The previous commit referenced a non-existent generatedAt and slipped
past a grep-based gate (grep exit 0 on finding the error). Build gate
checked by exit code from here on.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
When the live connection drops, a slim notice under the header shows
the timestamp of the last snapshot — the page keeps working from it.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The Methodology page lists each recompute (the finished match that
triggered it) with the biggest title-odds move it caused — the model's
thinking made visible, feeding the openness story.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
/compare (also in the mobile More sheet): pick any two of the 48 —
defaults to the top title favourites — and get the model's
neutral-ground matchup plus the full head-to-head from the historical
archive (wins/draws/goals and the last six meetings).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
A compact card up top: yesterday's scores at a glance and the teams
whose championship odds moved most over the last day (from the model's
recompute history), with up/down trends linking through to team pages.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
A keyboard-first palette (plus a header button) that jumps to any team
page or match page — arrow keys, enter, esc; flags and kickoff days in
the results. Pure client, searches the snapshot already in memory.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
buildRatings samples each WC team's rating along the full walk (yearly
historically, monthly since 2022) into elohistory.json (~72KB, 48
teams). Team pages plot it as a clean SVG line with year and rating
gridlines — Germany's century, 1908 → today, in one glance.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The scoreboard gains a Biggest surprises card — finished matches whose
actual outcome the frozen model priced under 30%, sorted by shock
value, each linking to its match page. Pure client derivation from
data the scoreboard already had.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The CV harness gained a calibration experiment: per-outcome isotonic
maps (pool-adjacent-violators) fit walk-forward on the five folds and
judged leave-one-fold-out against a pre-registered bar. It cleared it:
LOFO ECE improves 16% (0.0146 → 0.0122) with RPS slightly better too
(0.1724 → 0.1722). The final maps (fit on all folds) are committed as
a research artifact, embedded into ratings.json by buildRatings when
the shipped config matches, and applied in matchMatrix by scaling the
score-matrix outcome regions — scorelines, Monte Carlo and displayed
probabilities all stay mutually consistent. Without the field, output
is bit-identical (golden tests).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Once a group reaches its closing stretch (≤3 open matches), every
remaining win/draw/loss combination is enumerated and ranked with the
real tiebreaker logic. Under each table: Top 2 secured / Top 2 with a
win / Still possible (n of N outcomes) / Can't reach the top 2 — with
an honest footnote that tight cases hinge on goal difference and third
place runs through the cross-group best-thirds race. Pure, tested
enumeration in src/lib/scenarios.ts.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
A star on the team page (localStorage, no accounts) pins that team's
next or running match in a My team section at the top of the Live tab.
EN/DE strings + aria-pressed on the toggle.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
A wand toggle turns the bracket interactive: tap a side to send that
team through; W/L placeholders resolve down the tree (pure, tested
logic in src/lib/whatif.ts). Real results always override picks; picks
that stop making sense after an upstream change are pruned. Each open
pairing shows the model's advance probability (ratings.json fetched
lazily), and once your final resolves, a summary banner names your
champion with the model's chance of exactly that run. EN/DE included.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The odds history we have been capturing every 3h finally gets a face:
de-vigged DraftKings probabilities drift as solid lines toward
kickoff, with the model's numbers as dashed references — the honest
benchmark framing, right on the match page. Renders only when at least
two pre-kickoff lines exist. EN/DE copy included.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
While a match runs, every poll records the in-play model probability
per game state (insert-if-new on minute+score) into inplay_history.
