ARENA: AI features that lead to useful actions

ARENA should make a specific promise: understand the evidence, test a decision, take the next step and see what changed afterward. That is more defensible than another predictions tab or general sports chatbot. The recommendation is to build a Decision Flip Lab first, extending the existing captain planner and AI brief rather than buying a new feed.

The findings below use official pages reviewed on 9 September 2026 and a read-only audit of the current application. Positive competitor capabilities are verified advertised features. Proposed ARENA advantages are hypotheses about connected workflows; this review does not establish worldwide novelty, better predictive accuracy or measured demand. Search results from unofficial companion apps and public user comments were excluded.

What competitors already offer

Flashscore provides a further warning against invented ratings: its official explanation says its football grade combines more than 70 statistics and requires at least ten minutes. ARENA's transparent fantasy-points index remains a different measure, not a reconstruction of that system. Rating explanation

Five connected journeys worth testing

1. Decision Flip Lab: “What would change my choice?”

Journey: select two captain candidates → inspect the supplied historical sample → adjust per-fixture minutes → see the exact condition where the comparison changes → ask AI to explain the deciding assumption → save a dated decision receipt → review it after the round.

Proposed distinction: make the sensitivity of a decision usable, rather than ending with “pick A.” This combines a testable condition, source limitations, an explicit choice and a preserved review. Personalised comparison itself is FPL parity; lineup recommendations are Sorare parity. The exact connected sensitivity-and-receipt experience is the hypothesis.

Feasible first version: football Classic only, using the existing planner's per-90 calculation, fixtures, minutes inputs and server-built AI context. Compute thresholds in code; the model explains supplied values. A local decision receipt can test usefulness cheaply, but must be labelled device-local and cannot prove an unaltered pre-lock timestamp. A later server-saved receipt can establish chronology.

Measure: whether users can identify the assumption that changes the choice, scenario-to-save conversion, later receipt review and unsupported AI statements. Do not measure success by one round's winning captain or call the scenario a forecast.

2. Personal Impact Inbox: “This changed; here is what it affects”

Journey: a source update or scoring correction arrives → join it to an owned player, saved lineup, following item or note → show the before/after fact and affected decisions → offer Refresh plan, Review receipt or Prepare a discussion draft → preserve the original evidence.

Proposed distinction: the useful unit is an affected decision, not another headline. Player alerts are already common and ARENA already stores receipt revisions. The missing connection is explaining which personal objects depend on the revised fact and presenting the appropriate next action.

Feasible first version: start with existing ARENA settlement revisions and explicit page refresh. A deterministic dependency map identifies affected objects; AI condenses the consequences. Schedule and availability alerts require reliable fresh sources; the inbox must not turn an older snapshot into a newly reported event. Real push delivery is a later dependency.

Measure: correction comprehension, time to reach the relevant receipt/plan, false impact joins, duplicate notices and scoring support requests. Never automatically swap, sell, trade or post.

3. Scout for This Slot: “Find me an eligible alternative I can actually use”

Journey: tap an empty/replaceable lineup slot → describe constraints such as sport, position, owned-only or credit ceiling → receive a small set of eligible candidates with source/sample limitations → compare their supplied records → preview signing or substitution → explicitly confirm through the existing action.

Proposed distinction: link discovery to a particular playable slot across ARENA's five supported fantasy sports. Budget awareness and recommendations are incumbent capabilities; the hypothesis is a useful cross-league, coverage-aware handoff that removes the current burden of interpreting catalogue memberships.

Feasible first version: reuse canonical player IDs, exact competition/team/position memberships, owned cards, fixed signing price and historical receipts. The server validates eligibility and price. AI translates the request into permitted filters and explains the resulting options; it must not invent a starting probability or “best player” ranking for sparse records. Permit no result when the evidence is insufficient. Catalogue coverage is broader than performance coverage.

Measure: completed valid lineups, invalid-selection attempts avoided, time from slot to confirmed choice and reasons users reject candidates. Test known names against unfamiliar alternatives so the system does not merely repeat popularity.

4. Evidence Debate: “Can this claim survive its strongest counterexample?”

Journey: select a match/player fact → write a short claim → AI separates observation, inference and prediction → display supporting and conflicting supplied evidence plus what is missing → choose whether to revise the claim → publish a reviewed source-linked draft in the existing room.

Proposed distinction: make conversation more useful by helping a fan formulate and test a claim. Adjacent chat, polls and contextual stats are parity. ARENA already has source-linked drafts; the proposed addition is the explicit counterexample and scope check before posting, with a later corrected-source marker.

