Dataset Preview
Duplicate
The full dataset viewer is not available (click to read why). Only showing a preview of the rows.
The dataset generation failed
Error code:   DatasetGenerationError
Exception:    TypeError
Message:      Couldn't cast array of type
struct<name: string, x: int64, y: int64>
to
{'name': Value('string')}
Traceback:    Traceback (most recent call last):
                File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1827, in _prepare_split_single
                  for key, table in generator:
                                    ^^^^^^^^^
                File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 613, in wrapped
                  for item in generator(*args, **kwargs):
                              ~~~~~~~~~^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 343, in _generate_tables
                  self._cast_table(pa_table, json_field_paths=json_field_paths),
                  ~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 132, in _cast_table
                  pa_table = table_cast(pa_table, self.info.features.arrow_schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2378, in table_cast
                  return cast_table_to_schema(table, schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2312, in cast_table_to_schema
                  cast_array_to_feature(
                  ~~~~~~~~~~~~~~~~~~~~~^
                      table[name] if name in table_column_names else pa.array([None] * len(table), type=schema.field(name).type),
                      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                      feature,
                      ^^^^^^^^
                  )
                  ^
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 1861, in wrapper
                  return pa.chunked_array([func(chunk, *args, **kwargs) for chunk in array.chunks])
                                           ~~~~^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2158, in cast_array_to_feature
                  raise TypeError(f"Couldn't cast array of type\n{_short_str(array.type)}\nto\n{_short_str(feature)}")
              TypeError: Couldn't cast array of type
              struct<name: string, x: int64, y: int64>
              to
              {'name': Value('string')}
              
              The above exception was the direct cause of the following exception:
              
              Traceback (most recent call last):
                File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 1369, in compute_config_parquet_and_info_response
                  parquet_operations, partial, estimated_dataset_info = stream_convert_to_parquet(
                                                                        ~~~~~~~~~~~~~~~~~~~~~~~~~^
                      builder, max_dataset_size_bytes=max_dataset_size_bytes
                      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                  )
                  ^
                File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 948, in stream_convert_to_parquet
                  builder._prepare_split(split_generator=splits_generators[split], file_format="parquet")
                  ~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1694, in _prepare_split
                  for job_id, done, content in self._prepare_split_single(
                                               ~~~~~~~~~~~~~~~~~~~~~~~~~~^
                      gen_kwargs=gen_kwargs, job_id=job_id, **_prepare_split_args
                      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                  ):
                  ^
                File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1880, in _prepare_split_single
                  raise DatasetGenerationError("An error occurred while generating the dataset") from e
              datasets.exceptions.DatasetGenerationError: An error occurred while generating the dataset

