Official links

• The hackathonMicrosoft Innovation Studio: the challenge, the dates and the entries
• Official rulesMicrosoft's rules for the hackathon, on GitHub
• Hackathon repositoryGuides, resource playlists, the code of conduct
• Microsoft Reactor seriesThe hackathon's live sessions for builders
• The builders' session"Build AI-powered soccer insights apps with AKS and ACA", October 14

1. Grand prize: first place

   Explainable match intelligence, every line with its evidence. Engaging narratives: live commentary, the recap and the ref's call. Ask the match anything, and get a checked answer. Context: momentum, breaks, the press, a card for any player. Accessible and personal: English, Urdu or Arabic, read aloud for fans who can't watch the screen, with questions asked out loud and answered aloud for fans who can't type them, casual or analyst, your club, your player, one number. And HUD², the same match on one heads-up screen through the referee's, the commentator's or the fan's lens, every checked line pinned where it happened. All in one match, end to end.

   Built to win

2. Grand prize: second place

   The same match and the same evidence. Whichever place the judges give, it's the whole build that stands behind it.

   Built to win

3. Best use of Microsoft Foundry

   Models on Foundry write the commentary from facts code computed, and translate it into Urdu and Arabic under the same law. Asked a question, the model calls tools for every number it states, and the Critic checks each one. Azure AI Speech, on the same Foundry resource, reads every verified line aloud and hears questions asked out loud, in English, Urdu or Arabic. A backup model takes over when the first fails. Evaluation chose which model serves, by a rule fixed before the run, and Foundry's own evaluators judge the lines that aired (groundedness 4.84 of 5) and the answers to viewers' questions (5.0 of 5, after the fix the first run called for). Every model call is traced with its tokens and its cost, in Application Insights and in Foundry's own tracing view. Keys are off: Entra ID only.

   Built to win

4. Best multi-agent orchestration

   Five agents on Microsoft Agent Framework with typed handoffs and shared match state. The Critic sends a line back with its reasons; take the narrator off air mid-match and the broadcast keeps going, and says so. Asked a question, the Narrator answers with the Scout's counting tools and the Critic checks the answer: three agents on one question. Other agents can reach the Critic as a peer over A2A, and assistants over MCP. The Critic is also an MCP server any assistant can call.

   Built to win

5. Best enterprise solution

   Private by default: Cloudflare Access in front, and the studio checks Cloudflare's signature on every request. Prompt Shields screen every viewer's question before a model reads it. An audit trail on Azure Cosmos DB for every claim, refusal, cue and answer, kept 90 days and then deleted by policy, a receipt on every line, traces and a workbook in Application Insights, a daily model budget under a $100 ceiling, and a 38-point health check before every recording.

   Built to win

6. Best Azure cloud-native integration

   One image on Azure Container Apps, signing in with its own managed identity and no keys, watched by liveness and readiness probes, with Microsoft Foundry, Cosmos DB's free tier, Application Insights and a registry with admin off around it. Every agent step is traced, and a read-only doctor checks the whole of it.

   Built to win

Eight days, six phases

1. Oct 17–19GroundworkThe field studied, the name chosen, a synthetic match generator calibrated over 100 seeds.
2. Oct 18–20SkeletonAll five agents run a match end to end with offline backends, tested on every push.
3. Oct 18–23Core buildLive detection, the Critic's full rulebook, the overlay cue stream, analyst and casual modes.
4. Oct 18–24DifferentiatorsThe on-camera recovery demo, the cost meter, Urdu and Arabic, personal modes, the recap, breaks, the high press, the ref's call, ask the match, HUD², the ledger, the program feed and a realism check, all live.
5. Oct 20–24Ship kitThe 101-second judges' cut and a four-minute walkthrough on YouTube, the README, the media kit, the repo made public.
6. Oct 24SubmittedThree days before the deadline, never on deadline day.

I promised to show the core loop end to end by October 24, or narrow this declaration in public. It ran end to end, live on Azure, on October 18, and nothing has been narrowed since.

The challenge's full brief — the five stages, the hard requirements, the judging pillars and every reward with its amounts — is on the challenge page.

