Growth since the 6-hour pilot
Normalized comparison between the pilot snapshot and the latest 40.71-hour extract.
All result rows remain unique; no duplicate or infrastructure-error growth accompanied this scale-up.
A CPU-first research arena for interactive reasoning: it watches unfamiliar game worlds, builds compact causal hypotheses, chooses informative legal actions, learns from state changes, and searches for better reasoning policies over many resumable experiments.
Imagine giving an AI a small unfamiliar video game with no instructions. It can see the screen and press legal controls, but it is not told what the objects mean, what the goal is, or how the world reacts. It must learn by carefully trying things.
The benchmark presents new interactive environments. A capable agent must explore, infer the goal, learn the rules, remember what happened, and act efficiently.
Minimum Description Length rewards a compact model that explains more observations, predicts what will happen next, and needs fewer exceptions. Here “compression” means a shorter causal explanation—not simply a smaller image file.
The Digital Claustrum Controller decides which perception, probing, planning, memory, or repair specialist deserves compute right now. It can reduce, redirect, or stop unhelpful work.
Thousands of candidate policies are evaluated on matched internal worlds. Better or interesting specialists are retained; weak methods are ranked low rather than permanently erased, so a later representation can make them useful again.
The frozen v0.5.5 R2 agent used the official SDK/API environment, visited all 25 public games, and completed one level in lp85. Gate D finished with clean integrity, zero illegal executions and zero errors.
This is real public-game evidence, but it is not a Kaggle leaderboard score and it is not evidence that the system solves ARC-AGI-3 generally.
The multi-day search evaluates candidate reasoning policies on historical internal families and six synthetic causal programs. The run contract explicitly forbids final-structural, official-public, and private-answer feedback from driving candidate selection.
Therefore 102/102 on a Forge panel is an internal development result, never an official ARC percentage.
| Stage | Result | What it establishes |
|---|---|---|
| Gate A | PASS · 313 tests | 0 failures, 0 errors; baseline, storage, legality and secret checks passed. |
| Fresh smoke | PASS | Fresh Windows C5-versus-C6 smoke recomputed rather than resumed from shipped rows. |
| Gate B | PASS · 574/576 | Historical C6 compatibility and a calibration gain of +0.8403 with interval [0.8056, 0.8750]. |
| ACTION6 gate | PASS · 35 files | Coordinate actions are parameter-complete; the earlier click-action crash is closed. |
| Official Gate D | PASS · 25/25 games | One public level completed; integrity CLEAN; 0 illegal; 0 errors. |
| Long Forge | RUNNING | 899,028 unique result rows, 2,543 archived candidates, 0 duplicate rows, 0 infrastructure errors. |
The live extract is internally consistent and hash-manifested. A small snapshot timing caveat remains: the checkpoint and generation report can straddle one generation boundary while every captured file remains individually complete and verified.
030fc719c5c4c2e09803 still holds the best stored score, 7.461517. After roughly 40.7 Forge hours and 2,543 archived candidates, no later policy has exceeded it. More generations are producing useful ablations, but not a new global breakthrough.
148614318096c3501843, created by switching mdl_action_risk from 0 to 1, reached a best score of 7.458517—only 0.003 below the champion. This is promising, although not yet a stable superiority result.
The active population still contains every representation mode, but mode 1 now holds 10 of 18 places. Diversity survives, while selection is visibly concentrating around the current successful regime.
The long plateau suggests that the next progress will come from a new capability or challenge family—especially rotational invariance and latent delayed effects—not simply from extending the same search objective indefinitely.
