Games & the Match Cockpit

Golavo opens on games, not on an audit form. The workbench is organized around four surfaces: a Matchday home, the Match Cockpit for any indexed match, My Season for your score-picking race, the Leagues browse hub, and Model Lab for expert audit machinery. Every match also carries source-backed Match Notes. Below is what each does today.

Games (the home)

The landing page is football, offline, from the first launch. It shows:

• a recent results rail and an upcoming fixtures rail drawn from the local index;
• search over every match in the bundled index (~100,000 games — internationals, the five top European leagues, and UEFA club competitions), by team or competition;
• league shortcuts into the Leagues hub.

A fresh install with an empty record is still a full, useful page — the app opens on the games, not on an empty ledger. The header carries Search, an Aa reading-comfort control, and a Settings gear.

Match Cockpit

Open any indexed match — past or upcoming, club or international — and Golavo computes a leak-safe multi-model read on demand, at the seal’s own kickoff − 1s cutoff. A validated, content-addressed local cache makes repeat exploration immediate; it is still a live read, never a sealed artifact. It is presented as a six-chapter matchday programme: form, fitted style, history, model deliberation, verdict and your pick, then the optional analyst’s column.

• Replay — a played match, reconstructed using only data available before kickoff. It is not a forecast that existed at the time and never enters the track record; it just shows what the models would have said with the pre-kickoff picture.
• Preview — a scheduled match, with no later data to exclude.

Either way you get:

• a Conditions Snapshot before the programme chapters: city coordinates and elevation, venue-local kickoff only when the source has an exact instant, pre-match rest days, and great-circle travel from each side’s previous indexed match. Routes use a bundled Natural Earth map; GeoNames and map credits stay visible. It is labeled Context, not a model input, every missing stadium/place remains Unknown, and weather stays an explicit blocked row until a lawful issued-before-kickoff history exists;
• a gold Your call ticket on upcoming matches. Save a score, reveal the five deterministic rivals, change it until kickoff, then let the final result decide the points;
• venue-aware last-five form, opponents on focus or hover, deterministic streaks, and a compact goal-difference trend for each side;
• a result-fitted attack and defence profile, plus guarded goal-timing and penalty-share facts only when the engine’s notebook contains them;
• Two voices — Elo (ratings) and Dixon–Coles (goals) — shown side by side, with whether they agree or disagree. The Poisson variants are disclosed but never counted as extra opinions, and a climatology baseline is shown for reference. Nothing is averaged into a fake consensus — honest disagreement is the point.
• an Analysis explainer that names history support, the exact percentage-point gap between the two voices, known capability coverage, missing evidence, and hypothetical changes worth exploring. History support is not confidence, capability coverage is not accuracy, and the hypothetical list never changes a seal or quantifies missing lineups/injuries;
• A glanceable Score Outlook with the balanced over/under line, clean-sheet edge, and goal peak. Expert adds double chance, every total-goal threshold, clean-sheet comparisons, the total-goal distribution, the coherent exact-score grid, and the exact home/draw/away split in the probability mass beyond that grid.
• An honest abstain state when either side has too little history to model.

Casual and Expert

Both modes keep the same chapter order and forecast values. Casual is a concise editorial read: essential charts and one-line takeaways, with technical machinery kept out of the way. Expert visibly adds fitted parameters, precise style values, model-range bands, baseline and variant detail, all market rows, the exact-score matrix, source proof, and the full sealing audit. The mode summary beneath the match header states which depth is active.

The cockpit is machine-checked leak-safe: the cutoff proof rejects any training row dated at or after kickoff − 1s.

My Season

My Season is your private points race against five model rivals, scored only on matches you call. It keeps the pick history, standings, cumulative points, exact scores, correct outcomes, bonuses, and streaks together. See Picks, points & My Season for locking, fingerprints, and the 3 / 1 / +1 scoring rules.

Seal before kickoff

The cockpit is the live read; sealing is how you put a genuine pre-kickoff prediction on the record — from inside the app, no CLI required. For an eligible fixture (men’s senior full internationals, still scheduled, inside the seal window) the cockpit shows a Seal before kickoff action; it runs the same deterministic engine as the CLI (byte-identical), freezes the model version, seed, parameters, cutoff, and inputs, and writes an immutable ForecastArtifact.