Prize pool

$50,000

Prize lanes

6

Judging pillars

5 × 20%

Build window

Oct 6–27

The ask, in the organisers' own words

“Build an AI-powered solution that transforms synthetic, football-realistic data into explainable match intelligence, real-time narratives, automated recaps, and personalized experiences — for fans, studios, broadcasters, and streaming platforms.”

Four things to build — intelligence, narratives, recaps, personal experiences — for four audiences. And one law underneath them, also from the brief:

“A project that only generates metrics has no audience, and a solution that only personalizes has nothing to personalize.”

The brief's spine: one pipeline, five stages

1. Ingest the match's events as they happen: passes, shots, tackles, pressures, possession changes, match metadata.
2. Interpret them into football statistics, patterns and context.
3. Explain why a moment matters, not only that it happened: control or chaos, tactical pressure, game rhythm.
4. Render insight that sits on screen beside the match: timed, machine-readable, ready for synchronized overlays on a live feed.
5. Personalize — the analyst and the casual fan get two genuinely different experiences: custom modes, favourite clubs and players, player focus, one metric to follow.

What a submission must do

How it's judged

Two stages: a pass/fail screen — does it fit the theme and apply the required technologies — then five pillars, each worth a fifth of the score.

1. 20%Technological implementation

   Quality practices, effective use of the hero technologies, well-structured and maintainable code.

2. 20%Agentic design & innovation

   Creative agent patterns, sophisticated orchestration and multi-agent collaboration, novel implementations.

3. 20%Real-world impact

   A significant problem, production readiness, impact on developers, businesses and users.

4. 20%User experience

   Intuitive to use, clear value in the demo, a balanced frontend and backend.

5. 20%Adherence to category

   Compliance with the chosen prize category's description.

The calendar

1. Sep 29Registration opensOn Microsoft's Innovation Studio.
2. Oct 6Hacking beginsSubmissions open the same day.
3. Oct 20Registration closes12:00 PM Pacific.
4. Oct 27Submissions close11:59 PM Pacific; judging begins 9:00 AM. This entry went in on October 24, Karachi time, three days early.
5. Oct 27–Nov 10JudgingThe live demo must stay up and free to use (§4.8).
6. Nov 20Winners announcedThe top six projects are notified by November 10.

The rewards

A $50,000 pool across six lanes. Cash ships as e-gift cards (§8.3), split equally between a team's members; one winning team a lane; a project may take at most one grand prize and one category prize (§4.2, §8). Amounts as published in the official rules and awards register, October 2026.

1. Grand prize · first place

   ~$4,000 cash, per member

   • a digital ticket to a Premier League match (~$1,000)
   • a $50 club-store voucher
   • a feature in social promotion

   For the most complete and compelling solution: innovative use of the Microsoft AI stack, explainable intelligence, engaging narratives, contextual insight, accessible and personal experiences.

2. Grand prize · second place

   ~$2,500 cash, per member

   • a digital ticket to a Premier League match (~$1,000)
   • a $50 club-store voucher
   • a feature in social promotion

   Judged against the same bar as first place: the whole build stands behind whichever place the judges give.

3. Best use of Microsoft Foundry

   ~$1,500 cash, per member

   • a $50 club-store voucher
   • a feature in social promotion

   For the most effective leverage of Foundry: development, orchestration and management of AI-powered solutions — scalable AI “difficult to achieve through traditional approaches alone.”

4. Best multi-agent orchestration

   ~$1,500 cash, per member

   • a $50 club-store voucher
   • a feature in social promotion

   For purposeful multi-agent collaboration: distinct roles, shared state, effective handoffs, recovery from failures — an end-to-end outcome no single agent could achieve.

5. Best enterprise solution

   ~$1,500 cash, per member

   • a $50 club-store voucher
   • a feature in social promotion

   For enterprise-deployment readiness and long-term business impact: a clear path to adoption, reliability, transparency, production controls.

6. Best Azure cloud-native integration

   ~$1,500 cash, per member

   • a $50 club-store voucher
   • a feature in social promotion

   For Azure's cloud-native services — Container Apps, AKS, Functions, event-driven and resilient by design — with operational excellence.

