The current research checkpoint has played 80,000 self-play games along its active lineage. It reached the 2048 tile in 184/200 fixed evaluation games (92.0%), 459/500 validation games (91.8%), and 917/1,000 games (91.7%) on a final seed set run for the first time after the first two sets passed. All three measured rates clear the predeclared 90% threshold under these simulated game rules.

The playable 2048 website is a separate deployment. At this report's cutoff it still served the earlier 24,000-game model; the 80,000-game checkpoint and all results below are from the local training environment.

What was trained

The browser game and training program use the same 4 × 4 Python rules. A valid move slides and merges tiles, scores the values created by merging, then spawns a 2 with 90% probability or a 4 with 10% probability. Games continue after reaching 2048 and end only when no legal move remains. Accordingly, the metric in this report is the fraction of simulated games that ever reached 2048, not the fraction that ended at that tile.

The agent scores each legal move using the board after sliding and merging but before the random tile arrives. This is an afterstate. Its decision rule is:

move value = merge score now + predicted future score of the afterstate

The predicted future score is a sum of learned N-tuple lookup values. The first model used 17 overlapping patterns: four rows, four columns, and nine 2 × 2 regions. The later experiment added six 2 × 3 and six 3 × 2 regions, giving 29 patterns. There is no neural network and no human game dataset.

Each self-play game starts from a new board. The agent chooses a legal move, receives the merge score, observes the next board, and updates the previous afterstate toward the next reward plus the new afterstate estimate. At game over, the final afterstate is updated toward zero future reward. The objective is cumulative score; there is no explicit bonus for first reaching 2048. Early training used a 0.1 learning rate and 10% random exploration for the first 17,000 games. The 17,000-to-24,000 stage set exploration to zero. Later runs continued from saved weights; the current learning rate is 0.05, discount 1.0, exploration 0, and stage length 7,000 games.

new game → legal moves → afterstates → choose reward + estimated value
         → random tile → next move → temporal-difference update → game over

How results were measured

Every checkpoint below uses the same 200-game evaluation seed set. From 31,000 games onward, a separate 500-game seed set was also evaluated. Both sets have been inspected repeatedly to compare models, so neither is an untouched final holdout. The reserved 1,000-game seed set was run once, only after the 80k checkpoint passed both monitored sets. Its result is now reported and that seed set is no longer unused. The three thresholds were 180/200, 450/500, and 900/1,000.

A training-log window is a moving observation during learning. It is not a formal checkpoint result: the weights change throughout that window, and those games were part of training. For example, the last 500 training games in the 80,000-game stage reached 2048 at 92.2%, while the saved checkpoint scored 92.0% on the fixed set, 91.8% on validation, and 91.7% on the final set.

Chart 1 — fixed 200-game reach rate

Each block represents about 3.3 percentage points; the counts and percentages are exact. The 3k, 10k, and 17k figures come from the original published training note. Results from 24k onward are preserved in checkpoint reports.

3k         █                                4/200  2.0%
10k        ██                              11/200  5.5%
17k        ██                              11/200  5.5%
24k        ███████████████                101/200  50.5%
31k        ██████████████████████         144/200  72.0%
38k        ██████████████████████         145/200  72.5%
45k        ███████████████████████        150/200  75.0%
52k        ███████████████████████        152/200  76.0%
59k        █████████████████████████      169/200  84.5%
66k base   ███████████████████████        156/200  78.0%
66k 2x3    ███████████████████████████    180/200  90.0%
73k 2x3    ███████████████████████████    179/200  89.5%
80k 2x3    ████████████████████████████   184/200  92.0%

The long early plateau ended with a jump from 11/200 at 17k to 101/200 at 24k. That stage changed two things at once: it added 7,000 games and turned random exploration off. It does not isolate how much of the gain came from either change.

Chart 2 — second-seed 500-game validation

This set begins at the 31k checkpoint. It was reused for model comparison and is distinct from the untouched final set.

31k        ████████████████████           341/500  68.2%
38k        █████████████████████          349/500  69.8%
45k        ███████████████████████        380/500  76.0%
52k        ███████████████████████        389/500  77.8%
59k        █████████████████████████      422/500  84.4%
66k base   ████████████████████████       402/500  80.4%
66k 2x3    ███████████████████████████    444/500  88.8%
73k 2x3    ████████████████████████████   461/500  92.2%
80k 2x3    ████████████████████████████   459/500  91.8%

The 73k checkpoint improved validation from 444/500 to 461/500, while the fixed set slipped from 180/200 to 179/200. At 80k, fixed evaluation recovered to 184/200 while validation remained above threshold at 459/500. This was the first checkpoint to pass both monitored gates.