The following counts describe the best 100 archived development candidates. They are strong associations on the current internal panel. The matched ablations provide the more causal evidence.
| Mechanism | Present in top 100 | Latest matched evidence | Interpretation |
|---|---|---|---|
| DCC mode 3 | 100 / 100 | 95 solved → 52 with mode 1; 95 → 50 with mode 2 | Strongest current routing signal; changing it causes a large collapse. |
| Temporal graph | 100 / 100 | Consistently retained | Tracking events and state transitions is central to the surviving policies. |
| Transform recovery v2 | 100 / 100 | 95 → 89 and 100 → 94 when disabled | Clear positive evidence for recovering from wrong transform hypotheses. |
| Executable DSL | 99 / 100 | Earlier ablations concentrate failures in semantic repair | Executable models appear to help more than prose-like or non-executable hypotheses. |
| Diagnostic probing | 95 / 100 | Widely retained | Information-seeking actions are useful, though probe policy still needs broader validation. |
| Cross-level capsules | 93 / 100 | Widely retained | Compact transfer of verified mechanisms across levels is supported. |
| Online RDE policy | 0 / 100 | 100 → 97 in a matched recent mutation; more actions and compute | The archive principle is valuable, but the current online resurrection routing is not. |
| Predictive state / VOC | 0 / 100 | No positive promotion signal yet | Not disproven universally; the current panel and implementation do not reward them. |
In reproduced failure seeds, a rot180 presentation was interpreted as “risk,” routed to the frozen C5 specialist, and fell back to repeated ACTION5 instead of exploring legal coordinate clicks. Other flips of the same causal world were routed to the causal planner and solved.
General fix: D4-canonical perception and a contract-aware rule that preserves coordinate exploration when a legal coordinate action exists.
Some worlds require the same correct action four times: the first actions appear to do nothing, and only the later repetition reveals the effect. The current agent repeats it three times, marks it exhausted, then spends the remaining budget on another action.
General fix: a bounded run-length hypothesis: repeated semantic no-ops may be evidence of a delayed threshold, not proof that the action is wrong.
All 2,543 candidates remain available. However, the 163 recorded resurrection events concentrate heavily:
The same old champion was resurrected 87 times, and the action-risk candidate 25 times. The next selector should add cooldown, novelty, uncertainty, and rotation over method × representation × regime cells instead of repeatedly selecting the same high-ranked dormant method.
The full archive remains wider: modes 0–5 contain 414, 866, 336, 285, 320 and 322 candidates respectively. This is a useful reservoir for future representation-triggered re-evaluation.
Compact episode records at approximately 838.8 bytes per result row.
Exactly aligned with 899,028 unique result keys.
Linear projection of shards and index to the full 14-day target.
Includes a rough trace allowance; this remains an estimate because full trace blobs are intentionally excluded from the live extract.
The previous compact refresh described the evidence but removed the visual layer. This dashboard restores a visual reading of the current v0.5.5 R2 run, while the full historical ARC-grid atlas is preserved directly below.
Normalized comparison between the pilot snapshot and the latest 40.71-hour extract.
Association within the current internal development archive; matched ablations remain the stronger causal evidence.
All six lenses survive, although mode 1 now occupies more than half of the active population.
The long run is now producing specific architectural diagnoses rather than merely more candidate counts.
40 runs, 5,284 run-task evaluations and 39 unique train-fit exact tasks. It established that category × primitive-family routing mattered more than blind global composition.
Interactive mechanisms, staged Forge search, durable resume, matched seeds, budget enforcement, failure mining and compact live extracts became operational.
Added causal observation envelopes, temporal scene graphs, executable transition models, probing, planning, memory scopes, repair, legal-action projection and a dormant-method archive.
Closed ACTION6 parameter bugs, hardened narrow-grid handling, aligned authorization and run contracts, made resume source-aware, reduced storage, and passed the first clean 25-game official-public diagnostic with non-zero capability.
The complete v0.4 visual atlas and analytical record are embedded below, clearly labelled as historical evidence.
The three large input → target → prediction visual cards, seven historical tables, chain2 analysis, primitive-family evidence, failure buckets and reproducibility notes are preserved rather than summarized away.
The local arc_report.html pages are valuable because they show train input, target, winner prediction, exact/near-miss status, categories, candidate counts, and failure buckets. Instead of embedding every run report, the v0.4 package should carry a compact atlas plus links back to per-run reports.