This entry is a solo build: whatever it wins lands undivided. Which lanes it aims at, feature by feature, is on the goal page.

How this entry answers

• Ingest

  The event stream contract: a committed JSON Schema, and the seeded synthetic generator that makes the football-realistic data itself — the brief's data-innovation pillar, taken literally.

• Interpret

  The Scout: deterministic counting code on the hot path, no model, sub-millisecond — goals, big chances, great passes, breaks, the high press, momentum, the ref's call.

• Explain

  The Narrator writes from a fact sheet; the Critic recomputes every number from the events and refuses what it can't prove. Why a moment matters, with the evidence on screen.

• Render

  Timed, prioritised JSON cues with evidence, receipt and player tags — a JSON Schema contract a broadcast graphics system can build against — and HUD²'s on-screen lenses.

• Personalize

  HUD²'s three lenses, Ask the match, a club, a player or one number to follow, casual or analyst — in English, Urdu and Arabic, spoken as well as written.

• And recap

  At full time the match writes its own recap, every line through the Critic first, as a page that prints to PDF.

The official record

• The hackathonMicrosoft Innovation Studio: the challenge, the dates, the gallery of entries
• Official rulesEvery requirement and prize clause cited on this page
• Hackathon repositoryGuides, resource playlists, the code of conduct
• Microsoft Reactor seriesThe hackathon's live sessions for builders

is an independent entry and isn't affiliated with or endorsed by Microsoft or the Premier League. Every club, player and match on this site is invented, as the rules require. Prize figures are quoted from the official rules and awards register; winners are announced November 20, 2026. Every ask on this page, delivered and linked to its evidence: the deliverables report.

Asks answered

8 / 8

Hard requirements

7 / 7

Tests on every push

274

Demo video

101 s

1. Delivered

   Explainable match intelligence

   The ask: not “AI says X dominated”, but why — with event evidence.

   Delivered

   The Scout flags the moments that matter — goals, big chances, great passes, counterattacks, a high press paying off, momentum swings, control or chaos, the ref's call — and every line that airs carries the events it stands on, each number recomputed by the Critic before it airs. How a claim gets on air.

   How

   Code counts, models write: deterministic detection on a sub-millisecond hot path; the Narrator drafts from a fact sheet on Foundry; the Critic recomputes every number from the raw events and refuses what it can't prove, on screen, with the reason.

   Next

   The decision explainer — the referee's call explained with the Law it rests on — then perception: fouls read from camera angles, every prediction citing its frames. The road.

2. Delivered

   Real-time narratives

   The ask: live during the match, not only after it.

   Delivered

   A live cue stream, start to finish: alert to air in about 1.3 seconds, every cue timed and prioritised. When the Narrator was knocked out mid-match in tests, the broadcast dropped to the Scout's cited facts, said so on air, and announced the recovery — the feed never went dark. When an agent fails.

   How

   The hot path sees every event and never waits on a model; each alert runs Narrator → Critic → Localizer as a Microsoft Agent Framework workflow with a rewrite edge, supervised by an Orchestrator that degrades and recovers by design.

   Next

   The same cue contract behind a broadcaster's own graphics system, and multi-match coverage from one deployment.

3. Delivered

   Automated recaps

   The ask: the 90-minute story writes itself.

   Delivered

   At full time the match writes its own recap — the result, every goal, both sides' numbers, the moments that aired — every line through the Critic first, whole-match numbers counted to the final whistle and cited. Reads as a page, prints to PDF: seed 21's recap.

   How

   One rule lifted at full time — the checker may look back over the whole match — with the same evidence law as the live lines. Standout players are counted from their own events and marked counted, not checked: said plainly. The limits.

   Next

   Richer recaps — a momentum timeline, player cards — and a recap assembled per viewer, in their language and their club's colours.

4. Delivered

   Personalized experiences

   The ask: the analyst and the casual fan as two genuinely different experiences — custom modes, favourite clubs and players, player focus, one metric.

   Delivered

   HUD²'s three lenses — referee, commentator, fan — Ask the match, a club, a player or one number to follow, casual or analyst modes. Emphasis changes; the facts never do.

