RegalFire commited on
Commit
a6e2a06
·
verified ·
1 Parent(s): 3ba2635

Release verified agent recovery benchmark with lossless Parquet viewer, evaluation interface and full metadata

Browse files
MANIFEST.json CHANGED
@@ -1,116 +1,145 @@
1
  {
2
- "product": "Agent-Failure-Recovery-Benchmark",
3
- "created_at_utc": "2026-10-01T16:21:46.886533+00:00",
4
- "files": [
5
- {
6
- "path": "COMMERCIAL_EXTENSION.md",
7
- "bytes": 1385,
8
- "sha256": "111fc9d422ec7310dec17bf4561a87e722101301e8df36c3352a4972fafbfc4d"
9
- },
10
- {
11
- "path": "dataset.jsonl",
12
- "bytes": 52614965,
13
- "sha256": "c0a92838f16d2fb13ec372cc2b0aa32c986529837c8b350bb71e87474cafbae2"
14
- },
15
- {
16
- "path": "holdout.jsonl",
17
- "bytes": 17495594,
18
- "sha256": "4875ee7178477886f97c73e361996ca4b704c1283f31580c76327b9600180b33"
19
- },
20
- {
21
- "path": "LICENSE",
22
- "bytes": 484,
23
- "sha256": "a8036a8e172f9fcff1484d861b9d6903c6675396884d88bc308179782776c67e"
24
- },
25
- {
26
- "path": "QA_REPORT.md",
27
- "bytes": 1173,
28
- "sha256": "9e8686dd40a15431e734b84f469f968cd92fcdfedf04fb6f20c4135727cf8582"
29
- },
30
- {
31
- "path": "README.md",
32
- "bytes": 4148,
33
- "sha256": "f6252e1dba8728682bb22cf1f4cf1db8a036a8d5e20ce4499b20258deb0cacd4"
34
- },
35
- {
36
- "path": "rejected.jsonl",
37
- "bytes": 122722,
38
- "sha256": "7467f7bdcdbc9350e36e78aae7fb5bf0632580fe0eb1a642dc7b709ddd948515"
39
- },
40
- {
41
- "path": "robustness_report.json",
42
- "bytes": 11965,
43
- "sha256": "f8122cd495e3be8c4cd1e0263eb5b163600d8a068bf25f3643a854d6ded17772"
44
- },
45
- {
46
- "path": "schema.json",
47
- "bytes": 2643,
48
- "sha256": "87c8bc827e0ed2c7bc2fc185d09603de1b037f73b77ed93cf60f6d470f871660"
49
- },
50
- {
51
- "path": "SOURCE_LOCK.json",
52
- "bytes": 5370,
53
- "sha256": "63970a1a21eda958281d153e84d256201c93d97778541275ad76741c83ac9b16"
54
- },
55
- {
56
- "path": "stats.json",
57
- "bytes": 1168,
58
- "sha256": "e4aafbe217b6d3788f3fac11b4226027f9441f253901e99630f5123f0049707d"
59
- },
60
- {
61
- "path": "test.jsonl",
62
- "bytes": 291201,
63
- "sha256": "7f1d90b4c240d32796b498c2a149ec288262aa837a84ece25f984a9a2dc3f68b"
64
- },
65
- {
66
- "path": "THIRD_PARTY_LICENSES.md",
67
- "bytes": 3321,
68
- "sha256": "a8869bc8ba7fa034e60592d8f2ecd7592e9758a6aa8d3d1886bd78df733fad7e"
69
- },
70
- {
71
- "path": "train.jsonl",
72
- "bytes": 32358142,
73
- "sha256": "012442fff17b8b8f01fc22169363e76a13740e6ae33f8beb8929d27e9660d978"
74
- },
75
- {
76
- "path": "validation.jsonl",
77
- "bytes": 2470028,
78
- "sha256": "79be61f28b6f88ec93b3503a8c4cbf278fb3f9a3cbddfb6003d9c9ef263b4e4f"
79
- },
80
- {
81
- "path": "validation_report.json",
82
- "bytes": 705,
83
- "sha256": "b8e3d295b22055292a678fcca54fb5763f58c55453dc9eebe109e3ac0c2f8c08"
84
- },
85
- {
86
- "path": "validator/LICENSE",
87
- "bytes": 1087,
88
- "sha256": "7c82eb9f40f567300f2a945943695da0f6b64eb400bb267032d1aa73ff177726"
89
- },
90
- {
91
- "path": "validator/validate.py",
92
- "bytes": 16694,
93
- "sha256": "0a8b440ad924e065de0f438d4a473baf496c69ad681146962860249c599f136b"
94
- },
95
- {
96
- "path": "validator/vendor/common.py",
97
- "bytes": 7182,
98
- "sha256": "194b6e0cc92e5d4f40c0ed96f77627e61cc98b64c3b3754de594b82f6b096b81"
99
- },
100
- {
101
- "path": "validator/vendor/horizon.py",
102
- "bytes": 6434,
103
- "sha256": "c124ae50b1f1c1a2666e0aa9201edf0822c1ff684ab56f4f17d54cd219dda8e2"
104
- },
105
- {
106
- "path": "validator/vendor/nav_check.py",
107
- "bytes": 15810,
108
- "sha256": "6093b4f2e1a6f1a730398a43539f5cee665335ff48725cfe0256cbbf9bd7664c"
109
- },
110
- {
111
- "path": "validator/vendor/physics.py",
112
- "bytes": 5706,
113
- "sha256": "863b324344a53f71c5c638aec66ebda2e16ed34e8d4591ac1ac16a44d96a93c7"
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
114
  }
115
- ]
116
  }
 
