Research archive / preserved development history
The full development record.
Historical scores, model experiments and dated system descriptions. Each result keeps its original protocol; this is not a controlled learning curve.
LittlePly is an experiment in AI research guided by Andrei Savu, working with his OpenAI dot, Glimmer, on a local Lambda Vector Workstation.
LittlePly combines an 87-million-parameter transformer trained from scratch with a harness for search and deterministic assistance. CM-14 is the deployed model; the harness can evolve without changing those weights.
Play LittlePly on LichessPlay the agent on Lichess
The online agent adds uncertainty-guided search, caches, bounded Stockfish assistance and clock-aware execution. Browser play is the smaller, unassisted model-and-legality demo. See how the agent works.
Model + harness
Learn the game. Design how to play it.
01 / The learned model
Chess intuition
CM-14 was trained from scratch on chess games. Its transformer learns a policy over moves and a win/draw/loss estimate from game history. Training shapes what it predicts and how useful those predictions are.
What the model learns02 / The harness
Decisions under constraints
The harness maintains the legal game state, chooses where to search, reuses cached work, coordinates assistance and decides when to stop thinking.
How a move is chosen03 / The complete agent
Play, measure, improve
It acts in a live game, handles clocks and service events, and records what happened. We evaluate the model and the assembled agent separately.
Inspect games and analysisThe research has two connected parts: train a better model, and build a better harness around it. A change to either can change the agent's behavior.
Recorded progress / updated 9 November 2026
The model and the harness develop together.
72% → 99% observed score
CM-14's Sunfish development history spans 68W / 8D / 24L over 100 games with mate-only help, then 49W / 1D / 0L over 50 games with the improved harness and full helpers. Both use Sunfish at 1,000 nodes.
Exact counts and protocol changesBlitz 1,536 · Rapid 1,719
Both rose 92 Lichess rating points from just before the November 8 deployment to the November 9, 23:47 UTC snapshot. Bullet fell 7 to 1,645. A dated observation, not a controlled attribution.
Ratings by time controlCM-14 remains the model
Harness work changes how the agent uses its learned policy and value estimates. CM-16's multi-day auxiliary-objective experiment did not establish a replacement.
The unsuccessful training experiment99% is the observed score in one 50-game cohort, not a verified reliable over-95% rate on a new 100-game suite. The system protocols differ. The CPU result and public ratings do not establish deployment or strength gains for a new GPU harness.
CM-16 / November 4–9 / Not promoted
A multi-day training experiment that did not replace CM-14.
CM-16 tested auxiliary objectives, continued over several days, and later returned to next-move plus WDL training. It did not establish a stronger replacement. The idea that extra objectives dilute useful learning under a limited compute budget is a working hypothesis, not a controlled result.
Read what happened in CM-16 · Earlier CM-15 evidence
Research note 02 / PDF
LittlePly as a chess agent
Early draft v0.2, 10 November 2026, not peer reviewed. CM-14 and the harness that turns predictions into play: a five-page account of the combined agent, dated benchmark and Lichess evidence, and measured lessons from CM-15 and CM-16. It separates model learning from system performance and treats the auxiliary-objective explanation as a working hypothesis.
Read the research draft (PDF, 5 pages)
Preserved earlier draft: v0.1, 6 November 2026 (PDF, 3 pages). Its proposed evaluation predates the completed 50-game full-system result. See the dated evidence and protocol differences.
Follow the model and the harness
1. Understand
How moves become tokens, what training optimizes, and where rules and search enter.
Learn the system2. Inspect
Read successful and unsuccessful experiments, then replay the evidence.
Explore experiments3. Reproduce
Give an agent a recipe, versioned source, datasets and explicit verification gates.
Build your ownSunfish development benchmarks
Progress, with the method changes visible.
Score = (wins + half of draws) / scheduled games. Bars share a 0–100% scale. Protocol groups are separated deliberately: this is a research history, not a single controlled learning curve or Elo scale; not a general Elo estimate.
October 29 · data and continuation
Frozen three-ply helper; 100 scheduled games each. Unfinished games stay unfinished.
