Technical research notes

Chess-True Architecture

Chess-True is built as a long-term, persistent chess analysis tree. The system stores calculated positions, propagates evaluations through mini-max, and keeps the data compact enough to support very large scale analysis.

600B+ calculated positions in the current test log
8 bytes per tree record
5B positions per inspection snapshot

Compact Record Format

Each tree record is stored in only 8 bytes. The record contains the position evaluation, the number of moves from that position, and the address of the first child record.

2 bytesposition evaluation
1 bytemove count
5 byteschild address

A 5-byte address gives 40 address bits, enough for 2^40 records, or about 1.1 trillion positions. If the tree needs to grow further, one bit can be borrowed from the evaluation field and one bit from the move-count field, extending the address space to 42 bits, or a little more than 4.3 trillion positions.

Snapshot Metrics

The test log records long-term snapshots of the tree. The most complete rows track calculated positions, discovered draws, draws per billion calculated positions, the position-to-draw ratio, the snapshot date, evaluation thresholds above 100, 150, and 200, and empty or extra records.

Integrity Pass

In the current workflow, the cluster is usually paused after about every 5 billion new calculated positions. A separate tree-inspection program then scans the analysis tree with a linear algorithm. At the current database size, this produces a full snapshot in about one hour.

The inspection pass is not only a statistics tool. It checks data integrity and isolates records that no longer have a parent in the analysis tree, so parentless records do not pollute the main structure.

Pruning And Cleanup

Some historical drops in the data are intentional. Chess-True includes a cleanup tool that removes redundant cluster records. Because the data has grown very large, cleanup is slow and is used only occasionally. After cleanup, stored position counts can decrease and high-evaluation counters such as positions above 200 may reset.

Research Question

One research hypothesis behind Chess-True is that, as the stored analysis tree becomes more saturated, new draw discoveries per billion calculated positions should decrease, while the ratio between calculated positions and discovered draws should increase.

The current data has not clearly reached that stage yet. This may be a question of scale, or it may reflect a deeper chess property: if perfect chess is a draw, then draw regions may be large, distributed subtrees rather than isolated points.

Current Test Hardware