The action is honest about scope. A fixture that can’t be sealed shows a reason-specific message — already played, seal window closed, internationals-only, or pack unavailable — rather than a dead button. After full time, the seal is scored against the actual result as a separate successor; the original seal’s bytes never change. Golavo never shows a retro-forecast for a match that already kicked off — a forecast is only honest if it was sealed before kickoff.

Upcoming rails and seal windows depend on what approved sources genuinely publish. Settings exposes each source’s freshness and a manual refresh. With consent, Golavo can check on launch and periodically while it is open. It does not monitor after the app closes. The five bundled leagues carry their full 2026–27 schedules, so club fixtures appear from the first matchday; their results arrive when the packs are rebuilt, not from a live feed.

Follow this match

Use Follow on Games or Match Cockpit to keep a fixture in a local watchlist. Golavo targets approved-source checks at followed matches and records deduplicated kickoff, venue, score and settlement-availability changes without changing the match identity. Conflicting or unverified results cannot settle a forecast. Optional local notifications require explicit OS permission.

The exact product promise is: “Golavo checks followed matches on launch and periodically only while the app is open.” Closing Golavo stops checks. No daemon, Login Item or LaunchAgent is silently installed. Offline mode keeps the watchlist and event history available and shows the last verified source freshness.

Corrections and selected-source research

From a match, you can propose a fixture, kickoff, alias, venue or final-score correction. A source URL and captured evidence are required before validation. Proposals are append-only untrusted candidates; conflicts fail closed, and local acceptance is display context only. An explicit export action is required before anything leaves the Mac.

Optional research can discover a Wikimedia page or entity for you to select. Exact source text, retrieval time and hash are retained, deterministic parsers run before optional local AI, and candidate facts route into the same correction queue. Search never makes a fact authoritative.

Leagues

A browse hub with mutually exclusive International tournaments, Domestic leagues, and UEFA club competitions sections. The domestic leagues are a historical backtesting surface; the Champions League, Europa League, and Conference League are historical browsing and competition-local analytics surfaces. Their pages do not imply a complete future schedule or live club forecast. Forward sealing covers internationals only.

Domestic league pages add competition-local strength trends and workload context, plus a season-outlook gate that refuses to simulate an incomplete schedule. Covered pages also carry a collapsed historical research disclosure. It names the competition and era before showing team-only Pappalardo/Wyscout aggregates; those numbers are never mixed with current players, current matches, forecasts, or simulations.

Model Lab

The sealing, provenance, calibration, and evaluation machinery moved here, behind the product:

• Track record — the real forward calibration record, recomputed over genuine sealed → scored chains. It starts small because history is not available on back-order.
• Backtests — held-out chronological fold metrics (log loss, Brier, ECE, RPS) and reliability diagrams, kept strictly separate from the forward record.
• Methodologies — why three of the five model families are really one voice, and how abstention works.
• Sealed forecasts — the immutable forecast list.

Old #/ledger and #/eval links redirect into the Lab, so existing bookmarks keep working.

Match Notes

Every match page carries Match Notes: deterministic, source-backed facts computed at the same kickoff − 1s horizon, so it can never read the match’s own result or anything later. Facts are labelled predictive / context / coincidence, each with its sample, base rate, source, and freshness. Coincidences are capped and quarantined (“for the pub, not the forecast”) and are never shown to the AI. Nothing here changes a forecast — the fact package has a machine-checked rule that forbids it from importing any forecast, model, or calibration writer.

The magazine-style read also surfaces signature form stats most scoreboards never show — both-teams-scored rate, scoring momentum, clean-sheet rate, and the goal character of the head-to-head — and it is de-duplicated from the “three things to know” insight cards, so the Notebook is the deeper cut rather than a repeat. See the Fact & Coincidence engine and Match Notes & optional enrichment pages.

AI Analyst Read (optional, off by default)

An optional narrative and scenario synthesis, always subordinate to the deterministic numbers. It reads and connects the match’s notes and model council (and sealed forecasts), cites only the engine’s numbers, and can never change one. The result opens with a verdict, groups the strongest findings and scenarios for a quick read, keeps exact evidence behind disclosure, and separates opt-in web findings under a not-engine-verified badge.

Choose Fast for a short read from the smaller assigned model, or Deep analysis for more evidence and connected scenarios from the larger model (usually 5–8 minutes). The panel reports whether local AI is ready and includes a compact Get or manage local models guide. The web research checkbox lives beside the read so network use is visible when you request it. It runs locally (Ollama / llama.cpp) or via BYOK cloud. See AI providers.