.
• No backend for the site, no cookies, no localStorage. The chat is an optional serverless function.

mise install            # or: use Node ≥22 + pnpm
direnv allow            # optional: auto-loads the toolchain + local secrets (.envrc)
pnpm install
pnpm dev                # http://localhost:4321  → redirects to /en/
pnpm build              # builds the corpus, then the static site into dist/
pnpm preview            # serve the production build
pnpm check              # astro check (types)
pnpm sanity             # sim-math sanity checks (MC convergence, Bayes, Grimes)

1. Edit the relevant result in src/data/bracket.ts (fill in a semifinal/final, advance the stage) and/or refresh src/data/models.ts.
2. Bump lastUpdated in src/data/meta.ts.
3. git commit && git push — the host auto-deploys. Tab 4 (and the Monte Carlo) re-render for the new bracket state automatically, including an "Argentina eliminated" state that keeps the joyful-but-honest tone.

One command: mise run update runs the whole loop end-to-end and is the natural thing to repeat — by hand or on a cron. It's additive: it pulls the current state, does more online research (accumulating new verified facts/sources, not starting over), decides whether the run is worth a Log entry, does the rioplatense Spanish pass, scores predictions, refreshes the chat's RAG corpus, builds, and deploys to Cloudflare. Scope it when you only need a piece:

mise run update                    # everything (additive): research → log → translate → score → build → deploy
mise run update:research           # additive research refresh (+ the Log decision)
mise run update:translate          # just the Spanish editorial pass
mise run update:publish            # just build + deploy (rebuilds the RAG corpus too)
mise run update -- research build  # an arbitrary subset, in order
mise run update -- --no-deploy     # everything except the deploy
mise run update:scratch -- "topic" # bootstrap a brand-new research topic from scratch

Which one do I want? (run mise tasks to see all of these with descriptions)

Rules of thumb: update is additive and safe to repeat — it accumulates research and merges, it never wipes and restarts (only update:scratch does that). Steps run in order and stop before deploy on any failure, so a broken step never ships. Nothing auto-commits — you always review git diff and commit yourself.

The research/translate steps run headless claude under the hood (edit + web-search only, no arbitrary shell) against the scoped prompts in scripts/prompts/; the rest run the plain scripts. The Blog is judged, not automatic: the research step writes a recommendation, and the log step either asks you to confirm the entry (when run interactively, with the recommendation shown), auto-adds it (non-interactive / cron), or adds nothing when the change was cosmetic. build regenerates functions/api/corpus.json, so the RAG side the chat is grounded in refreshes on every publish. Nothing auto-commits — review git diff, then commit. Doing it by hand instead? The same steps, unrolled:

1. Re-run the Phase 0 research (web search over the same checklist). Drop a new research/YYYY-MM-DD/ folder with findings.md + snapshot.json.
2. Update src/data/ and any content that changed. Add a Blog entry in src/content/updates.json (bump the run number; note what changed).
3. Once outcomes are known, fill the outcomes block of an earlier snapshot and run node scripts/score-predictions.mjs research/2026-07-12/snapshot.json — it prints a Brier score + hit/miss table. Paste that into the new Blog entry's scorecard.
4. Optionally link the previous version: Cloudflare Pages keeps a persistent URL per deployment (wrangler pages deployment list) — put it in the entry's archiveUrl.

Unanswered chat questions (logged to KV, see below) are a good source of gaps to research on the next run.

functions/api/ask.ts is a Cloudflare Pages Function. On build, scripts/build-corpus.mjs compiles the content layer into functions/api/corpus.json; the function grounds Claude in that corpus and answers only from it (speculative labels + H1≠H2 distinction enforced in the system prompt). Questions it can't answer are logged to a KV namespace for the next re-run. It degrades gracefully when the endpoint or API key is absent (e.g. on the GitHub Pages fallback), showing a fallback message.

Configure in the Cloudflare Pages project (see Deploy):

• ANTHROPIC_API_KEY (secret) — required for the chat.
• CHAT_MODEL (var, optional) — defaults to claude-opus-4-8.
• QUESTIONS (KV binding, optional) — logs unanswerable questions.

pnpm render:video       # writes video/out/la-alegria-{en,es}.mp4 (1080×1920 H.264)

Composes hook → the story in 3 beats → the four hypotheses at a glance → who-wins snapshot → vamos closing, from the content layer, on-brand from the shared tokens, with speculative labels on screen. Requires ffmpeg (installed via Homebrew here). Brand TTFs live in assets/fonts/ (fetched once); the renderer falls back to system fonts if they're absent.