The match page draws it as a stacked area chart — home fills from the
bottom, away from the top, the gap is the draw — with minute ticks and
a 50% guide. Shows live (under the win-probability bar) and after full
time as its own card. EN/DE strings included.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Five two-year walk-forward folds (2014-2024), each with its own
pre-fold parameter fit; selection by mean fold RPS; staged greedy over
Team-DC decay/shrinkage/weight, fast-Elo blend, form multiplier, and
new Elo-walk axes (home advantage, K scale); untouched 2024-26 final
test. Result: the shipped v4 config holds — the staged champion gained
0.0003, under the 0.0005 ship bar, and the final test agrees. v4 is no
longer a one-window result.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The first prod snapshot landed as a 0-byte file with a stale journal:
SQLite needs byte-range locking on the VACUUM INTO destination, which
the CIFS storage box refuses. Now: vacuum into a local temp file, then
a plain sequential copy to the archive (always CIFS-safe), cleaning up
the temp and any leftover journal.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The 'external drive' turned out to be a CIFS-mounted Hetzner storage
box — SQLite cannot run on a network mount (WAL shared memory and
locking break), and the live DB is only ~7MB anyway. So: the hot
database stays on the local volume, and a nightly VACUUM INTO snapshot
(transactionally consistent, a plain sequential file write — perfect
for CIFS) lands in /data/archive, which compose binds to the storage
box via HOST_ARCHIVE_DIR. 14-day retention, plus a snapshot at boot.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
BACKTEST_QUICK=1 skips the hyperparameter grid and scores only the
shipped + previous configs — same test metrics, a fraction of the
build time on the 1-CPU VPS. The full grid search stays a local
activity. Verified byte-level: quick output matches the grid run.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
- Bracket on phones: a segmented round picker with full-width cards
replaces the 1040px sideways scroll; desktop keeps the columns.
- Odds table: sticky team column; the three lowest-value columns hide
on phones so Team/QF/SF/Final/Champion fit without scrolling.
- Bottom nav: 4 primary tabs + a More sheet (vs Markt, Modell, Story,
Teams, Daten, plus language & theme toggles) — 7 tabs were cramped.
- Emoji → icons: timeline events use a ball icon, card-shaped color
chips and substitution arrows; unknown-team fallbacks use a shield.
Country-flag emojis stay (intentional).
- Readability: bigger scores on match cards, minimum text sizes bumped,
probability-number triple hides on the narrowest screens (the bar
carries the information).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The hardcoded 8-match cap is gone: all upcoming fixtures with a
prediction render in day-grouped sections (every match the model can
predict — slots resolve as the bracket fills). Filter chips narrow to
the next 3 days (default) or everything, and group stage vs knockout.
Match count shown in the heading.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Upcoming fixtures now appear grouped under localized day headers
(Coming up · Mon 15 Jun) regardless of whether something is on today —
not just when the day was empty. Pure grouping logic lives in
src/lib/fixtures.ts with day-boundary tests; prediction chips on the
cards come for free from MatchCard.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
A tiny typed i18n layer (no library): en.ts is the single source of UI
copy and the clarity rewrite lives there — shorter sentences, plain
words, the dense Methodology/Scoreboard copy untangled. de.ts mirrors
its shape exactly, enforced by the compiler, with placeholder parity
covered by tests. Every page now reads from the dictionary via useT();
dates flip locale through useFormat() (en-GB/de-DE); raw server values
(status, outcome, stage) map through total records.
The header gains an EN/DE toggle (persisted, defaults from the
browser language); <html lang> stays in sync.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
HOST_DATA_DIR in .env can now point /data at a host path (the 1TB
external drive) instead of the named volume, so odds history, caches
and snapshots stop eating the 128GB root disk. A daily prune drops
cache entries stale for over a week and ingest-log rows older than 14
days — faster polling no longer means unbounded DB growth.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Validated on the strict 2018-2022 window and confirmed on the untouched
2022-2026 test set (RPS 0.1703 vs 0.1721 over 4,448 matches):
- the Elo member now blends 30% of a 3x-faster Elo walk, so recent
results move ratings much harder
- ensemble weight shifts from 75/25 toward Elo to 45/55 toward the
time-decayed Team-DC member — the recent-form model now leads
- Team-DC refits nightly (and at boot) on the 15-year window plus every
finished 2026 match, via a new committed-at-build dcTrain.json
(server-only, excluded from the PWA precache)
- a recent-form goal multiplier was also tested and did NOT validate;
it ships disabled, and backtest.json records the whole decision
buildBacktest.ts grew the experiment harness: wider xi/k grids, the
fast-Elo and form variants, a pre-registered ship bar (>=0.0005
validation RPS), and a 7-day refit-cadence check. Older ratings.json
files still work — all new model fields are optional with golden
regression coverage.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The live staleness bug: liveTick refetched ESPN summaries every 20s but
the cache TTL was 5 minutes, so in-play stats/timeline barely moved.