1
  {
2
+ "format": "sha256",
3
+ "files": {
4
+ ".gitattributes": {
5
+ "sha256": "1827d203fd7ac780f845099f31b973ad8f9e00fc1ac2c73794a6d4ee57a22646",
6
+ "bytes": 2652
7
+ },
8
+ "COMMERCIAL_EXTENSION.md": {
9
+ "sha256": "111fc9d422ec7310dec17bf4561a87e722101301e8df36c3352a4972fafbfc4d",
10
+ "bytes": 1385
11
+ },
12
+ "dataset.jsonl": {
13
+ "sha256": "c0a92838f16d2fb13ec372cc2b0aa32c986529837c8b350bb71e87474cafbae2",
14
+ "bytes": 52614965
15
+ },
16
+ "evaluate.py": {
17
+ "sha256": "288bc2f0c3c95bf74272bbf861c73cbe79fa09f5a4dd104b28f2369335be4f28",
18
+ "bytes": 2436
19
+ },
20
+ "evaluation_smoke.json": {
21
+ "sha256": "676cf233f623a318b2ee2543feeefd178e5f9f014178d7a8cb4136eedd7ad68e",
22
+ "bytes": 246
23
+ },
24
+ "holdout.jsonl": {
25
+ "sha256": "4875ee7178477886f97c73e361996ca4b704c1283f31580c76327b9600180b33",
26
+ "bytes": 17495594
27
+ },
28
+ "LICENSE": {
29
+ "sha256": "a8036a8e172f9fcff1484d861b9d6903c6675396884d88bc308179782776c67e",
30
+ "bytes": 484
31
+ },
32
+ "PACKAGE_PROVENANCE.json": {
33
+ "sha256": "cb46d39c42d469ca0aec493d4cee25dfbce817ea0dd24ce080ff229fa5d7bc60",
34
+ "bytes": 498
35
+ },
36
+ "package_validation.json": {
37
+ "sha256": "5e32ae389932ce9a9b43ee7ace1170368c7c0e2887c8af234551c169b836be2a",
38
+ "bytes": 6326
39
+ },
40
+ "QA_REPORT.md": {
41
+ "sha256": "f8084e3fdbafc2f46a13fecceb5f37c5941debb7bf45a001740d4abaee9d7f18",
42
+ "bytes": 619
43
+ },
44
+ "README.md": {
45
+ "sha256": "0fb96c72bbdbb52c538deef737b2d6c9c3ef1c148312a0460f0c22c563faa6d8",
46
+ "bytes": 7494
47
+ },
48
+ "rejected.jsonl": {
49
+ "sha256": "7467f7bdcdbc9350e36e78aae7fb5bf0632580fe0eb1a642dc7b709ddd948515",
50
+ "bytes": 122722
51
+ },
52
+ "requirements.txt": {
53
+ "sha256": "91c83cc4a9d3fcac6ffa0d05dc21073401a69dd76be6a020fe81efed530289cd",
54
+ "bytes": 69
55
+ },
56
+ "robustness_report.json": {
57
+ "sha256": "f8122cd495e3be8c4cd1e0263eb5b163600d8a068bf25f3643a854d6ded17772",
58
+ "bytes": 11965
59
+ },
60
+ "schema.json": {
61
+ "sha256": "87c8bc827e0ed2c7bc2fc185d09603de1b037f73b77ed93cf60f6d470f871660",
62
+ "bytes": 2643
63
+ },
64
+ "serialization.json": {
65
+ "sha256": "c5d55d098628b3c8b7b655e23d5768166fa84fb332e9309bf86523003c356c6d",
66
+ "bytes": 423
67
+ },
68
+ "SOURCE_LOCK.json": {
69
+ "sha256": "63970a1a21eda958281d153e84d256201c93d97778541275ad76741c83ac9b16",
70
+ "bytes": 5370
71
+ },
72
+ "stats.json": {
73
+ "sha256": "e4aafbe217b6d3788f3fac11b4226027f9441f253901e99630f5123f0049707d",
74
+ "bytes": 1168
75
+ },
76
+ "test.jsonl": {
77
+ "sha256": "7f1d90b4c240d32796b498c2a149ec288262aa837a84ece25f984a9a2dc3f68b",
78
+ "bytes": 291201
79
+ },
80
+ "THIRD_PARTY_LICENSES.md": {
81
+ "sha256": "a8869bc8ba7fa034e60592d8f2ecd7592e9758a6aa8d3d1886bd78df733fad7e",
82
+ "bytes": 3321
83
+ },
84
+ "train.jsonl": {
85
+ "sha256": "012442fff17b8b8f01fc22169363e76a13740e6ae33f8beb8929d27e9660d978",
86
+ "bytes": 32358142
87
+ },
88
+ "validation.jsonl": {
89
+ "sha256": "79be61f28b6f88ec93b3503a8c4cbf278fb3f9a3cbddfb6003d9c9ef263b4e4f",
90
+ "bytes": 2470028
91
+ },
92
+ "validation_report.json": {
93
+ "sha256": "b8e3d295b22055292a678fcca54fb5763f58c55453dc9eebe109e3ac0c2f8c08",
94
+ "bytes": 705
95
+ },
96
+ "data/holdout.parquet": {
97
+ "sha256": "e865560c21eeeaa68c122c76a307eafb5589d4ee5e4b39bae9227a958cdef7b6",
98
+ "bytes": 1451009
99
+ },
100
+ "data/test.parquet": {
101
+ "sha256": "9ee434c82448c207b597d1dd166f7a17ac50966422230fcdf78246bc3c8eb593",
102
+ "bytes": 25820
103
+ },
104
+ "data/train.parquet": {
105
+ "sha256": "4799594f224d12380b085d3d06ef3982ccef0c21d0966fd283751b680a16be0b",
106
+ "bytes": 3025580
107
+ },
108
+ "data/validation.parquet": {
109
+ "sha256": "24afe17b55c1074e7b1b9f503057e1e6d6763e22e30542794c76ca230a7d0d31",
110
+ "bytes": 257729
111
+ },
112
+ "samples/examples.jsonl": {
113
+ "sha256": "c4c8f04752c1af478b96d5c68edb2986164ea9e19f436bdd860a48700a3de678",
114
+ "bytes": 9040
115
+ },
116
+ "validator/check_package.py": {
117
+ "sha256": "e4542037254a5aef1c3e49fd7cf483e96a738e265d5746ed4cae01a28d3a61f3",
118
+ "bytes": 2586
119
+ },
120
+ "validator/LICENSE": {
121
+ "sha256": "7c82eb9f40f567300f2a945943695da0f6b64eb400bb267032d1aa73ff177726",
122
+ "bytes": 1087
123
+ },
124
+ "validator/validate.py": {
125
+ "sha256": "0a8b440ad924e065de0f438d4a473baf496c69ad681146962860249c599f136b",
126
+ "bytes": 16694
127
+ },
128
+ "validator/vendor/common.py": {
129
+ "sha256": "194b6e0cc92e5d4f40c0ed96f77627e61cc98b64c3b3754de594b82f6b096b81",
130
+ "bytes": 7182
131
+ },
132
+ "validator/vendor/horizon.py": {
133
+ "sha256": "c124ae50b1f1c1a2666e0aa9201edf0822c1ff684ab56f4f17d54cd219dda8e2",
134
+ "bytes": 6434
135
+ },
136
+ "validator/vendor/nav_check.py": {
137
+ "sha256": "6093b4f2e1a6f1a730398a43539f5cee665335ff48725cfe0256cbbf9bd7664c",
138
+ "bytes": 15810
139
+ },
140
+ "validator/vendor/physics.py": {
141
+ "sha256": "863b324344a53f71c5c638aec66ebda2e16ed34e8d4591ac1ac16a44d96a93c7",
142
+ "bytes": 5706
143
  }
144
+ }
145
  }
PACKAGE_PROVENANCE.json ADDED
@@ -0,0 +1,7 @@
 
