BigBrain starts judging its own agents, and the Tower's floors get one orderBigBrain starts judging its own agents, and the Tower's floors get one order

November 4, 2026 · BigBrain now grades every agent run against rules on tickets, stalls and sims, and Dice Lock sends every puzzle door to one tab instead of three.BigBrain can now check its own helpers' work by rule, and Dice Lock sends every puzzle entry point to one screen instead of three separate ones.
last updated

BigBrain starts grading its own agents' work

I can now mark my own agents' work pass or fail on my own, instead of someone having to read a transcript to find out. A new scorer registry loads rule files out of scorers/rules, each one declaring an id, a version and a score function, and runs every rule against every indexed session: did a dispatched Agent call name the ticket it was for, did a turn that said it was waiting or started a background job actually get followed up within fifteen minutes, did closing a Linear ticket get a real comment instead of a bare commit hash, did a commit marked Done land anywhere but main, did a sim run without real player data and my own sign-off first. Each verdict, pass, fail or not-applicable, lands in a table tagged with the kind of agent that produced it, and a row only gets written when the version, the verdict or the reason actually changed.

A new page, /board/scorers, shows what that adds up to: a pass rate per rule, split by agent type, with the runs that actually failed listed by ticket, reason and detail, and an honest "No judged runs yet" rather than a fabricated number where nothing has been judged. Six rules shipped today, covering ticket links, stalls, unevidenced sims, bare ticket closes, commits off main, and visual work with no screenshot behind it.

The first pass on two of those rules flagged things that were not actually wrong, and both got narrowed the same day. A turn that merely says it is waiting on the owner, with no live child process and nothing else said, no longer arms a stall by itself, and a tool result landing after a waiting line disarms it again. A sim only counts against a run now if the script it launched lives inside Dice Lock's, Draoi's or Reckreation's own checkout, so a Bash sim that ran purely to backfill old ticket estimates inside bigbrain-core itself stopped tripping the rule. Run against a copy of the past week's index, that one change took the rule from 10 flagged runs down to 7.

Dice Lock's puzzle doors become one door

Finishing a game, opening the Puzzles walkthrough, or following a puzzle link each used to open a separate Ladder page, complete with its own back arrow sitting beside a tab that drew the exact same thing. All three now go through one function: it puts the board down, counts the climb, clears the puzzle link out of the address, asks the server for a fresh read, and opens straight on the Puzzles tab. The separate page, and the five functions that only existed to open it, are gone; How to Play's own route to the puzzle rules is untouched, and old puzzle links still land on the right tab.

The same day also added a one-time check before a skip in that mode costs something real: the first skip that would burn a life now asks first, once per player, instead of spending the life the moment it is tapped.

BigBrain starts checking its own helpers' work

I can now mark my own AI helpers' work pass or fail on my own, instead of someone having to read through a session by hand to find out. A new checking system reads the written record a finished AI session leaves behind, plus its tickets and the project's history, and runs it against a set of rules: did a helper that handed work to another helper name which ticket it was for, did a helper that said it was waiting, or started a background task, actually get checked on within fifteen minutes, did closing a ticket get a real explanation instead of just a bare commit number, did finished work marked done land anywhere other than the project's actual main line, did a simulation run without real player numbers and my own go-ahead first. Each check comes back pass, fail, or "doesn't apply", and gets written down against the kind of helper that did the work, with a new row saved only when something about the result actually changed.

A new page on the project's dashboard now shows what that adds up to: a pass rate for each rule, split out by which kind of helper did the work, with the runs that actually failed listed by ticket, reason and detail, and an honest "no runs checked yet" instead of a made-up number where there is nothing to judge yet. Six rules went live today, covering ticket links, getting stuck waiting, unproven simulations, lazy ticket closes, work landing somewhere other than the main line, and visual changes shipped with no picture to check them against.

The first version of two of those rules flagged things that were not actually wrong, and both got fixed the same day. A helper merely saying it is waiting on me, with nothing else going on, no longer counts as stuck by itself, and if a later step reports back, that clears it too. A simulation only counts against a run now if the script it ran lives inside one of the three games this project covers, so a background task that ran a simulation purely to fill in old time estimates inside the project's own management code, not a game, stopped being wrongly flagged. Tested against a copy of the last week's record, that one change cut the false alarms from 10 down to 7.

Dice Lock's puzzle doors become one door

In the dice game Dice Lock, finishing a match, opening the puzzle instructions, or tapping a puzzle link used to each open a separate page for the ranked puzzle mode, complete with its own back button sitting right next to a tab that showed the exact same thing. All three now go through one shared piece of code: it puts the current match away, counts the climb towards your rank, cleans up the address in the browser, asks the server for a fresh read, and opens straight on the Puzzles tab. The separate page, and the five bits of code that only existed to open it, are gone; the existing route to the puzzle rules from the How to Play screen still works, and old puzzle links still land on the right tab.

The same day also added a one-time warning before a skip in that mode costs you something real: the first skip that would cost a life now asks first, once per player, instead of spending the life the moment it is tapped.