LittlePlyModel + harness

EXP-18 / evidence study

CM-14: current, with clear boundaries

Which system produced the published 72%?

Keep in view. Single cohort; no matched comparator or confidence interval supplied for this older benchmark.

Published benchmark

100 games; mate-only assistance. Do not substitute this result for the newer 72.5% matched reference in CM-15 evaluation.

Published benchmark — complete recorded values
CohortWinsDrawsLossesUnfinishedScore / bounds
CM-14 u21039568824072%

Model and interfaces

Current release identity is preserved; this presentation does not modify weights or serving configuration.

Model and interfaces — complete recorded values
InterfaceRecorded configuration
InitializationFresh Muon/AdamW lineage; not CM-13 continuation
BenchmarkMate-only
Live botSeparate 97% endgame conversion enabled
BrowserUnassisted

Download the study’s recorded values. Exact source records, hashes and protocol details remain below. Selectors inspect recorded observations; no missing points are interpolated.

Full experiment record, provenance and original links
Study source files and SHA-256 provenance

Original narrative and source links are preserved in the full record below. A linked release inventory may document identity without supplying every study value.

EXP-18 / completed / CM-14 / continuation

EXP-18: CM-14 continuation and 100-game Sunfish evaluation

Update 210,395; promoted by user request after the 100-game Sunfish evaluation.

Evaluation

Sunfish benchmarkWins / draws / lossesScore
CM-14 continuation, 100 games68 / 8 / 2472%

Draws count as half a point. This is a development benchmark, not a general Elo estimate or unassisted browser strength. It does not establish a statistically significant advantage over another model.

The evaluated checkpoint used the matched Stockfish 19 mate-only helper protocol. The live bot retains its separately enabled 97% endgame conversion, which was not part of this evaluation. Browser play uses neither helper nor endgame conversion. Keep these configurations separate when interpreting results.

Model and lineage

The 86,966,016-parameter transformer reads clean, complete move history: one token per UCI move, including promotions. It has 12 layers, 12 heads, width 768, 512 context positions and 1,975 vocabulary entries. Move logits and fixed-White win/draw/loss logits are normalized separately. Outcome labels are not supplied as input, and the WDL diagnostic does not choose the browser move.

CM-14 began from fresh random initialization with a hybrid Muon/AdamW optimizer. This promoted checkpoint continues that CM-14 run; it is not a continuation of CM-13 weights.

Exact identity

Source SHA-256 31451d0d425ef14e020b847a3d1fefcbceacd32b0e1123adeab5f62877c379c2; inference export SHA-256 4cee1bf258f6061843f6073cd21d65f8981153ae47108551df617cb3053eea72.

Release metadata · Browser validation and source · Historical CM-13 evaluation

Related reading

Context for interpreting this experiment, not a claim that it implements these methods: Amortized Planning with Large-Scale Transformers: A Case Study on Chess; Policy improvement by planning with Gumbel; A general reinforcement learning algorithm that masters chess, shogi, and Go through self-play.