 
 
 
 
 
 
 
1
+ {
2
+ "ancestor_repo": "RegalFire/Agent-Failure-Recovery-Benchmark",
3
+ "ancestor_commit": "3ba263588c21c07acc96c1ead2a049b9915da15d",
4
+ "raw_dataset_sha256": "c0a92838f16d2fb13ec372cc2b0aa32c986529837c8b350bb71e87474cafbae2",
5
+ "change": "Lossless viewer-compatible Parquet, metadata, examples, evaluation interface, independent packaging checks. No new or duplicated trajectories.",
6
+ "license": "MIT; all four selected source datasets are MIT. No scientific community content is present."
7
+ }
QA_REPORT.md CHANGED
@@ -1,15 +1,7 @@
1
- # RegalFire — QA Report
2
 
3
- Validation status: PASS. 22,573/22,573 records checked independently.
4
 
5
- Unique IDs: 22,573; exact semantic duplicates: 0; grouped source/template units: 23.
6
 
7
- Controlled corruptions rejected: 14/14.
8
-
9
- Checks: strict UTF-8/JSON (duplicate keys/nonfinite numbers rejected), schema fields/types, source hashes/relations, attribution/license, labels, independent source oracles, witness replay, split file partition, shared source/text/template/grid groups and scientific near-question leakage.
10
-
11
- Near duplicates: scientific normalized text/code and 5-shingle near-question groups; synthetic template-held-out grouping. Exhaustive semantic detection is not claimed. Synthetic master seeds are not individual task seeds; this is not a fresh-seed independence claim.
12
-
13
- Rejected candidates: 1427. rejected.jsonl contains identifiers and exclusion reasons, not unapproved source bodies. Many exclusions are selection failures or redundant examples rather than corrupt records.
14
-
15
- Scientific answers remain community content and scientific code was not executed. Public holdout content is exposed. See README for all label and source limitations.
 
1
+ # Validation report
2
 
3
+ 22,573/22,573 records independently source-checked and oracle-checked: PASS. Zero exact semantic duplicates. 23 connected grouping units. All original split partitions pass.
4
 
5
+ 22,573 Parquet rows restored exactly: PASS. 40/40 deliberate trajectory corruptions rejected across four domains. Strict baseline scorer smoke tests: PASS. Local Hugging Face Datasets Parquet load: PASS. YAML metadata validation: PASS.
6
 
7
+ See validation_report.json, package_validation.json and evaluation_smoke.json for machine-readable evidence. No model performance or exhaustive semantic deduplication is claimed.
 
 
 
 
 
 
 