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

model
string
game
string
i
int64
ts
float64
action
dict
reset
bool
full_reset
bool
prev_level
int64
level
int64
level_up
bool
state
string
win
bool
game_over
bool
win_levels
int64
available_actions
list
n_anim_frames
int64
grid
list
prev_grid_changed_cells
int64
release-gpt
ar25
0
1,787,955,946.114
{ "name": "RESET" }
true
true
0
0
false
NOT_FINISHED
false
false
8
[ 1, 2, 3, 4, 5, 6, 7 ]
1
[ [ 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 10, 10, 10, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, ...
null
release-gpt
ar25
1
1,787,956,041.517
{ "name": "ACTION1" }
false
false
0
0
false
NOT_FINISHED
false
false
8
[ 1, 2, 3, 4, 5, 6, 7 ]
1
[ [ 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 10, 10, 10, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, ...
109
release-gpt
ar25
2
1,787,956,135.713
{ "name": "ACTION2" }
false
false
0
0
false
NOT_FINISHED
false
false
8
[ 1, 2, 3, 4, 5, 6, 7 ]
1
[ [ 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 10, 10, 10, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, ...
109
release-gpt
ar25
3
1,787,956,209.208
{ "name": "ACTION3" }
false
false
0
0
false
NOT_FINISHED
false
false
8
[ 1, 2, 3, 4, 5, 6, 7 ]
1
[ [ 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 10, 10, 10, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, ...
109
release-gpt
ar25
4
1,787,956,217.249
{ "name": "ACTION2" }
false
false
0
0
false
NOT_FINISHED
false
false
8
[ 1, 2, 3, 4, 5, 6, 7 ]
1
[ [ 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 10, 10, 10, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, ...
109
release-gpt
ar25
5
1,787,956,217.251
{ "name": "ACTION2" }
false
false
0
0
false
NOT_FINISHED
false
false
8
[ 1, 2, 3, 4, 5, 6, 7 ]
1
[ [ 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 10, 10, 10, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, ...
109
release-gpt
ar25
6
1,787,956,217.253
{ "name": "ACTION2" }
false
false
0
0
false
NOT_FINISHED
false
false
8
[ 1, 2, 3, 4, 5, 6, 7 ]
1
[ [ 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 10, 10, 10, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, ...
109
release-gpt
ar25
7
1,787,956,217.255
{ "name": "ACTION2" }
false
false
0
0
false
NOT_FINISHED
false
false
8
[ 1, 2, 3, 4, 5, 6, 7 ]
1
[ [ 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 10, 10, 10, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, ...
109
release-gpt
ar25
8
1,787,956,217.257
{ "name": "ACTION2" }
false
false
0
0
false
NOT_FINISHED
false
false
8
[ 1, 2, 3, 4, 5, 6, 7 ]
1
[ [ 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 10, 10, 10, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, ...
109
release-gpt
ar25
9
1,787,956,217.259
{ "name": "ACTION2" }
false
false
0
0
false
NOT_FINISHED
false
false
8
[ 1, 2, 3, 4, 5, 6, 7 ]
1
[ [ 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 10, 10, 10, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, ...
109
release-gpt
ar25
10
1,787,956,217.261
{ "name": "ACTION2" }
false
false
0
0
false
NOT_FINISHED
false
false
8
[ 1, 2, 3, 4, 5, 6, 7 ]
1
[ [ 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 10, 10, 10, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, ...
109
release-gpt
ar25
11
1,787,956,217.262
{ "name": "ACTION2" }
false
false
0
0
false
NOT_FINISHED
false
false
8
[ 1, 2, 3, 4, 5, 6, 7 ]
1
[ [ 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 10, 10, 10, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, ...
109
release-gpt
ar25
12
1,787,956,217.264
{ "name": "ACTION2" }
false
false
0
0
false
NOT_FINISHED
false
false
8
[ 1, 2, 3, 4, 5, 6, 7 ]
1
[ [ 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 10, 10, 10, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, ...
109
release-gpt
ar25
13
1,787,956,217.266
{ "name": "ACTION2" }
false
false
0
0
false
NOT_FINISHED
false
false
8
[ 1, 2, 3, 4, 5, 6, 7 ]
1
[ [ 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 10, 10, 10, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, ...
109
release-gpt
ar25
14
1,787,956,221.269
{ "name": "ACTION3" }
false
false
0
0
false
NOT_FINISHED
false
false
8
[ 1, 2, 3, 4, 5, 6, 7 ]
1
[ [ 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 10, 10, 10, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, ...
109
release-gpt
ar25
15
1,787,956,241.273
{ "name": "ACTION3" }
false
false
0
0
false
NOT_FINISHED
false
false
8
[ 1, 2, 3, 4, 5, 6, 7 ]
1
[ [ 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 10, 10, 10, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, ...
108
release-gpt
ar25
16
1,787,956,299.299
{ "name": "ACTION3" }
false
false
0
0
false
NOT_FINISHED
false
false
8
[ 1, 2, 3, 4, 5, 6, 7 ]
1
[ [ 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 10, 10, 10, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, ...
108
release-gpt
ar25
17
1,787,956,299.302
{ "name": "ACTION3" }
false
false
0
1
true
NOT_FINISHED
false
false
8
[ 1, 2, 3, 4, 5, 6, 7 ]
3
[ [ 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 10, 10, 10, 9, 9, 9, 9, 9, ...
710
release-gpt
ar25
18
1,787,956,379.202
{ "name": "ACTION4" }
false
false
1
1
false
NOT_FINISHED
false
false
8
[ 1, 2, 3, 4, 5, 6, 7 ]
1
[ [ 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 10, 10, 10, 9, 9, ...
487
release-gpt
ar25
19
1,787,956,567.862
{ "name": "ACTION5" }
false
false
1
1
false
NOT_FINISHED
false
false
8
[ 1, 2, 3, 4, 5, 6, 7 ]
1
[ [ 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 10, 10, 10, 9, 9, ...
30
release-gpt
ar25
20
1,787,956,753.508
{ "name": "ACTION2" }
false
false
1
1
false
NOT_FINISHED
false
false
8
[ 1, 2, 3, 4, 5, 6, 7 ]
1
[ [ 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 10, 10, 10, 9, 9, ...
181
release-gpt
ar25
21
1,787,956,753.51
{ "name": "ACTION2" }
false
false
1
1
false
NOT_FINISHED
false
false
8
[ 1, 2, 3, 4, 5, 6, 7 ]
1
[ [ 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 10, 10, 10, 9, 9, ...
181
End of preview.