14% · 13 / 2 / 85 W/D/L · n=100
CM-4Broader human data; one unfinished game
25.5–26.5% · 21 / 9 / 69 W/D/L · 1 unfinished · n=100
CM-5Elite continuation; one unfinished game
31.5–32.5% · 27 / 9 / 63 W/D/L · 1 unfinished · n=100
October 29–31 · approaches that did not establish a gain
Different system protocols, not a matched group. CM-6/CM-7 tested state input and adaptive custom mate assistance; CM-8 tested aggressive masking. See each record for its own configuration.
CM-6State input; research-only, not promoted
16.5% · 13 / 7 / 80 W/D/L · n=100
CM-7State curriculum; two unfinished games, no clear CM-5 superiority
30.5–32.5% · 21 / 19 / 58 W/D/L · 2 unfinished · n=100
CM-8Lower training loss did not yield stronger play
7% · 6 / 2 / 92 W/D/L · n=100
October 31 · return to full history
Separate promotion suites; 100 games per reported checkpoint. Their matched references differ from the earlier suite.
CM-9CM-5 lineage; engine/human mixture and masked-gap training
32% · 23 / 18 / 59 W/D/L · n=100
CM-10Longer-gap continuation; mixed head-to-head evidence
37.5% · 25 / 25 / 50 W/D/L · n=100
November 2–3 · full-move tokens and Stockfish mate assistance
Stockfish 19 mate-only assistance; 100 games per row. No endgame takeover. CM-13 final had a freshly matched reference; CM-14 used archived references.
CM-13 finalFresh AdamW run; clean full-move history and separate WDL diagnostic
67.5% · 63 / 9 / 28 W/D/L · n=100
72% · 68 / 8 / 24 W/D/L · n=100
November 4 · CM-15 initial experiment
New matched CM-14 reference: 72.5%. Both used identical mate checks and 97% endgame conversion. Difference -4 percentage points; paired 95% interval -15 to +7. Separate from the older mate-only cohorts above; no improvement established.
CM-15 initialIsolated FUTURE_8 target; completed research, not promoted
68.5% · 64 / 9 / 27 W/D/L · n=100
These selected development suites are small and repeatedly inspected. More data, training, input changes and helper changes can all affect results. Matched comparisons belong inside their own experiments; do not infer a causal gain between every neighboring row. Historical unfinished scores show outcome bounds, not confidence intervals.
All experiments, including unsuccessful approaches · Checkpoint identities · Chart data and evidence links
Two connected lines of research
Early checkpoints predicted three tokens per move. The state-input experiments tested a different representation without establishing a strength gain. CM-8 showed why lower loss need not mean stronger play. CM-9 returned to full-history training with engine and human games. CM-13 restarted with one token per complete move and a separate WDL diagnostic. CM-14 tested a new optimizer and continued on human games. CM-15 then tested learning a future move alongside the next move; its completed initial evaluation did not justify promotion. That result remains part of the research, even though CM-14 stays live.
Harness design also includes uncertainty-guided search, caching and clock allocation. The historical serving path below explains earlier comparisons; it is not a complete description of the current agent. Current agent architecture.
Serving evolved separately: bounded custom mate checks were replaced by asynchronous Stockfish 19 mate reports. The online bot later added a seven-piece, two-confirmation, above-97% endgame rule. That conversion mode is absent from the older mate-only Sunfish cohorts and from browser play. The November 4 CM-15 comparison includes it for both models; its matched CM-14 reference is a different evaluation.
See how the pieces fit togetherBefore challenging LittlePly
Standard-start chess; availability depends on the host, current games and Lichess limits. The challenge dialog remains editable and needs your confirmation. Play is available on Lichess. Hardware and serving context.
The playing record
One place to follow every kind of play.
Browse public Lichess games, tournament checkpoints, evaluation suites and workstation runs by model and context. Replay older games, and inspect detailed analysis where it was recorded.
Explore games & analysis
Unknown metadata and unfinished games stay visible. Historical analysis has not been backfilled; the hub documents its source coverage and gaps.