LittlePlyModel + harness

CM-14 / model + harness / recorded 9 November 2026

Better use of the same learned model.

A 50-game CPU benchmark and a dated Lichess rating snapshot show the agent’s progress. Each result keeps its own protocol and limitations.

Harness development / same CM-14 weights

From roughly 70% to a 99% observed score.

Building the harness made the same trained model more effective in recorded development games. The historical 100-game mate-only cohort scored 72%. A later 50-game full-system cohort scored 49 wins, 1 draw, 0 losses: 49.5/50 points, or 99%. Its win rate was 98%.

CM-14 against Sunfish at 1,000 nodes. Different harnesses and cohorts.
RecordedSystemGamesW / D / LScore
November 3Mate-only assistance10068 / 8 / 2472%
November 610-second uncertainty search, no helpers10073 / 12 / 1579%
November 8Cached selective search and full helpers, CPU5049 / 1 / 099%

The 99% result is 50 games, not 100. It is above 95% in this observed sample, but does not establish a reliable over-95% rate on unseen games. These rows change the harness, assistance and/or opening cohort. They are a system-development history, not a controlled estimate of the search contribution. Score counts draws as half a point; it is not Elo.

The November 8 protocol

CM-14 update 210,395; 25 fixed openings with both colors, 16 forced opening plies, then real 300+2 clocks. Sunfish uses 1,000 nodes and Hash 16. The agent uses CPU inference, a 10-second search cap, six-ply depth and 4,096 nodes, with Stockfish mate, conversion and draw-save assistance. The run completed at 17:14:53 UTC.

All 50 games were independently replayed. There were 49 checkmates and one automatic fivefold repetition, with no illegal moves, clock forfeits, inference fallbacks or cleanup failures. Four cgroup diagnostic copies were missing; the available records and enforced memory cap were checked. Shared CPU load and the repeatedly inspected opening suite remain limitations.

The November 6 no-helper result includes one clock loss; its exploratory 95% interval was 72.5–85.5%. It does not isolate the effect of more thinking time. The later 99% cohort includes deterministic assistance and cannot be described as the raw model's strength.

November 3 model and mate-only result · November 6 report · November 8 protocol and audit summary · All 50 PGNs · Game results · Opening pairs

Live play / snapshot 9 November 2026, 23:47 UTC

The progression also appears in Lichess ratings.

Blitz and rapid rose over the same period as the harness work. Bullet did not. These are Lichess Glicko-2 ratings, separated by time control, not FIDE ratings or a conversion of the Sunfish score.

Ratings before the November 8, 09:49:51 UTC deployment and at the dated snapshot
FormatBefore deploymentNovember 9 snapshotChangeRated games started after deployment
Blitz1,4441,536+9285: 33W / 3D / 49L
Rapid1,6271,719+9255: 23W / 7D / 25L
Bullet1,6521,645−717: 3W / 0D / 14L

The longer public history places blitz at 1,203 on October 29 and rapid at 1,489 on November 2. Both reached new established peaks on November 9: blitz 1,541 and rapid 1,719. Early 3,000-point provisional starting ratings are initialization artifacts and are excluded from this comparison.

This is a before/after observation, not a controlled attribution to one decoder. Model and harness versions, opponents, assistance, time controls and host conditions can change. The 157 rated games are a temporal window, not a verified single-release cohort; 152 were against bot accounts. All three current ratings are non-provisional with rating deviation about 45, but underlying strength remains uncertain and future ratings will move.

Live profile · Blitz history · Rapid history · Bullet history · Dated snapshot and sources