Kepler 1.0 ARC-AGI-3 trace corpus

Run artifacts from Kepler 1.0, an open-source agent harness for the 25 public ARC-AGI-3 games. A stock CLI coding agent encodes its theory of each game as an executable world_model.py, certifies it against the full recorded interaction history, plans inside the certified model, and acts through a guarded channel that voids the plan on the first misprediction.

Project page · Code · Paper · Integrity record

The canonical release contains two single-configuration boards:

  • Claude Opus 5: 100.00. One frozen configuration, one run per game, no score-conditioned reruns. ARC Prize's official server replay re-executed all 25 games to 100. Scorecard: 91aa2f10.
  • GPT-5.6 Sol (max): 95.97. One frozen configuration, with the two non-perfect games retained. Scorecard: c9f087f3.

Across these two frozen release boards, 48 of 50 game-model cells reached 100. On the Opus board, 181 of 183 completed levels used no more actions than the corresponding median-human baseline. This is final-attempt action efficiency, not learning cost or evidence of human-like cognition.

Token and cost accounting was recovered from provider-side session records retained locally on the execution host and deduplicated by provider message ID. Those provider records are not part of this dataset; only the captured CLI session logs are. They put the Opus campaign at 858,041,926 raw tokens, 97.37% cache reads, and $777.72 at current API list-equivalent rates. That is 74.0% below the $2,986 API-equivalent estimate Retrodict published for Tycho. Tycho discloses no cost of its own, and neither do AVO or VISTA, so this is a comparison against one third-party estimate and not a ranking of the field. Retrodict's lower-scoring 99.86 costs less at $654. The GPT board is $1,312.14 current API list-equivalent.

The retained Opus board runs used 8,256 environment actions, with 7,292 in the original local scored-level results and 7,202 on ARC's public replay card. The GPT equivalents are 8,400 and 8,220. Replay selects the last full-reset-to-end ledger segment and counts a new opening reset, so these are different recorded denominators rather than score disagreements. Full campaign logs contain at least 13,688 non-reset actions and omit 22 prefix events, so 8,256 is not labeled learning-inclusive. None of these counts is ranked against other systems, whose action counts cover different stages and definitions.

Public-set scores are not a measure of AGI progress. The ARC-AGI-3 technical report (§4.3.1) says so explicitly and ships a human-replay harness scoring 100% to make the point. This dataset supports a reproduction question: does a published result hold up? It does not support a general capability claim.

The ablation is retracted. Read this before citing anything about the harness's contribution. We ran the same model with the methodology stripped out, but the control workspaces lived inside the repository, so every one of the six baseline agents found harness/ws_tools/ on disk and rebuilt the methodology they were meant to be a control for (9,148 tool invocations on the worst one). That measured harness against harness. Every conclusion drawn from it is withdrawn, including a "net advantage is roughly zero" headline we had published. When historical roots are selected, those runs appear under baseline-* labels so the contamination is inspectable. The harness's contribution is currently unmeasured, not small and not large.

The source repository separately preserves a source-reading incident that produced a natural-looking 100 before scoring 46.91 in a clean rerun, plus evidence that a bundled planner was broken across five experimental boards while agents silently routed around it. High outcome scores therefore do not establish tool health.

