A tournament's standings page now ranks players by the number chosenA tournament app's rankings page now shows the right numbers.
August 26, 2026 · Reckreation's standings and profile pages went live, catching a rating printed to thirteen decimal places along the way, while bigbrain-core stopped repeat sessions re-deriving what an earlier one already verified.A tournament app's rankings page now works, and along the way it caught a rating displaying with ten extra decimal places nobody asked for.
last updated
Standings that rank on the number an org actually chose
Reckreation, my tournament tracking project, shipped its standings and player profile pages today. An org can already choose whether its ladder ranks players by set wins or game wins (that setting has existed since JIO-557), but nothing read it. An org that picked set got a table ranked on something else entirely, because the column it wrote to and the column the page sorted on had never been the same code path.
Fixing that meant settling something nobody had actually written down: what a W-L number is even supposed to mean when it drives a ranking. Wins alone puts a 20-4 record above a 15-0 one. A win rate rewards the smallest sample on the page. Points makes a loss free. The one that agrees with the rating sitting next to it in the same row is wins minus losses. Against a field of equals, rating moves by a constant times exactly that number. The ranking and the rating badge now can't disagree, because both come out of one function instead of two.
The player profile page turned up a sharper bug on the way in: an early proof run against a scratch database printed a rating of 1499.4535888003202. Nothing had ever rounded a number before it hit a page, so a rating that should read 1500 could carry ten extra digits nobody asked for. Display rounding now lives in one place, and the underlying number stays exact everywhere it's used for math. The same page adds a career sparkline, drawn as an inline SVG rather than a charting library, with explicit handling for a flat career (divide-by-zero) and a single game (one point isn't a trend).
None of this touched production data. Every proof ran against a scratch database with real Clerk sessions and disappeared afterward; reckreation itself still holds zero live event rows. It landed alongside the rest of the day's 79 commits in the same repo: three independent rating pools, a Challonge bracket importer, and the event-sourced match core all of that stands on, most of it building toward this being the first page a player actually sees.
A handoff between AI roles now carries what the last one already checked
Separately, in bigbrain-core, a token audit from the day before had found that 122 of 266 Linear tickets ran through more than one AI session by design (triage, then delivery, then an auditor, sometimes back to delivery), and each one started cold, re-reading the whole issue to rebuild context a prior session already had. That repeat work summed to 1.26 billion cache-read tokens over two weeks, about a quarter of all ticket spend.
A handoff between roles now carries four short structured lists alongside its note: files touched and why, decisions made, facts already verified that the next session must not re-check, and questions still open. Every later role on that ticket gets the full accumulated ledger, oldest entry first, plus the diff the handing-off session actually committed, so the next session starts from a scoped diff and a list of settled facts instead of the whole repository. A follow-up commit the same day caught that the underlying inherit logic hadn't actually shipped working in production, despite the warning around it firing correctly.
Reckreation, a side project of mine for tracking tournament results, got its standings and player-profile pages working today. An organization running a ladder can already choose whether players get ranked by match wins or by individual game wins, but that choice wasn't actually being read anywhere. Pick one, and the page ranked people by something else entirely.
Fixing it meant deciding what a ranking number should even mean. Counting wins alone would put a 20-loss record above a spotless 15-0 one. A win percentage rewards someone who's barely played. The fair measure turned out to be wins minus losses, since that's the same math the skill rating next to it already uses, so the ranking and the rating badge can no longer disagree with each other.
Building the player page also caught a sharper bug: a test run showed a rating displaying as 1499.4535888003202 instead of a clean 1500. Nothing had ever rounded a number before showing it on screen. That's fixed now, and the raw precise number is still what's used for any actual math behind the scenes. The profile page also got a small chart showing a player's rating over time, with care taken for someone who's only played one game or never won at all, cases that would otherwise break a simple chart.
All of this was tested against a scratch copy of the database, not real data, and it's part of a much bigger day in that project: 79 commits building the skill-rating system, a way to import brackets from another tournament site, and the underlying event log all of it depends on.
Separately, in the background-worker system behind a lot of this work, an audit found AI sessions handing off a ticket kept re-checking things an earlier session had already confirmed. A quarter of two weeks' spend was sessions repeating each other. A handoff now carries a short record of what was touched, decided, and already verified, so the next session skips past it instead of starting cold.