Feasible first version: constrain analysis to the selected match's normalised facts, not unrestricted web fact-checking. For example, more shots alone cannot establish better chances when xG or shot quality is absent. The model can flag that inference limit; it cannot manufacture the missing measure. Factual correction links require stable source-field/revision references.

Measure: evidence opened by readers, claims revised before posting, unsupported statements caught and useful replies. Votes must never turn an unsupported claim into verified fact. Posting always remains a separate deliberate action.

5. My Twenty Minutes: “Fit the matchday around my time”

Journey: choose a time window, followed teams/players and spoiler preference → build a schedule of matches, lineup deadlines and relevant rooms → explain why each item matters → open one destination → later receive a cutoff-safe catch-up of missed facts.

Proposed distinction: organise scarce attention around the fan's available time and fantasy obligations. Alerts, favourites, schedules and personalised content are parity. A coherent plan that respects time budget, cross-sport deadlines and a spoiler cutoff is the hypothesis.

Feasible first version: planning for known future fixtures, using deterministic clash detection and user priorities. AI supplies a short explanation or translates it. Cached/uncertain kickoffs remain labelled and unscheduled items are not silently assigned times. A “most exciting live game” ranking, live catch-up and broadcasting links require additional fresh event data or media rights. Spoiler filtering must happen before data reaches the model or hidden UI.

Measure: plan completion, avoidable missed deadlines, accepted/rejected suggestions and accidental spoilers. Total screen time is not the objective; successful use within the chosen window is.

Build the Decision Flip Lab now

Place it immediately under the existing captain comparison, not in a separate AI destination. The compact result should say “What changes this comparison?”, show the threshold and its mathematical basis, and expose Try a scenario, Explain, and Save this decision. The current planner already has the needed arithmetic, source status and selected minutes. This is an incremental feature, not a new forecasting system.

An illustrative calculation: A has 4.8 historical ARENA points per 90; B has 4.0 and is assigned 75 minutes. B's linear scenario is 3.33 points. A ties at 62.5 minutes and exceeds it above that level. At 60 minutes A's scenario is 3.20. These are invented demonstration inputs, not current player predictions. When both players already belong to the same XI and play, the extra captain-choice difference is A minus B, not twice that difference.

For doubles, expose each fixture's minutes and hold the other assumptions fixed. Handle zero/negative rates, impossible thresholds, ties, missing samples, unknown schedules and expired deadlines explicitly. A known blank round can contribute zero; an unavailable schedule cannot. A stale source may support a clearly dated arithmetic illustration, but cannot certify current eligibility or a live recommendation.

Keep every number in a deterministic UI field. Send the computed threshold, alternatives, evidence dates and limitations to the existing AI brief route; never ask the model to invent the threshold. Save the selection fingerprint and invalidate the answer when inputs change. Initially, saving the decision need not alter the squad. Any later Apply captain action must revalidate ownership, lineup eligibility and deadline on the server and require an explicit click.

Budget and validation

The first journey needs no additional data subscription if it remains inside the evidence already available. That statement concerns incremental integration cost, not existing data rights. The current AI helper caps requests at 15 per user/day and 200 globally/day, with a 180-word output limit; these are code settings, not measured demand or a quoted operating budget. Its configured connection and source freshness still determine availability.

Use one bounded explanation request after the user completes a scenario, not one per slider movement. Cache only against the complete evidence/input fingerprint and expiry, with private contexts isolated. Keep deterministic explanations usable when AI is unavailable. The price of a multi-sport commercial feed and the right to use its data in AI outputs remain separate supplier questions.

This implementation fits the established principle of using a predictable workflow for a well-defined task and adding autonomous complexity only when it improves measured outcomes. Anthropic's engineering guidance supports that principle, while warning that its older tooling examples have since changed. Engineering guidance

Before wider exposure, test supported and deliberately unanswerable examples. Separate four evaluation questions: Is the calculation correct? Is the explanation supported? Can the user complete the action? Does the group return? A more fluent explanation does not prove a better forecast, and a saved choice that loses does not by itself prove a poor decision.

References and evidence boundaries

The inspected ARENA code includes the planner model/UI, server-built AI contexts, bounded AI helper, athlete/Clubs structures, receipt revisions and match discussion integration. This is not an authenticated competitor UI audit or an ARENA deployment test. No superiority claim is justified until these journeys work reliably and their intended audience repeatedly chooses them.