How this is built

League results and team match stats come from Football-Data.co.uk and are labelled as such; the current-season API-Football migration is gated on a full results-and-stats coverage probe. League Cup and Scottish Cup fixtures, lineups/roles, player match stats and the squad come from API-Football. Aberdeen club fixtures come from the SPFL feed. Availability is curated official AFC team news; the provider injury snapshot is optional and labelled separately. See Sources for the layer truth table. Raw payloads are never committed; the app publishes derived tables, fixtures and listing metadata with attribution.

The next-fixture page projects both XIs from observed formations, starts, minutes and role fit, with 4–6 contenders. The style lean uses measured shot volume, shot share, shots on target and goals; possession is unknown and is not used. Form is league-only, and home/away splits use every league match. A missing formation or a small lineup sample is labelled unavailable rather than guessed.

The opponent refresh is bounded to the next fixture and cache-first. A failed refresh carries the previous snapshot as stale, or removes only the opponent XI, never the league context.

The transfer register publishes completed permanent, loan and free moves only. API-Football is the candidate feed; official AFC announcements and Wikipedia season records are the cross-checks. Fees are not in the feed and are curated only: undisclosed is never zero and is excluded from net spend. Loan end, recall and obligation terms are unknown unless confirmed.

The history layer joins the measured league record, curated staff and honours, transfer windows, style proxies and lineup-based selection. Manager records are league-only; a match in a gap is unattributed. Style is opponent-unadjusted match-stat proxies, not xG. Selection is unavailable rather than inferred when the historical backfill is missing.

“The read” is a labelled editorial assessment, not a measurement or the club's internal view. The 2023-24 League Cup was a runners-up finish and is not in the honours cabinet.

The squad is the current licensed squad plus current-season loanees out from the confirmed register. Each player has per-competition goals, assists, cards and the provider rating, plus shots/passes for outfield players and saves/conceded for goalkeepers. Starts, substitute appearances and on-pitch minutes are reconstructed from fixture lineups and substitution events, so a start and a substitute appearance are different numbers rather than one “apps” total.

Role is a primary dimension, not a decorative attribute. The licensed team sheet gives the formation and each starter's grid slot. A starting right-back, centre-back, centre-midfielder, winger or forward is read from that grid; a substitute has no grid slot and therefore no role. A missing grid or an unmapped formation is shown as unknown, not filled with the generic squad position. Kick-off shape, centre-back count and holding-midfielder count are recorded for the shape/balance view. Mid-game formation switches are not in the feed and are never inferred from substitutions; reported ones are kept as sourced context notes.

Positions are an eligibility set, not one label. The career layer uses FM-style slots (`D/WB L`, `DM/M C`, `AM R`, ...) from the last three seasons of starting-XI evidence, with primary, secondary and stretch tiers. Equal weight is given to starts; a one-off start is a stretch role, not a new primary. The broad filter is defenders/midfielders/attackers and left/centre/right. A full-back or wing-back counts as both a defender and a midfielder; a DM is a midfielder; a winger is an attacker. The current Aberdeen role is itself a multi-slot summary and only replaces the display label when it is clearly different. Curated role notes cover interpretations the kick-off grid cannot see, such as an inside forward or inverted full-back. This is a general football model for fluid formations and utility players, not an Aberdeen exception. Seedocs/career-roles.md.

The player page is a projection of one typed document, with five panels: identity, position inference, playing time, performance and records. Each fact has one home; unknown attributes are omitted rather than repeated as caveats. See docs/player-page-spec.md.

A measured zero stays zero; a missing row is unavailable and a row missing a key field is partial. Parent-club rows are never presented as loan-club output. Provider ratings are labelled as such and do not rank players. Goals conceded while on the pitch, clean sheets (60+ minutes, zero conceded on the pitch) and own goals are reconstructed from fixture lineups and events; a missing fixture is excluded and the coverage is labelled.

Preferred foot is the one profile attribute the team sheet does not contain. It comes from a separate provenance pipeline: Sportmonks when its licensed adapter is enabled, then an official page only if it states the attribute, then Wikipedia as a last resort, then Unknown. Wikipedia values carry a page revision and permalink; no article text is published. The front end never infers a value. Side is derived from the role, not from preferred foot. Player photographs are not published. Current player names link to their profile wherever they appear in structured data; former players link to the former-player register and rumour targets do not link. The 2026-27 rebuild and shape-instability context is recorded indocs/context/aberdeen-2026-27.md.

A rumour is an unconfirmed link with public sources, a lifecycle status and an editorial confidence. It becomes confirmed only when the move is in the transfer register, and then it links to that row. Status is not a probability.

The fan signal samples the first page of recent AFC Chat Rumour Mill topics. Names are extracted with a player gazetteer and labelled by cross-reference: press rumour, club output, recent confirmed transfer or current squad. Sentiment is rule-based VADER on sampled posts, press claims and club titles. It is descriptive and noisy, not a poll.

No raw post bodies, usernames or quotes are published. The forum advertises ai-train=no, use=reference; fitba does not train on the text. The active rumour cohort also produces an advisory recruitment profile, which never overrides the scouting fit. Community votes and submissions are a deferred separate project.

fxG is fitba's own team-match expected-goals estimate. It uses open-play inside/outside-box shots, shots on target, blocked shots and a fixed penalty value, fitted to provider xG. It publishes an 80% model range for the estimate and a separate 80% goals range for the same chances. It has no shot coordinates, distance, angle, pressure, goalkeeper position or game state.

Provider xG is carried separately where available and is never blended into fxG. The model is validated against 2026-27 provider xG; the gate is an in-season calibration, not an out-of-season holdout, and the values are descriptive rather than a performance verdict.

fxG model card → Results and fixtures →

• No raw source payloads, full event feeds or source-shaped tables.
• No raw forum post bodies, usernames or quotes.
• No raw market-value feeds; only bands and our derived scores.
• No club crests or player photographs without a licence.
• No odds in the public app.
• No affiliation with Aberdeen FC, the SPFL or any provider.

A source can be disabled in config/sources.yaml and the site rebuilt from the remaining sources. See docs/sources.md and DATA_LICENSE.md.

Every published number has a registry entry, a declared formula, source layer and coverage tier, and an audit result on Metrics and audit. A number whose mode is provenance_only is present and source-checked, not independently recomputed.