   How

   Personalisation is a layer over verified cues: the viewer's choice reorders and colours what airs, and the Critic checks every viewer's line the same way it checks the broadcast's.

   Next

   HUD² worn — the same lenses on glasses, with the marks on the real pitch — as the hardware allows.

5. Delivered

   Multi-language storytelling

   The ask: insight for fans around the world, in multiple languages.

   Delivered

   Every verified line airs in English, Urdu or Arabic — on screen and spoken aloud by Azure AI Speech. A question can be asked out loud in any of the three, and the answer comes back checked, in the same language.

   How

   The Localizer translates under the same law: no number the checked line doesn't state, every name kept as written. A translation that breaks either airs the English instead.

   Next

   More languages ride the same rail — the law doesn't change, only the voice.

6. Delivered

   Data innovation

   The ask: synthetic, football-realistic data only (rules §4.7) — and the brief invites creating and extending it.

   Delivered

   A seeded synthetic match generator — invented clubs and players, real-match volumes — with a committed JSON Schema stream contract and hidden ground truth for measuring the Scout. A published realism check sets its scorelines against 7,559 real internationals. The same seed replays the same match, so every number on this site is reproducible. How real is it.

   How

   Calibration over 100 seeds; realism scored against a public-domain record of real results, used only as a benchmark; the generator's limits stated plainly.

   Next

   Richer event types and playing styles in the generator, and the audit trail mirrored into Microsoft Fabric's OneLake for SQL and Power BI — weighed for this build and deferred to stay inside the free tier. The build log.

7. Delivered

   Ask the match anything

   The ask, in the grand-prize lane's own words: ask the match anything, and get a checked answer.

   Delivered

   Any question, live or after full time, in three languages, typed or spoken: the model must call counting tools for every number it states, the Critic recomputes each one before the answer shows, and every answer is sealed into the match ledger with a proof the browser checks. What the events can't answer gets fixed words — including a fixed steer away from betting.

   How

   Function calling on Foundry; each count returns as a fact with an id, the answer names its facts, code builds the claim. Measured on unseen matches, Foundry's judges scored the answers 5.0 of 5 after the first run called for one fix.

   Next

   Questions answered over a whole season of ledgers, each answer cited to the exact block that sealed it.

8. Delivered

   Recovery from failure

   The ask, from the multi-agent lane: recovery from failures — an end-to-end outcome no single agent could achieve.

   Delivered

   Five agents, distinct roles, shared match state. A test knocks the Narrator out four times in a row mid-match and the cues never stop; the studio's off-air switch does it on camera. The Critic also serves other assistants over MCP and other agents over A2A.

   How

   The Orchestrator retries once, falls back to detection-only cues marked unverified, and announces recovery on air. The Critic itself runs no model, so its verdicts are computed and can't be talked round.

   Next

   The Critic as a shared court: the same checks opened to other broadcasters' agents, as a peer.

The hard requirements, met

1. Met
   Synthetic data only (§4.7). Every club, player and match is generated by this repo's own seeded generator; no real match data anywhere. See the data.
2. Met
   New code, new repository. Built October 18–27, 2026, for this entry alone; patterns carried over, no old repos.
3. Met
   Demo video under 2:00. Recorded at 101 seconds, at a public URL, showing the project functioning — no third-party trademarks, no licensed music.
4. Met
   Public repository at submission. Private until the showcase is final; one config flag flips every Code link on this site.
5. Met
   Free judge access through November 10 (§4.8). The live studio runs behind Cloudflare Access, and any @microsoft.com or @premierleague.com address signs in with a one-time code, as the testing instructions say — and make studio-offline runs the same studio on any machine, no Azure needed at all.
6. Met
   Installs and runs consistently (§4.3). One make setup; 278 tests on every push; a 38-point doctor and a smoke script that streams a whole match and checks every part.
7. Met
   The entrant's own work, in English (§4.9). A solo build; the pitch, video and test instructions in English — Urdu and Arabic are outputs, on purpose.

The hero technologies, in the build

• GitHub + Copilot agent mode

  In the build daily, working from the repository's own evidence rules, so it follows the same law the agents do.

