Start with the evidence, not the sentence
An AI sports commentary platform may combine video, player tracking, scoreboard readings and a licensed event feed. These inputs answer different questions. Video can show movement and contact; tracking can estimate position; a structured feed can label an action. None should silently inherit the authority of the others. A system that receives the word “goal” still needs to know whether that record is provisional, confirmed or subsequently withdrawn.
Before evaluating a platform, ask for an input inventory. Record the source, permitted uses, update behaviour and expected failure modes. A public scoreboard, a broadcast picture and an official competition feed can disagree because they update at different times. A reliable workflow preserves those disagreements long enough for the desk to understand them instead of forcing everything into a single confident label.
Separate detection from interpretation
Detection identifies a possible action: a ball changes direction, a player falls or the scoreboard changes. Interpretation attaches meaning: a clearance, an injury or a lead. These steps should remain distinguishable in the data model.
A computer vision system may miss the ball behind a player or confuse a replay with live action. An event feed may correct the credited player after review. Preserve the original observation, the current interpretation and the reason for a revision. That gives a producer a trail to inspect when a tempting prompt looks wrong.
Follow one moment through the desk
The sequence below is a worked example, not a measured latency claim. Its timestamps describe positions on an illustrative match clock. The important feature is the change in evidential status, not the number of seconds between rows.
- Observe the attempt
Video shows a strike toward goal. Log a candidate shot without deciding its outcome.
- Receive a feed update
A structured record labels a save. Retain its source identifier and provisional status.
- Reconcile the picture
The desk checks whether a deflection or offside decision changes the usable description.
- Offer a sourced line
A prompt names the event and its source. The commentator decides whether it adds anything.
Resolve names before adding narrative
Player identification is not a decorative detail. Squad lists, substitutions, shirt numbers and duplicate surnames all affect whether a line refers to the right person. Store stable identifiers separately from display names. Keep a pronunciation field separate again, because the spelling required by a data provider is not always a useful guide for speech.
Identity confidence should affect output behaviour. If a wide camera cannot distinguish two players, the system can offer a team-level description or wait for another source. It should not use a familiar name merely because that player usually occupies the relevant position. A plausible identity mistake can survive several later processing stages and become a convincing but false statistic.
Make the sport's rules explicit
Football analysis needs to distinguish a visible incident from a decision under the Laws of the Game maintained by The IFAB. FIFA and UEFA competitions also bring their own operational contexts. Those names do not make a third-party event feed official; the contract and source description determine what the feed represents.
The same caution applies elsewhere. The NBA and FIBA do not provide an interchangeable rules context for every basketball event. Cricket workflows must account for the Laws of Cricket maintained by MCC and the relevant ICC playing conditions. Build competition and ruleset metadata into the workflow rather than asking a language model to guess from the team names.
Keep time in more than one column
Store event time, receipt time and any subsequent correction time separately. A match clock may stop, count down or include added time. A production timecode is a different reference. SMPTE timecode conventions can help production systems align media, while an interchange timestamp such as the IETF's RFC 3339 describes a date and time. Neither automatically resolves the sporting meaning of a clock reading.
This separation also helps explain duplicates. A late feed packet can describe an earlier action without being a new action. Event identifiers, version numbers and correction links are more dependable than treating every arriving message as fresh match information. If the feed reconnects, the desk needs reconciliation, not a burst of repeated commentary prompts.
Desk caveat: an event label is not an explanation
A “turnover” record can support a description of possession changing. It does not establish that a player panicked, ignored instructions or caused a tactical collapse. Those are additional interpretations requiring evidence and editorial care.
Test the awkward footage
A demonstration built from clean, uninterrupted play says little about difficult coverage. Test rain, occlusion, crowded penalty areas, replays, graphics changes and missing roster updates. Include examples where the correct answer is “unknown.” Measure false positives and missed events separately, and inspect performance by competition, camera setup and event type rather than relying on one aggregate score.
The output of this stage should be a traceable event candidate: what was observed, what was inferred, where the evidence came from and what remains unresolved. Only then is it useful to ask how quickly a commentary assistant can turn that candidate into something worth showing the desk.
Next: work against the clock