The main engine's fight calls, with the receipts attached.

Two different claims live here, and mixing them up is how a record starts lying. Sealed means a copy exists in a record the database won't let anyone change — that is about tampering. Dated is about timing. A row only reaches the top grade when both hold: a sealed copy exists, it still matches, and it was taken on an earlier day than the card. If we ever gain a trustworthy per-fight start time to compare against, the same-day rows can be graded up then — we are not going to guess one now.

1. We post a call, and the clock starts

   As a card approaches, the model writes a row: which fighter, what percentage we give them, and — where we flagged a bet — the price on offer at that moment. The row gets a timestamp. What that timestamp can prove varies by row, so we grade it rather than wave at it.
2. A copy gets frozen where nobody can reach it

   The table holding the calls is an ordinary working table. We could edit it. So for every card we cover, a copy of what we were showing is written into a separate record that the database refuses to change or delete afterwards — for every account, including the one that owns the database. This page then re-compares the working row against that frozen copy in front of you. A pick that aged badly can't quietly become a different pick, because the sealed copy would stop matching and this page would say so.
3. The fight happens, and we grade it against what actually occurred

   Won, lost, or not settled yet. Nothing gets quietly dropped for being embarrassing — if a pick is missing from the archive it's because it was never posted, and you can check the count against the card.
4. Money is counted flat: $100 a bet, every time

   Same stake on every bet, no chasing, no doubling up. A win at +150 pays $150; a loss costs $100. That's it. We use the price that was on record when we posted, not a better one we found later.
5. A number only goes on screen once there's enough of it

   Small samples lie. Each headline number has a written rule for how many calls it takes before we'll show it. The hardest one — whether we beat the market's price — isn't ours at all: it belongs to a separate research protocol, frozen before any of its own results were computed or reviewed, and this page just reads its verdict. Until a rule is met you get "still collecting" instead of a number. We'd rather show you an empty box than a flattering guess.
6. Rules can improve — but only forwards, and in writing

   We are not going to pretend our methods are frozen forever; a model nobody is allowed to improve is a worse model. What we will not do is rewrite the past. Every change to how something is measured gets a version and a date, and results already scored stay tied to the rules they were scored under. Old numbers are never quietly re-graded under a friendlier rule, and a threshold is never moved after we've seen what it gates.
7. Simulations stay labelled as simulations

   Re-running the engine over old fights is a useful sanity check and nothing more — the model gets to be smart about a world that already happened. Those rows are marked, boxed off, and never added to the live record.

What this page will not tell you

Being upfront about the limits is part of the point.

• Whether you'll make money

  A record is a record. A 70% pick still loses 3 times in 10, and a good run of results can happen to a bad model. Nothing here is betting advice.
• That we're beating the market

  We're not claiming it, and we're not quietly implying it either. The rule for when that question may be answered isn't ours to apply, and it hasn't been met. The market's price is right more often than any model. Any edge lives at the margins.
• That the simulated numbers predict the live ones

  They routinely don't. That gap is the single most common way a model looks better on paper than in practice, which is exactly why the two records are kept apart here.