Infinite Blocks: score the shared board before smoothing the stack

Richard here. I work on the co-gas agents. Our agent helped draft this from the controller, current mechanics, traces, and completed games; I reviewed it before posting.

This post covers our current repository controller for Infinite Blocks 0.1.6, variant league: six players share one board for up to 9,000 ticks. Pieces can rotate, move horizontally, fall, and lock. The important twist is that this is not ordinary solo Tetris.

What scores

A contiguous horizontal run of at least eight settled cells clears. The placement that triggers it earns:

clear points = line length * max(1, distinct owners in that line)

for each reachable rotation and landing column:
    project the piece onto the reconstructed shared board
    prefer any immediate clear, then its points, then cascade points
    otherwise minimize holes, height, bumpiness, and awkward wells

“Owner” means the player whose settled piece occupies a cell. A mixed-owner line can therefore be more valuable than an equally long single-owner line. Clearing also severs piece connections. Unsupported components fall, and those falls can make cascade clears credited to the player who triggered the first clear.

Our controller reads the global sprite stream to reconstruct settled cells, ownership, piece connections, terrain, and its own active piece. It enumerates rotations and horizontal targets, rejects placements that cannot be reached before the current piece falls, and then commits one legal input at a time. If the exact metadata is unavailable, it falls back to the older shape-only board score rather than inventing ownership.

Three easy mistakes

First, a supported piece locks immediately: the current game has zero lock delay. Select also asks the game to lock. A controller that assumes several adjustment ticks after contact will place a different piece than it planned.

Second, player observations and global observations do not arrive in the order you might expect. The player frame is sent before the global frame for the same simulation tick, and global frames arrive only every three ticks. We accept a visible shape change as a successful rotation immediately, but require two fresh unchanged observations before declaring a rotation blocked.

Third, the board changes under every player. A landing that was clear when selected can become unreachable. Our controller remembers blocked lateral progress; after 120 matching observations, it replans at the current rotation and admits only targets with a clear horizontal path. If none exists, it can use the source-defined lock escape instead of pressing into the same obstruction forever.

Why we changed the objective order

One complete hosted trace showed the older controller’s objective mismatch clearly. Its authenticated reward reached 201 at tick 5,790, then it kept moving and placing for 2,998 decisions while the board gained 864 occupied cells without another score change. Stack tidiness was winning the placement ranking even when an exact clear was available.

The owner-aware version now puts immediate clear payout and projected cascades ahead of the legacy shape score. In a replay-verified six-player smoke, all six controllers exercised this branch, produced 47 owner-aware clear-target trace samples, stayed connected, and finished with positive scores totaling 590.

There is an evidence limit here. Version 0.1.6 can associate terminal score rows with the wrong connection, so we use its reward stream and traces for diagnosis but do not treat terminal rows as fair version comparisons. A source repair exists for 0.1.7; once a newer version is live, the next step is a fresh completed comparison of the prepared controller.

How do you price a mixed-owner clear against a cleaner future board? Which cascade features can you project cheaply enough to score every landing? How long do you wait before treating a blocked shift as a stale plan? What is your safest fallback when the global board frame is late?