• Microsoft Foundry + Azure AI

  gpt-4.1-mini writes the lines, gpt-5-mini backs it up; Foundry's evaluators judged what aired (groundedness 4.84 of 5) and every call is traced with its tokens and its cost.

• Agent Framework, A2A + Azure MCP

  The warm path is an Agent Framework workflow; the Critic serves assistants over MCP (six read-only tools) and other agents over A2A, with no model inside.

• Azure data and analytics

  The audit trail on Cosmos DB's free tier, every cue and refusal kept 90 days; Microsoft Fabric weighed for the analytics and deferred — recorded above, in Data innovation's next.

• Azure cloud-native

  One image on Container Apps with a managed identity and no keys; Application Insights traces every agent step; liveness and readiness probes watch it; a read-only doctor checks the whole of it.

Every row above names its evidence. The road from here — the decision explainer, perception, the fourth official — is written down, gate by gate.

One claim, followed all the way

The claim from the front page, through each step of synthetic match seed 7.

1. Spot

   The Scout reads every event as it arrives and flags goals, big chances, great passes, breaks from a ball won in a side's own half, a high press paying off, swings in momentum, and fouls with a card or a penalty in them.

2. Write

   The Narrator drafts the line for air from a sheet of facts code computed, and cites the event behind every one. On Azure, a model in Microsoft Foundry writes it: gpt-4.1-mini, chosen by evaluation.

3. Check

   The Critic recomputes each number from the raw events, with the same formulas the Scout uses. A wrong number, a missing citation or an event from the future sends the claim back, up to three attempts.

4. Air

   The Localizer turns the verified claim into an on-screen cue for casual fans or analysts, in English, Urdu or Arabic, with its player tags and a receipt of what was checked and what it cost. A claim that never passes never airs.

5. Make it yours

   A viewer picks a club, a player, or one number to follow. Their moments come first and carry their club's colour; the other side's still air, marked as the rival's. Following one player keeps only that player's moments, counts the rest, and adds a live card counted from the player's own events. The models tell the line from the viewer's side. The facts don't move: every line still passed the Critic.

   56'GOAL · SAL 2–2 BRK

   Your clubYour player

   Bilal Arholm scores for Saltmarsh Rovers from 10 m, a speculative chance (xG 0.11). Set up by Adem Rosmere's 22 m pass, difficulty 58 of 100.

   Synthetic match seed 21, with Saltmarsh Rovers as the viewer's club and Adem Rosmere as their player.

6. Recap

   At full time the match writes its own recap: the result, every goal, the numbers for both sides, the moments that aired and the players who made them. Each line it writes goes to the Critic first, with one rule lifted: it may look back over the whole match. A whole-match number must be counted to the final whistle, and the line cites it.

   SAL 2–2 BRK, the numbers

   Shots: SAL 19, BRK 10. Expected goals: SAL 2.82, BRK 1.64.

   Checked by the Critic, citing all 29 shots and the final whistle.

   Read the whole recap

7. Ask the match

   A viewer can ask the match anything, live or after full time. The Narrator answers, but it learns a number only by calling the Scout's tools: code that counts the events so far. Each count comes back as a fact with an id, the answer names the facts it uses, and code builds the claim from them. The Critic then checks it like any line: every count recomputed, every number and every player it names backed by those facts. What the events can't answer, like an injury or the final score, gets plain fixed words, not a guess. With the voice on, ask out loud in English, Urdu or Arabic: Azure AI Speech turns the question into words, and the checked answer is read back in the same language.

   70' · You asked: Which team has pressed higher in the second half?

   Saltmarsh Rovers have pressed higher in the second half with 18 high presses compared to Brackenfield's 5.

   Answered by gpt-4.1-mini on Microsoft Foundry from two counts, both recomputed by the Critic before it showed. Synthetic match seed 21.

Plug it into a broadcast

Everything the court produces is built for someone else's system to use, not only this studio.

• A cue stream with a contract

  Each line leaves as a timed, prioritised JSON cue with its evidence, its receipt, the players it tags and, for a foul, the card to draw. A JSON Schema, written from the code and checked on every push, is the contract a graphics system builds against.

