GMS · v2.5 RELEASE NOTES

00 · TIES

Level, and it says so.

Two decks holding the same three cards have the same power at every instant. The pairing card showed that as neither ahead nor behind — the same bare surface it uses for a power it has not resolved yet. Two situations, one appearance, and only one of them was actually undecided.

A level pairing now says Level — decided by entry order. It appears while the round is in play, both powers are known, and they are equal. That is the only case it fires on.
It does not say who wins, and it cannot. The tie is settled by which of the two decks entered first, and that order is not published before the round settles. The line states that a tie exists and that a rule decides it, and names neither side. The entry index is not on the wire the app reads.
Settlement itself was never wrong. Sixteen live tied pairings were checked against the tournament's own records — every one resolved to the deck that entered first, exactly as the rule says.
The colour and the border are untouched. On that card colour means standing, and a tie is not a standing.

01 · ROUNDS

One round number, everywhere.

Enter said "Round 1 started at 19:00" while Battles said "Round 3" — same tournament, same instant. Four places in the app each worked the round out for themselves, from a number that meant something slightly different on each side.

The server resolves the live round once and serves the answer. The Deck line, the Predict chip, the survival grid and the tournament screens all read that one value instead of four copies of the same formula.
The countdown points at the round's real comparison time, not at a time derived from the schedule it was expected to keep.
The round schedule re-anchored on the last round actually recorded. A stopgap from an earlier release, kept alive past the condition it was written for, is gone.
Talking to an older server, v2.5 falls back to working it out the old way — so the number is never blank, only ever better.

02 · CHANCES

The odds belong to the deck on screen.

Swapping one card for another keeps your selection at three cards, so the chance grid never cleared. The previous trio's survival percentages stayed up — with no spinner, looking exactly like a fresh reading — until the new ones arrived.

An estimate is now tied to the three cards that produced it. Change a card and the grid shows that it is working, rather than the outgoing deck's odds wearing the incoming deck's name.
Order still does not matter. The same three cards in a different arrangement are the same deck to the estimate, exactly as they are to entry.
The final round of a bracket is derived in one place and clamped, so the End button and the survival curve can no longer disagree about where the tournament ends.

03 · WORDS

Nine sentences that could never be translated.

A handful of failure messages were assembled as English text inside the app's code and rendered in the middle of an otherwise translated screen. Not translated badly — untranslatable, by construction.

Seven wallet-scan messages moved into the app's string catalogue. Out of charges · not available on your tier · reconnect your wallet · rate-limited · card service unavailable · connect your wallet first · and a plain retry for everything else.
The game-wallet check and the update check got the same treatment.
Which means they can now be changed without a new build. The catalogue is the layer both the translations and the server-side text corrections reach. They read English until each language lands; before this release they read English permanently.
The raw technical cause is no longer part of the message. It goes to the log, where it is useful, instead of onto the screen, where it is not.

04 · LINKS

A link is not worth a release.

The Discord invite was a line of code. Changing it — a server migration, a revoked link — meant a build, a submission and a wait, to alter one URL.

The Discord button asks the server for the current invite. The next change to it is a config line, not a release.
The built-in link stays as the fallback, so the button still works offline or before the first answer arrives. It is a safety net now rather than the record.
GMS invites carry their own link labels. A partner entry that offers a download says Download link rather than a generic label the app had to guess.
Four partner projects joined the invites directory this cycle — SeekerTapper, Mine / Offline Protocol, Sol Incinerator and Trench.

05 · REWARDS

Half the field finishes ahead.

The tournament payout ladder was flattened. It pays less at the very top and pays it to twice as many people. This is server-side and already live — it does not wait for you to update.

Surviving round two now pays more than your entry cost. Break-even is round 2, so the quarter of the field knocked out there joins the paid population instead of leaving empty-handed.
Half of every field ends in profit, up from a quarter. Paid and ahead are now the same event — no rung pays exactly your stake back.
The champion multiplier still climbs, from 13× in a small bracket to 260× in the largest. It is about a third smaller than before at every depth, and that third is what pays the wider field.
Every figure in the ladder is frozen against a self-test, so a later edit cannot move a payout quietly.

06 · FORECAST

Four readings, not one.

The 7-day projection ran off a single lookup table. It now blends four views of the same move — the raw week, the week relative to the rest of the board, the last day, and the move measured against the coin's own volatility.

Measured, not asserted. Thirteen months of walk-forward testing, refitting on the past and scoring on the month after it, on a price history assembled from two independent vendors.
The worst months got half as wrong. Rallies are where a projection breaks; the two worst months in the sample halved their miss. The typical month improves by less than a percent — real, and small.
The blend is the point. No single one of the four beat the old table on its own; averaging them did.
A projection is a projection. It is a measured tendency over a fortnight of hours, not a promise about tomorrow.

07 · POWER

A year of history that was never there.

One of the card ratings is described everywhere in the app as a 365-day reading. It had never held more than about eight weeks — the daily pass that was supposed to fill the rest had never written a surviving row. It was scoring a year's worth of behaviour on 54 days of it.

The year exists now. Average history per card went from 54 days to 339, and the qualifying bursts it looks for from 158 to 793. Scores that were all squashed into the bottom eighth of their range now use it.
Rebuilt from two price sources, not one. Each day's own trading turnover is used where a vendor recorded it — 84% of rows — instead of a recent quiet-period average being applied to price moves up to a year old, which is what flattened the historical peaks by roughly sevenfold.
Every row carries where it came from. Measured, bounded, estimated or fallback — so a weak row is findable rather than indistinguishable from a strong one.
Your card letters and bands did not move. Not one card, checked before and after. The power tier reads a different table over a 60-day window and never consulted this one — this feeds the long-horizon ladder ranking, the historical peak the sell signal compares against, and the quality column.
The live database was snapshotted before the rebuild and the result verified from the data afterwards, not from the console output.

08 · ENTRIES

Money that landed, counted as landed.

A transaction that reached the chain could still be sitting in the app as pending, because the only thing that ever promoted it was the app being open at the right moment.

A five-minute sweep re-checks pending transactions against the chain and promotes the ones that landed. It found and cleared three purchases that had been stranded for three weeks.
An entry-fee check that could not be read no longer counts as "not an entry fee". A rate-limited or unreadable read now retries instead of quietly deciding against you, and a verdict that was reached once is remembered rather than re-derived.
The tournament you can actually enter is the one shown as upcoming. A tournament whose registration had already closed could outrank it simply by starting sooner.
Round 1 is anchored on when entry closed, not on the hour the tournament was scheduled to open — so the whole schedule after it lands where it should.

09 · ALSO FIXED

Quietly not working.

The download link on this site pointed at the previous release for eight days. Sideloaded builds were told an update existed and handed the old file.
The card listings mirror recovers from a silent feed. When the live listings stream goes quiet, a watchdog now backfills the window it missed instead of leaving a hole in the desk.
The player and deck lists on the tournament screens read the same live round as everything else, so a spectator and a player see the same bracket.
The app's About line reports v2.5.