Snapshot of 17 Nov 2026, 18:00 Vienna time, proof of concept, version 2 (same data as version 1; what changed).

What changed from v1

Version 2 is the same frozen snapshot as version 1, rebuilt with the parts of the developer brief that fit a proof of concept. Any difference between the two links is these changes, not new data.

Adopted

  1. Six-check briefingbrief §8

    Each AI write-up answers why it matters, rivalry and recent history, current situation, who to watch, recent head-to-head and expected outcome, in 50 to 200 words. A check with no fact behind it says "No previous meeting in our sources" or "Not applicable" and is never filled in. The outcome is favoured, close contest or upset potential, set from the odds before the model writes, never odds as advice.

  2. Claim-level fact checkbrief §8, §14

    v1 checked that each cited fact exists. v2 also checks that every number, name, place and day is in a fact the sentence cites, that each number sits next to the side its fact is about, and then a second AI pass reads each check against only its cited facts. A write-up that fails twice is not published. This run: 21 events sent, 6 passed first time, 13 on the second draft, 2 not published; the checks caught a grounding problem in 13 of 21.

  3. Audit record per eventbrief §10

    Every briefing stores the model, the prompt version, the fact pack exactly as sent, the sources with when they were fetched, both checks' results for each draft, and two hashes. Each event page shows it under "Sources and checks". This is the source trail and audit log, without a CMS.

  4. Watchability, kept apart from importancebrief §7, §15, §22

    Its own component (10% of the base, the other five scaled to make room) and its own badge on every card: local start time over the event's length, and whether a viewing source is known. A 03:00 start reads "Middle of the night". The weights are a choice, not a finding, and stay in one config file.

  5. Hidden gem and varietybrief §7, §22

    One "Hidden gem" slot per day: the strongest event in a sport that is not big in your region, that you can watch at a sane hour. And no sport takes more than 3 of the top 10 tonight.

  6. Where to watch, only where a source saysbrief §11

    ESPN's schedule names US providers for 119 events; those show for US viewers. For Austria, the UK and the Netherlands no free source names a provider, so the page says "Viewing not verified" and nothing else: no guessed links, never an unofficial stream.

  7. Acceptance testsbrief §14

    Vienna and London times across a clock change; a postponed or cancelled event drops out of the lists and its page says so; no briefing is published without passing both checks; no viewing line for a region without a verified source.

Does the fact check bite?

A passing run proves nothing on its own, so each published briefing was corrupted on purpose and run through the checks again.

CorruptionCaught
One number changed to one no fact holds19 of 19
A number given to the other side (the fact is real, the team is wrong)15 of 15 (2 by the AI check)

Not yet, and why

From the briefWhy not now
Accounts, saved events, reminders, newsletterThey need a product with users. After the proof of concept shows the ranking is worth opening daily.
Human review of every eventThe aim is as much automation as possible: v2 reviews by exception (the fact checks, the audit record, write-ups that fail are not published). Whether the brand wants editors is an open question.
Full CMSThe audit record and the config files cover a proof of concept. A CMS comes with editors and a review queue.
Sportsbook odds comparison and affiliate linksNeeds legal advice per country and licensed feeds; the brief itself puts it after month 12. Odds stay a hidden signal, as in v1.
Affiliate and outbound-click trackingNo programme exists yet. Add it with the first real partner.
Creator pages, SEO landing pages, B2B widgetGrowth work, for once the product is worth growing.
Insider significance as its own score dimensionNeeds a source that is not popularity. Curated lists could stand in; the real version waits for specialist input.
Live data, simulated stakes, backtestStill next in line from v1. The backtest is what makes the weights defensible.
Native apps, 30+ sports, more languagesOn the brief's own later list.

19 events carry an AI briefing in this build. Every number on this page is read from the build itself. How the score works.