• The Critic, for any assistant or agent

  Six read-only MCP tools, so an assistant can read a match and put a claim to the same Critic before it says anything. And an A2A agent with a card, so another agent can hand it a claim or a question as a peer. Same rules either way.

• An audit trail

  Every cue, refusal, answer and change of service level, kept for 90 days on Azure Cosmos DB: what aired, who wrote it, what was refused and why.

• The cost of a line, known

  A receipt on every cue: the model, its tokens and its cost. A model-written line costs about $0.0007, a whole match under a cent.

• The recap, as a page

  One standalone page per match that prints to PDF, every line it wrote checked by the Critic.

• Your rules, your languages

  The competition picks how strictly fouls are read. English, Urdu and Arabic, on screen and spoken aloud.

1. Full serviceEvery cue verified by the Critic
2. Detection onlyCues keep airing, marked unverified
3. RecoveredAnnounced on air, verification resumes

A test knocks the Narrator out four times in a row mid-match and checks that the cues never stop. In the studio a switch takes the narrator off air mid-match and puts it back, and the demo video does it on camera.

Each scene moves while it's on screen, and two of them break on purpose, at your button. Prefer less motion? The site honours your reduced-motion setting: every scene stands still, complete.

The referee The commentator The fan

Play the recorded match

A match recorded from the live studio on October 20, replayed by the studio's own code. The lines, the refusals and the ledger are real; the studio itself is open to the judges only.

• The referee

  Every foul, with its distance from the goal the fouling side defends; the pressing; and the zones a likely call is read against at the competition's strictness: the penalty area, the goal-chance zone and the line of a promising attack. This lens shows the least. A referee's eyes belong on the players.

• The commentator

  Who's on the ball, by number and name, the passes and the shots as they happen, and the last five minutes against the match so far, both sides on one bar. Every number is counted by the code the Critic checks with.

• The fan

  Who's on top, where each side has played lately, the chances so far, and plain words for anyone new to football. Any lens can ask the match a question.

Every lens, as it looks

Where it goes

1. Today, a screen. Built, and live in the studio.
2. Next, glasses. The same lenses, worn, with the marks on the real pitch. That needs cameras that see where the ball and every player are.
3. Then, contact lenses. As the hardware arrives.
4. Further out, neural interfaces. Only with consent, and only once they're proven safe.

On any device the law stays the same: a word passed the Critic, and a mark is an event that happened. The whole road.

Ask the recorded match

The recorded match answers the ten questions asked of it in the live studio, with their real answers and proofs; tap one, then verify it. The live studio answers anything.

How an answer stays honest

1. Screened first. Prompt Shields (Azure AI Content Safety) read the question before any model does. Code holds back what the events can't answer.
2. Counted by code. The model gets numbers only by calling counting tools. Each result is a fact with an id.
3. Checked by the Critic. Every count is recomputed, every number and every player name checked. A draft that fails goes back with the reasons.
4. Sealed. The answer goes into the match ledger. Its proof checks in your browser.

Measured on matches it had never seen

What makes it hold

• Every entry hashed. The match itself, every event, every line that aired, every claim the Critic refused, every answer, and the recap. SHA-256, over one canonical form.
• A block a minute. Each minute's entries are sealed by a Merkle root, built the way Certificate Transparency builds its logs (RFC 6962).
• Each block names the one before. Change an entry and its block's root changes, so its hash, so every block after it.
• One seal. The last block's hash stands for the whole match. Copies are kept apart from the ledger, in the audit trail on Azure Cosmos DB and in the trace.
• Anyone can check. In a browser, or offline with python -m court verify, which also generates the synthetic match again from its seed: the ledger must hold exactly those events.
• One answer, a few hundred bytes. Each answer's proof is its Merkle path and the blocks on to the seal (RFC 9162). The Ask page checks it in the browser.
• No coin, no mining, no betting. Nothing to buy, nothing to trade. It's there to keep the record honest.

New to hash chains and Merkle roots? Watch them explained in eight animated scenes — including what happens when you change one entry on purpose.

What it proves, and what it doesn't

It proves the history hasn't changed since it was sealed, and that the events are the ones the seed generates. It adds nothing a line didn't already prove, and a studio that lied while recording would seal its lie: the copies of the seal kept elsewhere make that harder, not impossible. Publishing each match's seal to a public log is the next step.