Chart 3 — final release gates at 80k

These are three distinct seed sets. The final 1,000 games were evaluated only after the fixed and validation results met their thresholds.

Fixed 200       ████████████████████████████   184/200   92.0%  pass
Validation 500  ████████████████████████████   459/500   91.8%  pass
Final 1,000     ████████████████████████████   917/1000  91.7%  pass

The problems that changed the plan

More games were not always better. From the 59k checkpoint, 7,000 additional games with the original 17 patterns reduced the fixed result from 169/200 to 156/200 and validation from 422/500 to 402/500. We kept both checkpoints and stopped treating episode count as a proxy for quality. The cause of the regression was not isolated; it is an observed result on these seed sets.

A shortcut with symmetry failed. At 52k, averaging the saved model's values over eight board rotations and reflections after training reached only 9/200 (4.5%) on the fixed set, versus 152/200 (76.0%) for the unmodified model. We did not adopt that post-processing. A possible explanation is that the existing lookup weights had been learned in their original orientations; rotation averaging changed their meaning without training them for shared symmetry. That explanation remains a hypothesis, not a measured diagnosis.

We compared a larger board representation from the same starting point. We copied the 59k checkpoint into a separate branch, retained all 17 learned pattern weights, and initialized the 12 new 2 × 3 / 3 × 2 pattern tables to zero. Before training, its fixed evaluation was still exactly 169/200. Both branches then trained for 7,000 games with the same game seeds and learning rate. The feature set was the intended difference.

Chart 4 — controlled 66k branch comparison

F is the fixed 200-game set; V is the 500-game validation set. The two 66k branches descend from the same 59k model.

66k base F    ███████████████████████        156/200  78.0%
66k 2x3 F     ███████████████████████████    180/200  90.0%
66k base V    ████████████████████████       402/500  80.4%
66k 2x3 V     ███████████████████████████    444/500  88.8%

Adding the larger patterns moved the fixed result from 156/200 to 180/200 and validation from 402/500 to 444/500 at the same 66k training depth. This supports the feature expansion under this comparison, but it does not prove the larger architecture will outperform on all unseen games.

Score and 2048 reach rate can move differently. From the expanded 66k model to 73k, fixed-set mean score rose from 51,384.50 to 53,215.08, but the fixed 2048 count fell by one game. The current TD reward optimizes score, while the release gate measures reaching a tile. Reward changes are a candidate future experiment if this mismatch persists; none has been introduced yet.

Chart 5 — fixed-set mean score

Bars use a 60,000-point scale. Exact values are in the saved reports; whole numbers are shown here for readability.

24k        ████████████                   24,827
31k        ████████████████               31,673
38k        █████████████████              33,521
45k        ██████████████████             36,867
52k        ████████████████████           39,025
59k        █████████████████████          41,990
66k base   ███████████████████            37,751
66k 2x3    ██████████████████████████     51,385
73k 2x3    ███████████████████████████    53,215
80k 2x3    █████████████████████████████  57,951

The larger representation also costs storage. Compressed model size grew from about 3.5 MB at 59k with 17 patterns to 30.1 MB at expanded 66k, 39.2 MB at 73k, and 46.1 MB at 80k with 29 patterns. We have not measured a production inference-latency comparison.

59k / 17   ██                         3.5 MB
66k / 29   ████████████████████       30.1 MB
73k / 29   █████████████████████████  39.2 MB
80k / 29   █████████████████████████████  46.1 MB

Release status and reproducibility

The active research model follows the 80k expanded-feature branch. Across all branches, the trainer has executed about 87,000 self-play games: 80,000 along the active lineage plus a separate 7,000-game ordinary-feature comparison from 59k. Those are training attempts, not 87,000 games in one model.

The current local artifact passed all three predeclared 90% evaluation gates: fixed 184/200, validation 459/500, and first final evaluation 917/1000. Automated training was stopped at this checkpoint. This is evidence for the specified simulated environment and seed sets, not a guarantee for every game or a claim that the public website already serves this model.

The underlying local records are the checkpoint reports in artifacts/checkpoints, the eval-*-validation.json and eval-80000-2x3-final.json files, the stage logs, and config.json. The first three early points above are retained from the previous published version of this article. These measurements describe a simulated 2048 agent under specified seeds; they do not claim that a public user will see the same rate in any single game.