osazuwa/wrong-move-reference-v1

Wrong Move Reference v1

Wrong Move Reference v1 is a synthetic longitudinal Match-3 dataset for studying causal adjustment with learned world models. It contains paired natural and randomized assignment regimes generated by the same explicit causal simulator.

The natural regime adaptively serves harder content to stronger simulated players. The randomized regime removes that dependence while preserving the marginal assignment variance conditional on level and baseline tier. The two regimes use the same 2,400 synthetic player IDs, player-level train/validation/ test split, and deterministic player-attempt seeds.

The accepted release contains 84 payload files (115,342,801 bytes). Its root status is accepted_reference_dataset; qc.json reports pass.

Intended use

The release supports player-disjoint training and evaluation of:

• strict-history adjustment representations;
• Match-3 state, behavior-policy, and retention models;
• observational, direct-adjustment, and rollout-based causal estimators; and
• the contrast between adaptive natural assignment and its randomized control.

All people, player histories, gameplay traces, and outcomes are simulated. The dataset contains no human participants or personal information.

Release layout

accepted_benchmark.json
release-manifest.json
qc.json
splits.json
natural/
  shard-000/ ... shard-009/
    episodes.csv
    transitions.npz
    manifest.json
    oracle/attempts.csv
randomized/
  shard-000/ ... shard-009/
    episodes.csv
    transitions.npz
    manifest.json
    oracle/attempts.csv

Each regime has ten contiguous 240-player shards. Across both regimes, the release contains 95,455 attempt rows, 1,719,631 gameplay-transition rows, and 95,455 physically separated oracle rows.

splits.json assigns 1,680 player IDs to training, 360 to validation, and 360 to test using split seed 12017. These sets partition player IDs 0 through 2,399 exactly and apply to both regimes. Splits are by player, not by attempt or transition.

Logged and oracle boundary

Deployable training inputs are only:

• episodes.csv, containing observed attempt summaries and the realized churn_after outcome; and
• transitions.npz, containing observed game states, legal actions, counters, served difficulty, and row identifiers.

The deployable files exclude true latent skill, true expected-experience state, oracle win propensity, and churn probability. Files under oracle/ contain simulator-only latent state and probabilities. They are for held-out evaluation and diagnostics only and must not be used to train deployable models, choose hyperparameters, or construct adjustment representations.

Attempt schema

Every episodes.csv has these 38 columns:

level, n_colours, colour_entropy, tier, move_budget, nominal_goal_count,
baseline_logit, E, served_goal_count, R, moves_used, goals_cleared, reshuffles,
cascade_depth_mean, tiles_per_move, striped_tiles_created,
striped_tiles_activated, candidate_recall_mean, pattern_noise_scale_mean,
selected_setup_value_mean, selected_goal_cleared_mean, x_search_latency,
x_candidate_recall, x_hint_count, x_pattern_error_rate,
x_immediate_pattern_precision, x_distractor_resistance,
x_lookahead_choice_rate, x_setup_value_z, x_cascade_preparation,
x_goal_clear_share, x_goals_per_move, x_moves_left_efficiency, player_id,
attempt_id, active_before, completion_margin, churn_after

Transition schema

Every transitions.npz can be loaded with numpy.load(..., allow_pickle=False) and contains:

schema_version, board_before, board_after, specials_before, specials_after,
action, action_index, moves_left, moves_left_next, goals_left, goals_left_next,
goal_colour, level, tier, served_difficulty, episode_id, step_id, player_id,
attempt_id

Boards and special-kind grids are [rows, 8, 8]; actions are [rows, 4]; all other row-level arrays are one-dimensional. See each shard's manifest.json for exact row counts and SHA-256 digests.

Oracle schema

Every oracle/attempts.csv has these 11 columns:

player_id, attempt_id, mastery_before, mastery_after, completion_margin,
oracle_win_probability, churn_probability, k_search, k_pattern, k_planning,
k_strategy

Loading a pinned snapshot

Install huggingface_hub, pandas, and numpy. The revision below is the verified immutable commit containing the 84-file accepted payload. Confirmatory runs should pin this full Hub commit SHA rather than use main.

from pathlib import Path

import numpy as np
import pandas as pd
from huggingface_hub import snapshot_download

release = Path(
    snapshot_download(
      repo_id="osazuwa/wrong-move-reference-v1",
        repo_type="dataset",
      revision="8ad9a451109f6313ab4ed1f7795e1531a446d367",
    )
)

attempts = pd.read_csv(release / "natural/shard-000/episodes.csv")
with np.load(
    release / "natural/shard-000/transitions.npz", allow_pickle=False
) as transitions:
    boards = transitions["board_before"]

Regimes and provenance

Generation seed: 12011. Maximum attempts per player: 20. Release generator code SHA: e4f175e9e6c471e5096e86df01800c7038144372. The accepted benchmark records simulator revision 0c56017ac3a1cad42d551e35cc790bb7c8393a24 and its frozen calibration hashes in accepted_benchmark.json.

The natural regime uses level-specific skill gains [1.9, 2.0, 1.8] and noise standard deviations [0.8, 0.8, 0.8]. The randomized regime sets every skill gain to zero and uses noise standard deviations [2.0615528128, 2.1540659229, 1.9697715604] to preserve natural marginal assignment variance conditional on level and baseline tier.

The release was accepted retrospectively. The accepted benchmark reports that 15 of 15 held-out cells reproduce the recommendation reversal and positive causal regret; 13 of 15 pass the full conservative auxiliary audit. These facts make the release a controlled synthetic benchmark, not evidence about real players or deployed games.

Integrity

Treat the Hub commit SHA as the dataset version. release-manifest.json hashes the root metadata and every shard manifest; each shard manifest hashes its logged and oracle artifacts. Verify this complete chain before training. Do not silently repair, rewrite, concatenate, or reshard the accepted payload.

No license metadata is declared because the source repository contains no license file. Confirm reuse terms with the dataset owner before redistribution.