00d62c1b62ab2642007bbfb7Package rule: place the generated key bundle beside this page as arc_v04_key_bundle/. Keep the huge original out_arc_v04_overnight_w10 outside the MDLxDCC package unless needed locally.
| Run | Cat | Family cat | Tasks | Exact | HIGH | LOO avg | Exact/hour | Runtime | Chain2 exact | Exact families |
|---|---|---|---|---|---|---|---|---|---|---|
28_cat_symmetry_wide | symmetry | all | 160 | 14 | 12 | 0.091 | 6.5 | 2:09:37 | 0 | erode_color:1; fill_holes:2; line_connect:1; object_to_marker:1; outline_foreground:1; ray_cast:1; recolor_by_map:1; scale_up_integer:1 |
32_cat_easy_signature_wide | easy_signature | all | 160 | 14 | 12 | 0.084 | 150.5 | 0:05:35 | 0 | erode_color:1; global_color_replace:1; line_connect:1; recolor_by_map:2; remove_background:1; rotate:1; scale_up_integer:2; substitution_expand:1 |
21_cat_macro_wide | macro | all | 160 | 14 | 11 | 0.082 | 6.5 | 2:08:16 | 0 | crop_largest_object:1; crop_non_bg_bbox:1; diagonal_connect:1; extract_object:1; fill_holes:1; gravity_pack:1; line_connect:2; recolor_by_map:1 |
15_cat_crop_like_wide | crop_like | all | 160 | 11 | 10 | 0.080 | 1.9 | 5:40:03 | 0 | block_compress:1; compress_blank_rows_cols:1; crop_largest_object:2; crop_non_bg_bbox:1; extract_object:1; pad_resize:1; remove_background:1; tile_repeat:3 |
11_cat_in_canvas_wide | in_canvas | all | 160 | 11 | 9 | 0.066 | 2.6 | 4:12:26 | 0 | diagonal_connect:1; fill_holes:1; gravity_pack:1; line_connect:2; outline_foreground:1; recolor_by_map:1; rotate:1; symmetry_complete:1 |
10_cat_same_shape_wide | same_shape | all | 160 | 11 | 9 | 0.066 | 2.4 | 4:36:48 | 0 | diagonal_connect:1; fill_holes:1; gravity_pack:1; line_connect:2; outline_foreground:1; recolor_by_map:1; rotate:1; symmetry_complete:1 |
45_paint_family_color_morph | recolor_or_paint | color,morphology,line | 220 | 11 | 9 | 0.050 | 5.0 | 2:12:52 | 2 | chain2:2; diagonal_connect:1; erode_color:1; fill_holes:1; line_connect:2; outline_foreground:1; ray_cast:1; recolor_by_map:1 |
17_cat_sequence_wide | sequence | all | 160 | 10 | 9 | 0.064 | 3.9 | 2:35:35 | 0 | crop_largest_object:1; crop_non_bg_bbox:1; extract_object:1; fill_holes:1; gravity_pack:1; line_connect:1; recolor_by_map:1; substitution_expand:1 |
24_cat_frame_wide | frame | all | 160 | 10 | 9 | 0.063 | 4.7 | 2:08:05 | 0 | block_compress:1; crop_largest_object:1; crop_non_bg_bbox:1; extract_object:1; outline_foreground:1; substitution_expand:2; tile_repeat:3 |
51_crop_family_core_object | crop_like | core,object | 186 | 10 | 8 | 0.060 | 107.8 | 0:05:34 | 1 | chain2:1; crop_largest_object:2; crop_non_bg_bbox:1; extract_object:1; pad_resize:1; remove_background:1; tile_repeat:3 |
23_cat_multicolor_object_wide | multicolor_object | all | 160 | 10 | 8 | 0.057 | 4.4 | 2:15:57 | 0 | block_compress:1; crop_largest_object:1; extract_object:1; gravity_pack:1; pad_resize:1; recolor_by_map:1; rotate:1; scale_up_integer:1 |