 
README.md CHANGED
@@ -1,73 +1,136 @@
1
- ---
2
- license: mit
3
- language:
4
- - en
5
- tags:
6
- - regalfire
7
- - evaluation
8
- - reproducible
9
- configs:
10
- - config_name: default
11
- data_files:
12
- - split: train
13
- path: train.jsonl
14
- - split: validation
15
- path: validation.jsonl
16
- - split: test
17
- path: test.jsonl
18
- - split: holdout
19
- path: holdout.jsonl
20
- ---
21
- # Agent-Failure-Recovery-Benchmark
22
-
23
- RegalFire — AI Data Foundry · Omar Soliman · ootiris@gmail.com
24
-
25
- ## Problem
26
-
27
- Cross-domain deterministic failure/recovery trajectories. **22,573 records** in this source-derived release. Do not sum it with its public ancestors as unique examples.
28
-
29
- ## Sources and provenance
30
-
31
- SOURCE_LOCK.json pins approved releases and file hashes. Per-record source IDs and hashes link to source rows. Selected normalized native scientific-computing sources accompany the affected products in native_scicomp_source.jsonl; raw API scrape files are not included. Source URLs, author attribution and revision contributors are retained. Unverified Bionic inputs are excluded.
32
-
33
- ## Schema and extraction
34
-
35
- See schema.json for required fields/types. Nested source content and labels are independently validated. Navigation: reject a source path invalidated by a critical edit, then replan or mark unreachable. Multi-agent: reject premature stop and recover with an optimal joint plan or mark infeasible. Counterfactual: reject an inconsistent prediction and replace it with the analytical result. Planning: reject a wrong next-step policy action and apply the correct step; long-horizon completion is not claimed.
36
-
37
- ## What labels mean
38
-
39
- Success means the narrow recovery contract, not game completion or deployed-agent reliability.
40
-
41
- ## Validation and reproduction
42
-
43
- Install Python 3.10+ dependencies: `pip install jsonschema huggingface_hub`. Run in the repository directory: `python validator/validate.py .` Pinned public ancestors require network or a populated HF cache; credentials are unnecessary. The validator does not import the builder. Source computational checkers were implemented independently from source generators. See validation_report.json, robustness_report.json and QA_REPORT.md.
44
-
45
- ## Splits and leakage
46
-
47
- {"holdout": 6058, "test": 165, "train": 15150, "validation": 1200}. Scientific threads, normalized shared text/code and near questions (5-shingle Jaccard >=0.85) are globally grouped across the two scientific-derived products. Synthetic task families are kept in one split, with repeated source grid groups united. Source master seeds identify an entire generation run, not independent per-task seeds. These are template-held-out splits, not fresh-generation-seed experiments. Uneven splits are intentional. Public holdout labels are visible and are not secret or contamination-free. Exhaustive semantic near-duplicate detection is not claimed.
48
-
49
- ## Intended uses and limitations
50
-
51
- Research, regression and scoped private adaptation. No measured model performance, clients, ROI or clinical claims. Scientific code is not executed; community answers are not expert labels. Clinical flags and email-like content are excluded heuristically, not certified absent. Agent errors are controlled harness operations, not real-agent logs; counterfactual recovery is prediction correction, not physical intervention. Validation checks the specified contracts, not universal scientific correctness.
52
-
53
- ## License
54
-
55
- mit. Preserve attribution, revision contributors and source notices. Scientific content remains CC BY-SA 4.0; MIT synthetic rows retain MIT notices. See THIRD_PARTY_LICENSES.md. No private exclusivity is promised for public source adaptations.
56
-
57
- ## Custom / Private Dataset Work
58
-
59
- RegalFire builds custom AI datasets, evaluation sets and reproducible data pipelines for research and production systems.
60
-
61
- Available:
62
- - custom schemas
63
- - private sources
64
- - private holdout sets
65
- - custom validators
66
- - domain-specific hard cases
67
- - continuous generation
68
- - regression datasets
69
-
70
- Contact: ootiris@gmail.com
71
- Hugging Face: RegalFire
72
-
73
- Public license obligations remain applicable. Confidential private work requires client-owned or otherwise permitted sources.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ pretty_name: Agent Failure Recovery Benchmark — 22,573 Verified Trajectories
3
+ license: mit
4
+ language:
5
+ - en
6
+ task_categories:
7
+ - reinforcement-learning
8
+ - text-generation
9
+ tags:
10
+ - agents
11
+ - agentic-ai
12
+ - benchmark
13
+ - evaluation
14
+ - reasoning
15
+ - trajectories
16
+ - failure-recovery
17
+ - reinforcement-learning
18
+ - synthetic-data
19
+ - verified
20
+ - planning
21
+ - multi-agent
22
+ - text
23
+ size_categories:
24
+ - 10K<n<100K
25
+ configs:
26
+ - config_name: default
27
+ data_files:
28
+ - split: train
29
+ path: data/train.parquet
30
+ - split: validation
31
+ path: data/validation.parquet
32
+ - split: test
33
+ path: data/test.parquet
34
+ - split: holdout
35
+ path: data/holdout.parquet
36
+ ---
37
+
38
+ # Agent Failure Recovery Benchmark
39
+
40
+ **22,573 synthetic, source-verified failure → recovery trajectories across four domains.**
41
+ Evaluate whether a system can reject a failed plan, choose a recovery, or recognize that no recovery exists. Every row links to a hashed record in a pinned public source and is checked by independent computational oracles and trajectory replay.