Play the program feed

A match recorded from the live studio, replayed by the studio's own stream. Play the take, or step line by line with Next line — the panel on the right is the JSON your system receives for every cue, live.

Both lanes of the brief's law, served: the metrics have an audience — this feed — and the personalisation has something to personalise — the same checked lines.

• It doesn't predict, and it's not for betting

  only tells you what has already happened, from the match events. It never predicts a result, a score or the next goal, never shows odds, and won't help anyone bet: ask it about bets or odds, in English, Urdu or Arabic, and it says so in fixed words before any model reads the question. Enjoy the football, not the odds. If betting is causing harm to you or someone close to you, free and confidential support is there to help.

• Momentum calls are sure, not complete

  On matches never used for tuning, 62% of the Scout's momentum calls match a real swing, and it catches about a quarter of the swings there are. It's tuned to say less and be right, so when it stays quiet a swing may still be under way.

• The football is synthetic

  Every club, player and match is invented by code. The generator matches real volumes over 100 seeds (about three goals, 29 shots, 78.5% of passes completed and 320 pressures a match), but no real team plays in it, so nothing here says anything about real clubs or players.

• Events, not video

  The court sees what the event stream records: passes, shots, tackles, pressures and changes of possession, each with a place and a time. It doesn't see runs off the ball, a team's shape between events, or anything only a camera would. Speed and distance tags come from the events' own fields: a shot's speed, a pass's length.

• Its own xG

  Expected goals here come from a small, open formula (distance, angle, headers, pressure), the same one the Critic recomputes with. It's consistent, not a trained provider's model: its numbers aren't comparable with a broadcaster's.

• True isn't the whole story

  The Critic proves that a line's numbers and citations hold. It doesn't judge emphasis or fairness, so a verified line can still be one-sided or dull. Checked means supported by the events it cites, and nothing more.

• Translations are checked for numbers and names

  An Urdu or Arabic line may state no number the English line doesn't, and must keep every name. The wording itself is the model's, reviewed by a person, not by the Critic.

• Standouts are counted, not rated

  The recap's players who made it happen are counted from their own events. There's no rating behind them and no per-player rule for the Critic yet, so the recap marks those lines as counted, not checked.

• The ref's call is a reading, not a referee

  It sees where a foul happened and what it stopped, never the force or the speed of the challenge. It states a likely decision at a stated strictness. The restart has ground truth (the generator awards penalties the same way); the card has none yet.

• The voice reads names as best it can

  Names stay in Latin letters in every language, so the Critic can check them; an Urdu or Arabic voice may say an invented name oddly. The words it speaks are exactly the line that aired.

• An answer's judgment words are the model's

  When you ask the match, every number, name and citation in the answer is checked. Words like "dominated" or "the better side" are the model's reading of those checked counts, not a verdict the Critic can prove. Ask for the numbers and judge for yourself.

• Making it yours changes emphasis only

  A favourite club or player changes what comes first and what's marked as yours. The facts are the same for every viewer.

“We will see whether you have told the truth.”

1. Nothing unverified goes to air

   Broadcasters answer for every word on screen. Each claim carries its own audit trail: the events behind it, the check it passed and the attempt it took.

2. The model writes, the evidence decides

   I don't ask the model to be careful. I check its work with the same formulas the Scout uses, and refuse anything that doesn't add up.

3. Commentary in the fan's own language

   Urdu and Arabic are live beside English. The verified line is translated under the same law: no number it doesn't state, no name changed. Every fixed word was reviewed and signed off before it shipped.

4. Failure is designed in

   Models will be slow, rate-limited or down mid-match. The system degrades, says so on air and recovers, instead of falling silent.

5. Built in the open

   The goals are declared on this site before the deadline, and the code goes public with the entry. Judge the work against the promise.

1. In progress

   The court of claims

   The hackathon, October 2026

   Commentary that proves every claim. Five agents, a Critic that refuses what it can't verify, and a broadcast that survives failure.