30_cat_hard_wide | hard | all | 120 | 8 | 7 | 0.066 | 1.2 | 6:34:19 | 0 | block_compress:1; compress_blank_rows_cols:1; crop_largest_object:1; extract_object:1; pad_resize:1; tile_repeat:2; wire_connect:1 |
20_cat_recolor_or_paint_wide | recolor_or_paint | all | 160 | 8 | 6 | 0.047 | 3.7 | 2:10:09 | 0 | diagonal_connect:1; fill_holes:1; line_connect:2; outline_foreground:1; recolor_by_map:1; symmetry_complete:1; wire_connect:1 |
25_cat_holes_wide | holes | all | 160 | 6 | 6 | 0.048 | 2.1 | 2:54:49 | 0 | crop_largest_object:1; extract_object:1; fill_holes:1; mirror:1; pad_resize:1; tile_repeat:1 |
| Run | Cat | Family cat | Tasks | Exact | HIGH | LOO avg | Exact/hour | Runtime | Chain2 exact | Exact families |
|---|---|---|---|---|---|---|---|---|---|---|
43_expand_family_macro | expand | macro | 68 | 4 | 4 | 0.070 | 348.0 | 0:00:41 | 0 | scale_up_integer:2; substitution_expand:2 |
49_motion_family_motion_object | motion | motion,object | 19 | 2 | 2 | 0.105 | 346.6 | 0:00:21 | 0 | extract_multicolor_object:1; translate_object:1 |
60_full_integer_expand_macro | integer_expand | macro | 38 | 4 | 4 | 0.113 | 167.9 | 0:01:26 | 0 | scale_up_integer:2; substitution_expand:2 |
14_cat_integer_expand_wide | integer_expand | all | 38 | 4 | 4 | 0.121 | 154.4 | 0:01:33 | 0 | scale_up_integer:2; substitution_expand:2 |
32_cat_easy_signature_wide | easy_signature | all | 160 | 14 | 12 | 0.084 | 150.5 | 0:05:35 | 0 | erode_color:1; global_color_replace:1; line_connect:1; recolor_by_map:2; remove_background:1; rotate:1; scale_up_integer:2; substitution_expand:1 |
51_crop_family_core_object | crop_like | core,object | 186 | 10 | 8 | 0.060 | 107.8 | 0:05:34 | 1 | chain2:1; crop_largest_object:2; crop_non_bg_bbox:1; extract_object:1; pad_resize:1; remove_background:1; tile_repeat:3 |
27_cat_motion_wide | motion | all | 19 | 5 | 5 | 0.307 | 68.4 | 0:04:23 | 0 | crop_non_bg_bbox:1; mirror:1; remove_background:1; rotate:1; translate_object:1 |
44_color_family_color | color | color | 220 | 2 | 1 | 0.007 | 51.8 | 0:02:19 | 1 | chain2:1; recolor_by_map:1 |
46_object_family_object | object | object | 220 | 4 | 3 | 0.025 | 49.6 | 0:04:50 | 0 | extract_multicolor_object:2; extract_object:1; translate_object:1 |
48_morphology_family_morph | morphology | morphology | 220 | 3 | 2 | 0.017 | 45.0 | 0:04:00 | 1 | chain2:1; fill_holes:2 |
47_multicolor_family_object_color | multicolor_object | object,color | 220 | 3 | 1 | 0.009 | 33.1 | 0:05:26 | 0 | extract_multicolor_object:1; extract_object:1; recolor_by_map:1 |
62_full_color_family | color | color | 80 | 1 | 0 | 0.006 | 26.8 | 0:02:14 | 0 | recolor_by_map:1 |
32_cat_easy_signature_wide gives 14 exact in 5:35; 51_crop_family_core_object gives 10 exact in 5:34; 43_expand_family_macro gives 4 exact in 41 seconds. These are the lanes v0.5 should remember and route toward first.The 50-task global baseline with chain2 and the no-chain2 ablation both produced 4 exact / 3 HIGH. The no-chain2 run used roughly half the runtime. That keeps chain2 out of the global default.