42
+
43
+ **Tasks:** failure diagnosis, recovery-action prediction, planning regression, scoped agent evaluation and imitation learning.
44
+
45
+ **Limits:** controlled harness errors, not real agent logs; small template diversity; public answers; no measured model performance. Physics rows correct predictions rather than perform physical interventions. Planning rows verify one next step, not full task completion.
46
+
47
+ ```python
48
+ from datasets import load_dataset
49
+ import json
50
+ ds = load_dataset("RegalFire/Agent-Failure-Recovery-Benchmark")
51
+ row = ds["test"][0]
52
+ failure = json.loads(row["failure_state"])
53
+ recovery = json.loads(row["recovery_action"])
54
+ ```
55
+
56
+ ## Coverage
57
+
58
+ | Domain | Records | Failure and recovery contract |
59
+ |---|---:|---|
60
+ | Navigation | 8,000 | A critical grid edit blocks a legacy path; replan optimally or report unreachable |
61
+ | Multi-agent | 8,000 | A premature stop fails the goal; recover with a collision-free optimal joint plan or report infeasible |
62
+ | Counterfactual physics | 6,000 | An inconsistent prediction is rejected; replace it with the analytical transition |
63
+ | Planning | 573 | A wrong available policy action is rejected; apply the independently verified next step |
64
+
65
+ Includes recoverable and unrecoverable cases. There are 27 domain-specific task families and 23 connected source/template groups. See `stats.json` for exact family counts.
66
+
67
+ ## A trajectory at a glance
68
+
69
+ ```text
70
+ initial state → rejected action / prediction → failure state
71
+ → verified recovery (or no recovery) → final state
72
+ ```
73
+
74
+ Each record provides `initial_state`, `goal`, `action_history`, `failure_state`, `failure_type`, `recovery_available`, `recovery_action`, `recovery_sequence`, `final_state`, `success`, `cost_or_steps`, `oracle`, `provenance`, `group_id`, and `split`. Four complete examples are in [samples/examples.jsonl](samples/examples.jsonl).
75
+
76
+ ## Viewer and formats
77
+
78
+ The default configuration uses Zstandard Parquet. Heterogeneous object/list columns are **losslessly encoded JSON strings**, including JSON `null` for absent recovery actions. Decode them with `json.loads`; see `serialization.json` for the exact columns. This avoids incompatible nested Arrow schemas across domains.
79
+
80
+ Original nested records remain in `dataset.jsonl` and the four root split JSONL files. Use plain Python to read those; the default Hub loader uses Parquet. Samples are examples of existing rows, not additional data.
81
+
82
+ ## Evaluation without answer leakage
83
+
84
+ Run `python evaluate.py make-prompts test prompts.jsonl`, give only those prompts to your system, then save predictions as one JSON object per line:
85
+
86
+ ```json
87
+ {"id":"<record id>","recovery_available":true,"recovery_action":{"action":"<predicted action>"}}
88
+ ```
89
+
90
+ Run `python evaluate.py score test predictions.jsonl`. It reports availability accuracy and strict recovery-action exact match, overall and per domain. Missing predictions count as failures; unknown/duplicate IDs are rejected. **Exact match can penalize alternative equally valid paths/plans.** It is a reproducible baseline, not a semantic trajectory judge or a model-performance claim. The independent validator verifies dataset truth, separately from this baseline scorer.
91
+
92
+ Prompt exports exclude reference recovery, final state, oracle and provenance. Public split labels are still visible, so this is an exposed regression benchmark, not a secret evaluation.
93
+
94
+ ## Validation you can rerun
95
+
96
+ ```bash
97
+ pip install -r requirements.txt
98
+ python validator/validate.py .
99
+ python validator/check_package.py
100
+ ```
101
+
102
+ The first command checks all 22,573 records: strict JSON, schema, identity, exact semantic deduplication, source file and row hashes, source relations, source oracles, failure/recovery contracts, recovery witnesses, and grouped split partitioning. It does not import a builder. Pinned ancestors require network access or a populated Hugging Face cache; credentials are unnecessary.
103
+
104
+ The second independently compares decoded Parquet with original rows and deliberately corrupts ten trajectory fields in each of four domains. All 40 corruptions must be rejected. Reports: `validation_report.json`, `package_validation.json`; prior release tests are retained in `robustness_report.json`.
105
+
106
+ ## Splits and limitations
107
+
108
+ | Split | Rows |
109
+ |---|---:|
110
+ | train | 15,150 |
111
+ | validation | 1,200 |
112
+ | test | 165 |
113
+ | holdout | 6,058 |
114
+
115
+ Splits retain whole task-template families and connected grid groups. The uneven sizes and small test split are intentional consequences of grouping; validation and holdout should also be evaluated separately. Master seeds identify generation runs, not independent task seeds. These are template-held-out splits, not fresh-seed independence experiments. Exhaustive semantic near-duplicate detection is not claimed.
116
+
117
+ Real agents may fail in ways absent here. No browser/tool/API error traces, measured model leaderboard, real-world outcomes, contamination-free evaluation, or reinforcement-learning environment implementation are provided. Successful recovery means the narrow contract described above.
118
+
119
+ ## Provenance and license
120
+
121