Summaries now take a live flag (15s TTL in play), the scoreboard TTL
drops 20s->5s, and polling gets three tiers (8s in-play / 20s near
kickoff / 10m idle) via a pure, tested cadence module. The
football-data fallback is throttled to one attempt per 15s so the 8s
cadence can never breach its 10 req/min cap. Enrichment runs every 5m
(was 20m); client refetch intervals tighten to 10s (live match page)
and 30s (scoreboard).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
- scripts/processStatsbomb.ts: one pass over the 17GB StatsBomb lake (4,235
matches) -> lakestats.json (12.82GB events / 14.9M events / 6,919 players),
scorestate.json (goal-rate multipliers by game state — shipped as an insight,
NOT wired into in-play: the raw effect is selection-biased toward stronger
teams and would distort a calibrated model), fingerprints.json (40/48 nations'
shots/xG/set-piece share from real event data). Artifacts committed; raw lake
stays local (gitignored).
- /data page + /api/data/stats: the gigabytes made visible — lake counters,
historical archive, live-collection counts (odds lines, frozen forecasts,
enrichment), per-source health with freshness. Footer links.
- Term component: plain-language glossary tooltips (RPS, Brier, ECE, Elo, xG,
de-vig, ensemble) attached where metrics appear.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
- prediction_snapshots: model + de-vigged market frozen ≤15min pre-kickoff
(or at first sight of a live match, flagged 'late' and excluded from totals).
- src/lib/model/scoring.ts: ONE shared RPS/Brier/log-loss implementation for
the backtest and the live scoreboard.
- /api/scoreboard + /scoreboard page: per-match model-vs-bookmaker bars, RPS
for both after FT, 'closer' badges, running head-to-head with honest rules
printed (small-sample caveat included). Nav: 'vs Market'.
- Methodology page now shows the three-way split + ensemble bake-off variants.
- Verified end-to-end with a simulated match: snapshot froze model 80.7% and
DraftKings 67.0% (de-vig correct), both scored at FT, late-flag honored.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
- src/lib/model/teamDc.ts: per-team attack/defence Dixon-Coles via weighted MLE
(Maher iterations), exponential time-decay, shrinkage for thin histories,
learned home multiplier; log-pool of score matrices. 7 vitest cases.
- buildBacktest.ts rewritten as a strict three-way-split bake-off: params
pre-2018, tune (xi, k, w) on 2018-2022 validation, final scores on untouched
2022-2026 test (4,448 matches). Result: ensemble RPS 0.1721 beats Elo-DC
0.1726 and Team-DC 0.1801; ECE stays 0.01. Tuned: xi=0.5 k=5 w=0.75.
- Runtime now IS the backtested model: ratings.json carries teamDc+ensembleW;
matchMatrix() pools both members for predictions, Monte Carlo and advance
probs; ModelEngine re-rating preserves ensemble params; lambdas reported as
matrix expectations (drives in-play).
- scripts/buildSquadValues.ts: Transfermarkt FIWC participants page -> all 48
squad market values (France 1.52bn ... Qatar 20m). Display/availability layer
only — June-2026 values are NOT in the backtested model (would leak).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
- docs/V3-GIGA-PLAN.md: the v3 plan (data lake to 7+GB, GIGA ensemble, scoreboard,
UI clarity) with the honest reframe: perfect accuracy = perfect calibration;
'beat the betting sites' = a live public head-to-head, scored with proper rules.
- src/lib/odds.ts: American-ml → decimal, de-vig (normalize away the margin),
overround; 3 vitest cases incl. tonight's real opening lines.
- ESPN core odds adapter (server/src/ingest/espnOdds.ts): captures DraftKings
1X2 moneylines + O/U + spread per fixture; follows $ref pointer items.
- odds_history table (insert-if-changed → clean line-movement history);
scheduler odds sweep every 3h over a 14-day horizon, first sweep at boot;
/api/odds + /api/odds/:num. Odds are a benchmark ONLY — never a model input.
- Boot speed-up: fixture↔event mapping skips dates already fully mapped.