But routed chain2 produced 5 exact rows across 5 unique tasks, with 3 robust LOO=1.0 rows. The useful cases are mostly color/morph/crop compositions such as line_connect → component_recolor or fill_holes → global_color_replace.
--chain2-policy off|always|routed|exact-only. Default should be routed, with category evidence and per-family caps.| Exact win family | Rows |
|---|---|
tile_repeat | 20 |
fill_holes | 17 |
substitution_expand | 16 |
line_connect | 16 |
recolor_by_map | 15 |
wire_connect | 14 |
scale_up_integer | 12 |
crop_largest_object | 12 |
extract_object | 11 |
translate_object | 9 |
diagonal_connect | 8 |
outline_foreground | 7 |
gravity_pack | 6 |
symmetry_complete | 6 |
crop_non_bg_bbox | 6 |
pad_resize | 6 |
| Near-miss family | Rows |
|---|---|
global_color_replace | 747 |
identity | 690 |
pad_resize | 537 |
line_connect | 341 |
erode_color | 240 |
fill_holes | 232 |
translate_object | 187 |
diagonal_connect | 184 |
frame_extract | 172 |
crop_color_bbox | 170 |
row_col_project | 156 |
tile_repeat | 135 |
extract_multicolor_object | 125 |
crop_largest_object | 103 |
component_recolor | 91 |
chain2 | 82 |
Row counts are not unique-task counts because the same task can appear in several category slices. Still, the repeated winners are informative: tile_repeat, fill_holes, substitution_expand, line_connect, recolor_by_map, wire_connect, scale_up_integer, crop_largest_object, and extract_object are the current strongest families.
Across the 40-run bundle, v0.4 finds 39 unique exact train-fit task ids; 33 have best LOO=1.0. Baseline global runs found 4; the rest came from slicing and routing.
| First run that found task | New exact task ids | Robust LOO=1.0 | Task ids |
|---|---|---|---|
01_dev50_wide_all | 4 | 3 | 007bbfb7, 00d62c1b, 070dd51e, 0d3d703e |
10_cat_same_shape_wide | 8 | 7 | 1e0a9b12, 1f876c06, 22168020, 22eb0ac0, 25ff71a9, 3c9b0459, 4347f46a, 496994bd |
13_cat_expand_wide | 3 | 3 | 5b6cbef5, 60c09cac, c59eb873 |
15_cat_crop_like_wide | 11 | 10 | 1cf80156, 1f85a75f, 2013d3e2, 23b5c85d, 28bf18c6, 5614dbcf, 5bd6f4ac, 68b67ca3, 73182012, a740d043, be94b721 |
25_cat_holes_wide | 1 | 1 | 67a3c6ac |
28_cat_symmetry_wide | 4 | 4 | 623ea044, 6f8cd79b, 88a10436, a5313dff |
32_cat_easy_signature_wide | 3 | 2 | 9dfd6313, aabf363d, c8f0f002 |
44_color_family_color | 1 | 1 | 62ab2642 |
45_paint_family_color_morph | 2 | 2 | 0d87d2a6, 7b6016b9 |
48_morphology_family_morph | 1 | 0 | 91714a58 |
51_crop_family_core_object | 1 | 0 | b9b7f026 |
| Failure bucket | Rows |
|---|---|
needs_depth2_composition | 4985 |
near_miss_small_rule_gap | 2801 |
needs_natural_law_primitive | 1795 |
needs_in_canvas_rule | 1180 |
needs_relational_object_reasoning | 1086 |
needs_wire_crossing_or_signal_priority | 302 |
needs_crop_then_recolor_or_mask_cleanup | 220 |
needs_substitution_or_morphogenesis | 206 |
needs_crop_resize_or_canvas_transform | 137 |
unknown_or_small_missing_step | 25 |
needs_transform_then_crop_or_recolor | 17 |
| Run | Cat | Family cat | Tasks | Exact | HIGH | LOO avg | Exact/hour | Runtime | Chain2 exact | Exact families |