+ This release packages an existing verified RegalFire asset; it adds **zero new trajectories**. The previous release is pinned in `PACKAGE_PROVENANCE.json`. `SOURCE_LOCK.json` pins four MIT source datasets and file hashes:
122
+
123
+ - [GameNav-CriticalEdits-25K](https://huggingface.co/datasets/RegalFire/GameNav-CriticalEdits-25K)
124
+ - [GameCoord-JointPlans-25K](https://huggingface.co/datasets/RegalFire/GameCoord-JointPlans-25K)
125
+ - [GamePhysics-Counterfactuals-100K](https://huggingface.co/datasets/RegalFire/GamePhysics-Counterfactuals-100K)
126
+ - [GameAgent-Horizon-120K](https://huggingface.co/datasets/RegalFire/GameAgent-Horizon-120K)
127
+
128
+ Per-row provenance retains source ID, source hash, transformation and license. No unverified Bionic input was added. Do not count these rows and their source ancestors as independent examples. This dataset contains synthetic MIT rows; no scientific community/CC BY-SA content is present. Preserve `LICENSE`, `THIRD_PARTY_LICENSES.md` and validator notices.
129
+
130
+ ## Custom / Private Dataset Work
131
+
132
+ RegalFire builds custom AI datasets, evaluation sets and reproducible data pipelines for research and production systems.
133
+ Contact: ootiris@gmail.com
134
+ Hugging Face: [RegalFire](https://huggingface.co/RegalFire)
135
+
136
+ Private adaptation can include client-owned environments, new failure contracts, independent validators and private holdout sets. Public source adaptations remain public-license material.
data/holdout.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e865560c21eeeaa68c122c76a307eafb5589d4ee5e4b39bae9227a958cdef7b6
3
+ size 1451009
data/test.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:9ee434c82448c207b597d1dd166f7a17ac50966422230fcdf78246bc3c8eb593
3
+ size 25820
data/train.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:4799594f224d12380b085d3d06ef3982ccef0c21d0966fd283751b680a16be0b
3
+ size 3025580
data/validation.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:24afe17b55c1074e7b1b9f503057e1e6d6763e22e30542794c76ca230a7d0d31
3
+ size 257729
evaluate.py ADDED
@@ -0,0 +1,34 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Prompt export and strict reference-action baseline; stdlib only."""
2
+ import json,sys
3
+ from pathlib import Path
4
+ from collections import defaultdict
5
+ R=Path(__file__).resolve().parent
6
+ INPUTS=['id','domain','task_type','initial_state','goal','action_history','failure_state','failure_type']
7
+ def canon(x):return json.dumps(x,sort_keys=True,separators=(',',':'),allow_nan=False)
8
+ def read(path):
9
+ def pairs(items):
10
+ d={}
11
+ for k,v in items:
12
+ if k in d:raise ValueError('duplicate JSON key')
13
+ d[k]=v
14
+ return d
15
+ return [json.loads(x,object_pairs_hook=pairs,parse_constant=lambda x: (_ for _ in ()).throw(ValueError('nonfinite'))) for x in Path(path).read_text('utf8').splitlines() if x.strip()]
16
+ def score(rows,predictions):
17
+ refs={r['id']:r for r in rows};pred={};totals=defaultdict(lambda:[0,0,0])
18
+ for p in predictions:
19
+ if p['id'] not in refs or p['id'] in pred:raise ValueError('unknown or duplicate prediction ID')
20
+ if type(p.get('recovery_available')) is not bool:raise ValueError('availability must be bool')
21
+ if 'recovery_action' not in p or (p['recovery_action'] is not None and not isinstance(p['recovery_action'],dict)):raise ValueError('recovery_action must be object or null')
22
+ pred[p['id']]=p
23
+ for r in rows:
24
+ p=pred.get(r['id']);a=p is not None and p['recovery_available']==r['recovery_available'];b=a and canon(p['recovery_action'])==canon(r['recovery_action'])
25
+ for group in ['overall',r['domain']]:
26
+ totals[group][0]+=1;totals[group][1]+=int(a);totals[group][2]+=int(b)
27
+ return {'metric':'strict reference action match; alternative valid plans may score zero','predictions':len(pred),'missing':len(rows)-len(pred),'scores':{g:{'n':n,'availability_accuracy':a/n,'recovery_action_exact_match':b/n} for g,(n,a,b) in totals.items()}}
28
+ if __name__=='__main__':
29
+ mode,split,path=sys.argv[1:];assert split in ['train','validation','test','holdout']
30
+ rows=read(R/(split+'.jsonl'))
31
+ if mode=='make-prompts':
32
+ Path(path).write_text(''.join(json.dumps({'instruction':'Infer whether recovery is available and return recovery_available (bool) and recovery_action (object or null).','input':{k:r[k] for k in INPUTS}},ensure_ascii=False)+'\n' for r in rows),encoding='utf8')
33
+ elif mode=='score':print(json.dumps(score(rows,read(path)),indent=2))
34
+ else:raise ValueError('mode must be make-prompts or score')
evaluation_smoke.json ADDED
@@ -0,0 +1,11 @@
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "status": "PASS",
3
+ "tests": [
4
+ "perfect reference predictions score 1",
5
+ "missing predictions score 0",
6
+ "duplicate IDs rejected",
7
+ "unknown IDs rejected",
8
+ "nonbool rejected"
9
+ ],
10
+ "not_model_performance": true
11
+ }
package_validation.json ADDED
@@ -0,0 +1,249 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "status": "PASS",
3
+ "lossless_parquet_rows": 22573,
4
+ "controlled_corruptions": 40,
5
+ "rejected_corruptions": 40,
6
+ "tests": [
7
+ {
8
+ "domain": "navigation",
9
+ "field": "success",
10
+ "rejected": true,
11
+ "reason": "success/availability mismatch"
12
+ },
13
+ {
14
+ "domain": "navigation",
15
+ "field": "recovery_action",
16
+ "rejected": true,
17
+ "reason": "recovery trajectory mismatch"
18
+ },
19
+ {
20
+ "domain": "navigation",
21
+ "field": "recovery_sequence",
22
+ "rejected": true,
23
+ "reason": "recovery trajectory mismatch"
24
+ },
25
+ {
26
+ "domain": "navigation",
27
+ "field": "final_state",