This dataset is narrower than the full development archive. It contains exactly the two final 25-game boards. Earlier stages, failed experiments, quarantined incidents, and superseded runs remain documented in the source repository and are not part of this 50-run export.

Layout

  • runs.jsonl, one row per (run_root, model, game) run.
  • events/<label>/<game>.jsonl.gz, one row per recorded environment transition of that run, in order. This is the append-only ground-truth ledger written by the harness daemon; the agent could read but never write it.
  • agent_logs/<label>/<game>/sessions/*.log.gz, the harness's captured CLI output: 95 gzipped session logs, the only agent behavioural records in this dataset. Across this export, 50 of 50 runs have at least one retained log, but coverage depth is uneven. Some Claude tee files contain only final summaries or quota markers. These are the captured CLI records available on the execution host, not a platform attestation that every event was retained.
  • arc_agi_3_human_baseline_actions.csv, the per-level human action baselines used to recompute RHAE.
  • score_trajectories.py, verify_scores.py, and audit_integrity.py, dependency-free verification programs copied into the dataset.

Labels

label prefix meaning
release-opus canonical Claude Opus 5 board: 100.00, exact official replay
release-gpt canonical GPT-5.6 Sol board: 95.97, exact official replay

Internal run-root names are provenance identifiers, not public release versions. The public software and paper have one identity: Kepler 1.0.

runs.jsonl schema

field type description
model str exported board label, such as release-opus or release-gpt
game str 4-char public ARC-AGI-3 game id (e.g. ft09)
game_id str full versioned environment id
state str terminal state of the run (WIN, NOT_FINISHED, ...)
levels_completed / win_levels int levels cleared / levels needed to win
actions int real actions committed to the environment
elapsed_hours float wall-clock duration of the run
score float official per-game RHAE score (0..100)
level_scores / level_actions list per-level score / action count
note str how the run ended (win, budget exhaustion reason, ...)
notes_md str the agent's final lab notebook, verbatim
world_model_py str the agent's final executable world model, verbatim
scorecard json str the official local arc_agi scorecard for the run
n_events int number of rows in the matching events file
events_file str relative path to the matching events file
grids_compacted int/null --compact N used at export time, if any
run_root str internal source directory retained as provenance, not a public version
base_model str model id without the label prefix
complete bool false for aborted runs (score is null)
n_log_files / log_bytes_gz int archived agent session files and their gzipped size

events/*/*.jsonl.gz schema

One JSON object per line, fields as recorded live by the harness daemon:

field type description
model, game str added at export time (join keys to runs.jsonl)
i int event index, 0-based, contiguous (append-only ledger)
ts float unix timestamp
action obj {"name": "RESET"|"ACTION1".."ACTION7", ["x","y" for ACTION6]}
reset bool this event is a reset, not a scored action
full_reset bool reset restarted the whole game (vs the current level)
prev_level / level int levels completed before / after the event
level_up bool this action completed a level
state str NOT_FINISHED / WIN / GAME_OVER ...
win / game_over bool terminal flags
win_levels int levels needed to win the game
available_actions list[int] action ids the env accepts here
n_anim_frames int animation frames the env returned (last one is grid)
grid 64x64 list[int] or null the observed frame after the action (null only when exported with --compact)
grid_stripped bool present+true when --compact removed this row's grid
prev_grid_changed_cells int/null cells changed vs the previous grid

Verify the export

From the downloaded dataset directory:

python3 score_trajectories.py .
python3 verify_scores.py --traces-dir .
python3 audit_integrity.py --traces-dir .

score_trajectories.py independently recomputes both board scores from the event ledgers and the human baselines. verify_scores.py checks that no level was scored with fewer actions than its ledger records. audit_integrity.py scans the captured CLI session-log bytes for known leakage signatures. A clean result applies to the records present here; it does not prove that the client preserved every event or that the pattern set detects every possible violation. An export without agent_logs/ is reported as unauditable and exits nonzero.

Provenance & license

Produced by scripts/export_traces.py in the harness repository from the raw run workspaces. Environments are the 25 public ARC-AGI-3 games run locally via the arc-agi toolkit. MIT.

Generated: 2026-09-03. 50 runs, 58098 events, 95 session logs, labels: release-gpt, release-opus.

Downloads last month
141