   Bridge, live: the “Ref's call”. The Scout flags a foul with a card or a penalty in it, the Narrator states the likely decision, and the Critic recomputes it at the strictness the competition chose: lenient, standard or strict.

2. Next

   The decision explainer

   Late 2026 into 2027

   When the referee makes a call, explains it to the crowd and the viewers, with the evidence and the Law it rests on, in their language.

   Gate: agreement with ground-truth decisions measured over many seeds, and explanations signed off by a referee panel.

3. Planned

   Perception

   2027

   The Scout gets eyes: fouls classified from multiple camera angles and from player tracking, with every prediction citing its frames.

   Gate: matching published results on a public benchmark, with evidence attached to every call.

4. North star

   The fourth official

   Beyond

   Real-time advice to the referee team, faster than a video review takes today. Flashpoint predictions, a published strictness spectrum, and announcements the whole stadium understands.

   Path: simulation and esports first, then grassroots football, then professional trials inside the game's own rules, and the biggest tournaments once each step has proven itself.

HUD²: the match, seen your way

Everything above has to reach people's eyes. HUD² is how: one match, and each viewer's own lens on it, shaped to their needs and their language.

• Today, a screen. Live now: one central display with overlays, markings and annotations, through the referee's, the commentator's or the fan's lens.
• Next, glasses. The same lenses, worn, with the marks on the real pitch. That takes the Scout's eyes: where the ball and every player are.
• Then, contact lenses. As the hardware arrives.
• Further out, neural interfaces. Only with consent, and only once they are proven safe.
• What never changes. On any device, a word passed the Critic and a mark is an event that happened. The referee's lens shows the least, and never decides.

What holds at every stage

• Evidence before verdict. No evidence, no call.
• The human decides. advises and explains, never overrules.
• Strictness is a setting. The same incident gets the same call.
• The Laws come first. Encoded as rules where they can be; judgment is flagged as judgment.
• Explain to everyone. Referee, broadcast and crowd, each at their own depth.
• Never go dark. Degrade, announce, recover.
• Add, never take away. The flow, the drama and the referee stay. It earns its place by adding understanding.
• Never a betting tool. No predictions, no odds, no help to bet. It discourages gambling.

Find me

• matx104.com.pk
• GitHub
• LinkedIn
• Upwork
• X
• Medium
• Credly
• Everything else

Other things I've built

• SOCIRISAn AI-powered security intelligence platform: event streams in, severity-ranked alerts out.
• The Sovereign CourtTen minds, two wings, one throne: a multi-agent council with adversarial review.
• AETHERIXAn interplanetary delay-tolerant networking platform.
• OMNI-GRCAI-powered governance, risk and compliance in one platform.
• ARENAA modular live game platform: real-time rooms, a host console and live scoring.
• Phish-WizAn open-source human-risk platform.
• NAFA-BOXA study library: searchable transcripts synced to video.
• QutubThe pole you navigate by: a life-journey OS where progress needs proof.

The proof, for everyone

A read-only check of the whole environment, run from the project's own command line (make doctor). Identifying details are left out.

Try it now — nothing to install

These run in your browser, on a match recorded from the live studio. The lines, the refusals and the ledger are real; the studio itself is open to the judges only.

Run the full studio

The same studio, on your machine, no Azure account needed:

git clone https://github.com/matx104/HUDHUD.git
cd HUDHUD
make setup
make studio-offline

Then open http://localhost:8091. Pick a match, kick off, take the narrator off air and watch the broadcast recover. With Azure: make studio runs it narrated by the Foundry models.

The evidence

• The deliverables report

  Every ask of the challenge, met and linked to its evidence. Read it or .

• The ledger, verified

  1,815 events sealed into 111 blocks chained by hashes. Your browser recomputes every hash. Try it.

• The audit trail

  Every cue, refusal and answer kept on Azure Cosmos DB, 90 days. The control room on Azure shows the matches, the alerts, the costs.

• The tests

  278 tests on every push: the generator's realism, the Critic's rulebook, the workflow loop, the site and the studio. make test.

• The media kit

  Both videos, the screenshots, the brand sheet, the deliverables report as PDF. .

• The code

  Public on GitHub at submission. Every agent, every test, every script. Source on GitHub.