28
+ "rejected": true,
29
+ "reason": "recovery trajectory mismatch"
30
+ },
31
+ {
32
+ "domain": "navigation",
33
+ "field": "failure_type",
34
+ "rejected": true,
35
+ "reason": "failure/goal labels mismatch"
36
+ },
37
+ {
38
+ "domain": "navigation",
39
+ "field": "goal",
40
+ "rejected": true,
41
+ "reason": "failure/goal labels mismatch"
42
+ },
43
+ {
44
+ "domain": "navigation",
45
+ "field": "action_history",
46
+ "rejected": true,
47
+ "reason": "history mismatch"
48
+ },
49
+ {
50
+ "domain": "navigation",
51
+ "field": "cost_or_steps",
52
+ "rejected": true,
53
+ "reason": "recovery trajectory mismatch"
54
+ },
55
+ {
56
+ "domain": "navigation",
57
+ "field": "initial_state",
58
+ "rejected": true,
59
+ "reason": "initial state mismatch"
60
+ },
61
+ {
62
+ "domain": "navigation",
63
+ "field": "failure_state",
64
+ "rejected": true,
65
+ "reason": "failure state mismatch"
66
+ },
67
+ {
68
+ "domain": "multi_agent",
69
+ "field": "success",
70
+ "rejected": true,
71
+ "reason": "success/availability mismatch"
72
+ },
73
+ {
74
+ "domain": "multi_agent",
75
+ "field": "recovery_action",
76
+ "rejected": true,
77
+ "reason": "recovery trajectory mismatch"
78
+ },
79
+ {
80
+ "domain": "multi_agent",
81
+ "field": "recovery_sequence",
82
+ "rejected": true,
83
+ "reason": "recovery trajectory mismatch"
84
+ },
85
+ {
86
+ "domain": "multi_agent",
87
+ "field": "final_state",
88
+ "rejected": true,
89
+ "reason": "recovery trajectory mismatch"
90
+ },
91
+ {
92
+ "domain": "multi_agent",
93
+ "field": "failure_type",
94
+ "rejected": true,
95
+ "reason": "failure/goal labels mismatch"
96
+ },
97
+ {
98
+ "domain": "multi_agent",
99
+ "field": "goal",
100
+ "rejected": true,
101
+ "reason": "failure/goal labels mismatch"
102
+ },
103
+ {
104
+ "domain": "multi_agent",
105
+ "field": "action_history",
106
+ "rejected": true,
107
+ "reason": "joint history mismatch"
108
+ },
109
+ {
110
+ "domain": "multi_agent",
111
+ "field": "cost_or_steps",
112
+ "rejected": true,
113
+ "reason": "recovery trajectory mismatch"
114
+ },
115
+ {
116
+ "domain": "multi_agent",
117
+ "field": "initial_state",
118
+ "rejected": true,
119
+ "reason": "joint state mismatch"
120
+ },
121
+ {
122
+ "domain": "multi_agent",
123
+ "field": "failure_state",
124
+ "rejected": true,
125
+ "reason": "joint state mismatch"
126
+ },
127
+ {
128
+ "domain": "counterfactual",
129
+ "field": "success",
130
+ "rejected": true,
131
+ "reason": "success/availability mismatch"
132
+ },
133
+ {
134
+ "domain": "counterfactual",
135
+ "field": "recovery_action",
136
+ "rejected": true,
137
+ "reason": "recovery trajectory mismatch"
138
+ },
139
+ {
140
+ "domain": "counterfactual",
141
+ "field": "recovery_sequence",
142
+ "rejected": true,
143
+ "reason": "recovery trajectory mismatch"
144
+ },
145
+ {
146
+ "domain": "counterfactual",
147
+ "field": "final_state",
148
+ "rejected": true,
149
+ "reason": "recovery trajectory mismatch"
150
+ },
151
+ {
152
+ "domain": "counterfactual",
153
+ "field": "failure_type",
154
+ "rejected": true,
155
+ "reason": "failure/goal labels mismatch"
156
+ },
157
+ {
158
+ "domain": "counterfactual",
159
+ "field": "goal",
160
+ "rejected": true,
161
+ "reason": "failure/goal labels mismatch"
162
+ },
163
+ {
164
+ "domain": "counterfactual",
165
+ "field": "action_history",
166
+ "rejected": true,
167
+ "reason": "physics history mismatch"
168
+ },
169
+ {
170
+ "domain": "counterfactual",
171
+ "field": "cost_or_steps",
172
+ "rejected": true,
173
+ "reason": "recovery trajectory mismatch"
174
+ },
175
+ {
176
+ "domain": "counterfactual",
177
+ "field": "initial_state",
178
+ "rejected": true,
179
+ "reason": "physics initial mismatch"
180
+ },
181
+ {
182
+ "domain": "counterfactual",
183
+ "field": "failure_state",
184
+ "rejected": true,
185
+ "reason": "physics failure mismatch"
186
+ },
187
+ {
188
+ "domain": "planning",
189
+ "field": "success",
190
+ "rejected": true,
191
+ "reason": "success/availability mismatch"
192
+ },
193
+ {
194
+ "domain": "planning",
195
+ "field": "recovery_action",
196
+ "rejected": true,
197
+ "reason": "recovery trajectory mismatch"
198
+ },
199
+ {
200
+ "domain": "planning",
201
+ "field": "recovery_sequence",
202
+ "rejected": true,
203
+ "reason": "recovery trajectory mismatch"
204
+ },
205
+ {
206
+ "domain": "planning",
207
+ "field": "final_state",
208
+ "rejected": true,
209
+ "reason": "recovery trajectory mismatch"
210
+ },
211
+ {
212
+ "domain": "planning",
213
+ "field": "failure_type",
214
+ "rejected": true,
215
+ "reason": "failure/goal labels mismatch"
216
+ },
217
+ {
218
+ "domain": "planning",
219
+ "field": "goal",
220
+ "rejected": true,
221
+ "reason": "failure/goal labels mismatch"
222
+ },
223
+ {
224
+ "domain": "planning",
225
+ "field": "action_history",
226
+ "rejected": true,
227
+ "reason": "failure action not a wrong policy action"
228
+ },
229
+ {
230
+ "domain": "planning",
231
+ "field": "cost_or_steps",
232
+ "rejected": true,
233
+ "reason": "recovery trajectory mismatch"
234
+ },
235
+ {
236
+ "domain": "planning",
237
+ "field": "initial_state",
238
+ "rejected": true,
239
+ "reason": "planning initial/failure mismatch"
240
+ },
241
+ {
242
+ "domain": "planning",
243
+ "field": "failure_state",
244
+ "rejected": true,
245
+ "reason": "planning initial/failure mismatch"
246
+ }
247
+ ],
248
+ "scope": "Lossless serialization and deliberate trajectory corruptions across every domain; full source/oracle audit is validator/validate.py."
249
+ }
requirements.txt ADDED
@@ -0,0 +1,4 @@
 