- Verified: 54 fixtures captured with real moneylines locally.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
- Match cards surface the model's pre-match expectation (favoured team + win%).
- Everything is now clickable: group-table + odds-table team rows → team
profiles; bracket matches → match pages; match ↔ team cross-links.
- New /teams hub: all 48 nations ranked by title odds, linking to profiles
(reachable from Groups + every team link; back-links point here).
- Predict → Methodology link ('How & how accurate'); cohesive hover/transition
polish across cards.
- 26 tests pass; build clean.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
- src/lib/model/inplay.ts: in-play win probability — remaining goals modelled as
time-scaled Poisson processes convolved with the current score; recomputed each
tick. 4 vitest cases (locks at FT, lead grows decisive over time, sums to 1).
- ESPN summary normalizer extended with the live event timeline (keyEvents);
preview now carries liveStats + events. Scheduler refreshes LIVE matches'
summaries every live tick (~20s) for fresh stats/timeline.
- /match/:num is now a live match center: when live it shows in-play win
probability (driven by the WS score/clock), match stats bars and an event
timeline, polling while live; preview content for scheduled games, recap when
finished. Verified with a simulated live state (94% live vs 81% pre-match).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
- scripts/buildBacktest.ts: honest walk-forward validation — params fit on
pre-2018 internationals, tested out-of-sample on 7,988 matches (2018-2026)
predicting each game from prior data only. Proper scoring (Brier/log-loss/RPS/
accuracy) vs uniform / base-rate / Elo-only baselines + a calibration analysis
(reliability bins + ECE). Results: 60% accuracy, RPS 0.171 (beats all
baselines), ECE 0.01 (excellent calibration).
- Methodology page (/methodology): plain-language model walkthrough, the backtest
scorecard, a custom-SVG reliability diagram, and honest limits. Transparency is
the differentiator — no market odds, no overclaiming.
- ReliabilityDiagram component; 'Model' nav entry.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
- Persistence via Node's built-in node:sqlite (zero native deps) on a Docker
volume: response cache, source health, fixture↔provider id map, per-fixture
enrichment, ingest log. Runtime bumped to node:24-slim + --experimental-sqlite.
- Resilient fetcher: DB cache + per-source rate-limit + jittered backoff +
circuit-breaker (blocked sources trip + skip; UI reads DB, never breaks).
- ESPN hidden API as the PRIMARY rich source (works from the VPS where SofaScore
403s): scoreboard (live scores w/ clock) + summary (venue, H2H, recent form,
lineups, team stats). football-data / SofaScore are fallbacks.
- Scheduler: maps all 72 group fixtures to ESPN event ids, polls live on a
dynamic cadence, enriches imminent fixtures into match_ext. /api/sources/health.
- Verified: DB populates (real H2H + form for opening matches), 22 tests pass,
Docker image runs node:sqlite on the volume and persists across restart.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
- Mobile bottom tab bar (icon tabs) with the top nav hidden on small screens;
content padding so it never overlaps
- data: scripts run via bun (not tsx/node) so the oven/bun build image needs no
node — fixes the Dockerfile data:build step
- README: features, data sources, dev/test/build/deploy, env vars
- Validated the production Docker image end-to-end: build succeeds (378MB),
container serves /healthz, /api/snapshot (104 fixtures), /api/predictions
(20k sims), the SPA with deep-link fallback, and the story dataset
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
- buildStatsbombViz.ts: processes the 2022 WC Final (Argentina 3-3 France) from
StatsBomb open data → shots+xG, cumulative xG race, starting-XI pass networks,
totals, shootout result. Output committed (11KB) to avoid build-time fetch
- Custom SVG viz (no chart lib): reusable Pitch (StatsBomb 120x80 coords),
ShotMap (xG-sized dots, teams attacking opposite goals), XgRace (stepped
cumulative xG with goal markers, d3-scale), PassNetwork (nodes ∝ involvement,
links ∝ pass volume)
- StoryPage: narrative with hero scoreline, by-the-numbers bars, and the three
visualizations with explanatory copy
- Verified in browser: all three render correctly, no console errors
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
NilsBriggen