|---|---|---|---|---|---|---|---|---|---|---|
81_holdout_line_wide | line | line | 100 | 0 | 0 | 0.003 | 0.0 | 3:51:30 | 0 | |
41_sequence_family_line | sequence | line | 220 | 1 | 1 | 0.006 | 0.4 | 2:49:12 | 0 | line_connect:1 |
52_holes_family_morph_natural | holes | morphology,natural | 220 | 2 | 2 | 0.018 | 0.4 | 5:00:51 | 0 | fill_holes:2 |
31_cat_slow_risk_quick | slow_risk | all | 116 | 2 | 2 | 0.022 | 0.6 | 3:29:37 | 0 | crop_largest_object:1; wire_connect:1 |
30_cat_hard_wide | hard | all | 120 | 8 | 7 | 0.066 | 1.2 | 6:34:19 | 0 | block_compress:1; compress_blank_rows_cols:1; crop_largest_object:1; extract_object:1; pad_resize:1; tile_repeat:2; wire_connect:1 |
40_line_family_line | line | line | 220 | 4 | 3 | 0.019 | 1.4 | 2:50:41 | 0 | diagonal_connect:1; line_connect:2; wire_connect:1 |
61_full_line_family | line | line | 80 | 2 | 2 | 0.025 | 1.9 | 1:03:26 | 0 | diagonal_connect:1; wire_connect:1 |
15_cat_crop_like_wide | crop_like | all | 160 | 11 | 10 | 0.080 | 1.9 | 5:40:03 | 0 | block_compress:1; compress_blank_rows_cols:1; crop_largest_object:2; crop_non_bg_bbox:1; extract_object:1; pad_resize:1; remove_background:1; tile_repeat:3 |
26_cat_morphology_wide | morphology | all | 160 | 6 | 6 | 0.048 | 2.0 | 2:57:28 | 0 | crop_largest_object:1; extract_object:1; fill_holes:1; mirror:1; pad_resize:1; tile_repeat:1 |
25_cat_holes_wide | holes | all | 160 | 6 | 6 | 0.048 | 2.1 | 2:54:49 | 0 | crop_largest_object:1; extract_object:1; fill_holes:1; mirror:1; pad_resize:1; tile_repeat:1 |
needs_depth2_composition is too broad. It appears almost everywhere and must be split into object matching, macro-grid, line completion, marker instruction, crop/resize, color-role transfer, local cellular rule, and canvas transformation buckets. identity and global_color_replace near-misses are also too noisy and should be treated as diagnostic hints, not strong routes.
arc_az.json and optional task cache storing task signature hash, categories, winning families, failed families, near-misses, runtime cost, and best candidate summaries.--cat-report, and --chain2-policy routed; skip expensive lanes when shape evidence says impossible.--cat line,color --cat-mode all and not:slow_risk.exact_but_loo_bad, identity_fallback_suspect, and near_miss_small_rule_gap subtypes.arc_report.html inside the MDLxDCC package.Canonical source files remain the arena code, batch file, summarizer, and JSON outputs. The compact key bundle is the MDLxDCC-facing artifact.
python arc_mdlxdcc_arena_v0_4.py --data-dir D:\8Z\8z\ARC\ARC-AGI-2 --split dev --max-tasks 50 --outdir out_arc_v04_quick_w10 --mode both --fresh --budget wide --workers 10 --progress-every 5
python summarize_arc_overnight.py out_arc_v04_overnight_w10
python arc_collect_keydata_v04.py out_arc_v04_overnight_w10 --out out_arc_v04_keybundleExpected local package link target: arc_v04_key_bundle/arc_visual_atlas.html, arc_v04_key_bundle/arc_report_hub.html, arc_v04_key_bundle/run_index.csv, and arc_v04_key_bundle/task_index.csv.
These links point directly to ARC Prize Foundation, Kaggle and the official open-source toolkit.