1. Push this repo to GitHub under emmanueloga.
2. Cloudflare dashboard → Workers & Pages → Create → Pages → Connect to Git → select emmanueloga/fifa-2026-argentina. Build command pnpm build, output dist, framework preset Astro. Deploy. Every push to main auto-deploys after that, and every branch/PR gets a preview deployment.
3. For the chat: Pages project → Settings → Environment variables → add ANTHROPIC_API_KEY (encrypted). Optionally bind a KV namespace named QUESTIONS and set CHAT_MODEL. (Locally: wrangler pages secret put ANTHROPIC_API_KEY, wrangler kv namespace create QUESTIONS.)
4. Analytics: enable Cloudflare Web Analytics with one toggle in the Pages project (or paste the beacon into src/components/Analytics.astro). No base path needed — the site serves from the root; site is https://fifa-2026-argentina.pages.dev.

Serves from root, so astro.config.mjs keeps base: '/'.

A ready-to-use workflow ships as deploy/github-pages.yml.example (kept out of .github/workflows/ so the repo can be pushed without the GitHub workflow OAuth scope). To activate it: copy it to .github/workflows/deploy.yml, then repo Settings → Pages → Source: "GitHub Actions". It builds with a subpath base and sets SITE_URL=https://emmanueloga.github.io and BASE_PATH=/fifa-2026-argentina; all internal links go through src/lib/url.ts (import.meta.env.BASE_URL), so switching hosts is config-only — no hard-coded paths. The chat needs serverless functions and only works on Cloudflare; on GitHub Pages it degrades gracefully.

Verify both host configs locally:

pnpm build                                   # Cloudflare (root)
SITE_URL=https://emmanueloga.github.io BASE_PATH=/fifa-2026-argentina pnpm build   # GitHub Pages

Two cookie-free, consent-banner-free options are stubbed in src/components/Analytics.astro — activate one: Cloudflare Web Analytics (recommended on Pages) or GoatCounter. Google Analytics is intentionally not offered.

Everything renders from the content layer, so a new target is a new reader, not new content. To add e.g. LinkedIn/Twitter/Mastodon/Discord/Instagram cards or a slide deck: (1) read src/content/*.json (the graded fields — use headline/short for social, full for slides); (2) pick the locale key (en/es); (3) render — reuse scripts/brand.mjs for on-brand Canvas assets (see scripts/gen-og.mjs and video/render.mjs as worked examples); (4) keep the speculative labels + allegation phrasing that already live in the content. No content edits required.

Phase 0 verification corrections (Run 1, 2026-07-12) — verified facts overrode the source payload; hedging was never removed and no allegation was upgraded to a fact:

• Opta refreshed post-quarterfinals (25k sims): title France 34.0 / Spain 23.4 / England 21.9 / Argentina 20.6; the "49.1% ARG vs ENG" figure is Argentina's reach-final probability.
• Silver Bulletin ("PELE", Elo/SPI-style, 100k sims, daily) — per-team title numbers are paywalled, so we cite methodology + reputation and the public prediction-market co-favorite figure instead of inventing a number.
• Penalty superlative softened to "most in recent World Cups, ~double any other" — the exact "most over a 12-match span in tournament history" isn't authoritatively sourced. The "~23M views" Cape Verde figure was dropped (unverifiable).
• Xhaka's line is a paraphrase ("a decision where you kill the game"), attributed as such — not a verbatim quote.
• Financial thread: real and traceable to La Nación + the Miami Herald, but the $260M/$57M split and the named prosecutors are single-sourced to La Nación (attributed, not asserted); the private TourProdEnter individuals are referred to by role, not name; "two law-enforcement sources" is attributed to the Miami Herald; Grondona is DOJ's "Soccer Official #1" (identified in reporting, deceased 2014, never charged).
• Grimes exponential form confirmed as the constant-population special case of the corrected model; Garicano/Erikstad/Morgulev/Lamptey citations verified with DOIs.

Known limitations: Silver/PELE live per-team numbers (paywalled); Spain's QF winning scorer left generic ("late winner"); the Miami Herald original is paywalled (syndications used); some single-source figures carried only with explicit attribution. Full notes in research/2026-07-12/findings.md.

MIT. See IDEAS.md for the backlog (including turning this into a reusable "evidence, honestly" framework for other contested topics).