 
 
 
 
1
+ jsonschema>=4.20,<5
2
+ huggingface_hub>=0.30,<2
3
+ pyarrow>=15
4
+ datasets>=3
samples/examples.jsonl ADDED
@@ -0,0 +1,4 @@
 
 
 
 
 
1
+ {"action_history": [{"action": "submit_path", "path": [[4, 0], [5, 0], [5, 1], [5, 2], [5, 3], [5, 4], [5, 5], [5, 6], [4, 6], [3, 6], [2, 6], [1, 6]], "result": "rejected_atomically_blocked_cell", "state_unchanged": true}], "content_license": "MIT", "cost_or_steps": {"cost": null, "steps": null}, "domain": "navigation", "failure_state": {"costs": [[1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1]], "grid": ["....##..", "##..#...", "....#...", "..#.##..", ".#..##..", "....#...", ".#####.."], "position": [4, 0]}, "failure_type": "legacy_path_blocked_by_critical_edit", "final_state": {"costs": [[1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1]], "grid": ["....##..", "##..#...", "....#...", "..#.##..", ".#..##..", "....#...", ".#####.."], "position": [4, 0]}, "goal": {"position": [1, 6]}, "group_id": "56b52f4688a5079bec48b61f0746f3dc83015df6977469e2921e9779cd1a301d", "id": "AFR-b4f7ffc6c94405ff11386833", "initial_state": {"costs": [[1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1]], "grid": ["....##..", "##..#...", "....#...", "..#.##..", ".#..##..", "........", ".#####.."], "position": [4, 0]}, "oracle": {"failure_semantics": "Explicit evaluation harness rejection; failed actions are not claimed real agent observations", "source_expected": {"after": {"cost": null, "path": null}, "before": {"cost": 11, "path": [[4, 0], [5, 0], [5, 1], [5, 2], [5, 3], [5, 4], [5, 5], [5, 6], [4, 6], [3, 6], [2, 6], [1, 6]]}, "reachability_changed": true}, "source_family": "gap_close", "source_input": {"costs": [[1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1]], "edit": [5, 4], "goal": [1, 6], "grid": ["....##..", "##..#...", "....#...", "..#.##..", ".#..##..", "........", ".#####.."], "start": [4, 0]}, "source_rule": "RegalFire deterministic source logic"}, "provenance": {"master_generation_seed": 20261001, "procedural_group_id": "435c0144f2b547984f453117a4d6392cc861bad4f58438194b939d378750a8f1", "source_commit": "e8e8c8d504db6ccd621c4a067c47d0f797e4b70d", "source_data_file": "data/dataset.jsonl", "source_dataset": "RegalFire/GameNav-CriticalEdits-25K", "source_license": "MIT", "source_record_id": "NAV-0000000", "source_record_sha256": "9b4b9993c920b5726a160dc767cfbb2b689e2895530892b831bbfae5a608899b", "transformation": "RegalFire V3 deterministic transformation; no LLM truth labels"}, "recovery_action": null, "recovery_available": false, "recovery_sequence": [], "source_dataset": "RegalFire/GameNav-CriticalEdits-25K", "split": "holdout", "success": false, "task_type": "gap_close", "verification_method": "independent source oracle recomputation and recovery witness replay"}
2
+ {"action_history": [{"action": "stop", "result": "rejected_goals_unsatisfied", "state_unchanged": true}], "content_license": "MIT", "cost_or_steps": {"cost": null, "steps": null}, "domain": "multi_agent", "failure_state": {"grid": [".....", "##.##", "..##.", "#..##"], "positions": [[0, 3], [0, 2]]}, "failure_type": "premature_stop_goals_unsatisfied", "final_state": {"grid": [".....", "##.##", "..##.", "#..##"], "positions": [[0, 3], [0, 2]]}, "goal": {"absorbing_goals": true, "edge_swap_forbidden": true, "positions": [[0, 2], [0, 3]], "vertex_collision_forbidden": true}, "group_id": "e0afcc8a92a2ffa4764504d901356e395a98fb48b26a71db8471c199bd37f07e", "id": "AFR-5de685607b43cb3c8e00ceb8", "initial_state": {"grid": [".....", "##.##", "..##.", "#..##"], "positions": [[0, 3], [0, 2]]}, "oracle": {"failure_semantics": "Explicit evaluation harness rejection; failed actions are not claimed real agent observations", "source_expected": {"coordination_overhead": null, "independent_lower_bound": 1, "optimal_makespan": null, "plan": null}, "source_family": "adjacent_swap", "source_input": {"goals": [[0, 2], [0, 3]], "grid": [".....", "##.##", "..##.", "#..##"], "starts": [[0, 3], [0, 2]]}, "source_rule": "RegalFire deterministic source logic"}, "provenance": {"master_generation_seed": 20261001, "procedural_group_id": "5c2231f188e750d7634c88616b53663f8337ce3394fb816e0d0ce7716b597878", "source_commit": "5bdd6f516279c44d2495eb9084bba1c12f8a9390", "source_data_file": "data/dataset.jsonl", "source_dataset": "RegalFire/GameCoord-JointPlans-25K", "source_license": "MIT", "source_record_id": "COORD-0000000", "source_record_sha256": "7325a26b0fb79d152e6caed9e1e5a9604ae84b6c20764d08f2fc119035bd06ae", "transformation": "RegalFire V3 deterministic transformation; no LLM truth labels"}, "recovery_action": null, "recovery_available": false, "recovery_sequence": [], "source_dataset": "RegalFire/GameCoord-JointPlans-25K", "split": "train", "success": false, "task_type": "adjacent_swap", "verification_method": "independent source oracle recomputation and recovery witness replay"}
3
+ {"action_history": [{"action": "submit_prediction", "prediction": {"vy": 0.350586, "y": 47.434548}, "result": "rejected_analytical_inconsistency"}], "content_license": "MIT", "cost_or_steps": {"cost": 1, "steps": 1}, "domain": "counterfactual", "failure_state": {"predicted_state_t1": {"vy": 0.350586, "y": 47.434548}}, "failure_type": "analytically_inconsistent_vy", "final_state": {"predicted_state_t1": {"vy": 0.324368, "y": 47.434548}}, "goal": {"prediction_matches_analytical_transition": true}, "group_id": "39ee9df81233816910b6a0215843e1b29dda9715a6a264df46586ec29ff9e353", "id": "AFR-bfa196435fab1d0b66104ce3", "initial_state": {"action": {"action": "none"}, "dt": 0.1, "state_t0": {"vy": 1.305368, "y": 47.353061}}, "oracle": {"failure_semantics": "Explicit evaluation harness rejection; failed actions are not claimed real agent observations", "source_expected": null, "source_family": "gravity_fall", "source_input": null, "source_rule": "RegalFire deterministic source logic"}, "provenance": {"master_generation_seed": 20260930, "procedural_group_id": null, "source_commit": "3fadd21aac5aa54a3c293d086a48f5699a0d213b", "source_data_file": "gamephysics_counterfactuals_100k.jsonl", "source_dataset": "RegalFire/GamePhysics-Counterfactuals-100K", "source_license": "MIT", "source_record_id": "e93eb2d9c746bf3a12f42ab4b4dc737d", "source_record_sha256": "dfb814ec745da0ca4c55b5f61a2805565c7a7a7385b21e301dbff6a6fcb19613", "transformation": "RegalFire V3 deterministic transformation; no LLM truth labels"}, "recovery_action": {"action": "replace_prediction", "prediction": {"vy": 0.324368, "y": 47.434548}}, "recovery_available": true, "recovery_sequence": [{"action": "replace_prediction", "prediction": {"vy": 0.324368, "y": 47.434548}}], "source_dataset": "RegalFire/GamePhysics-Counterfactuals-100K", "split": "train", "success": true, "task_type": "gravity_fall", "verification_method": "independent source oracle recomputation and recovery witness replay"}
4
+ {"action_history": [{"action": "move_west", "result": "rejected_policy_mismatch", "state_unchanged": true}], "content_license": "MIT", "cost_or_steps": {"cost": 1, "steps": 1}, "domain": "planning", "failure_state": {"door_locked": true, "enemy_nearby": false, "inventory": [], "key_visible": true}, "failure_type": "next_step_policy_mismatch", "final_state": {"door_locked": true, "enemy_nearby": false, "inventory": ["key_0"], "key_visible": false}, "goal": {"full_long_horizon_goal_claimed": false, "short_horizon": "Pick up key_0"}, "group_id": "a7aa4c29523f9e259b9a1399b5497e6d1dbe953f0621540855ee8cba9ac6092d", "id": "AFR-7b6ee0f760aa37d8dcc15fca", "initial_state": {"door_locked": true, "enemy_nearby": false, "inventory": [], "key_visible": true}, "oracle": {"failure_semantics": "Explicit evaluation harness rejection; failed actions are not claimed real agent observations", "source_expected": null, "source_family": "key_then_door", "source_input": null, "source_rule": "RegalFire deterministic source logic"}, "provenance": {"master_generation_seed": 20260930, "procedural_group_id": null, "source_commit": "15af97f817891a5290afff28f4a49e86bafd6e64", "source_data_file": "gameagent_horizon_120k.jsonl", "source_dataset": "RegalFire/GameAgent-Horizon-120K", "source_license": "MIT", "source_record_id": "ab23656902b8a56ab990ba0fe9500b28", "source_record_sha256": "1498ca918a581a2e0913e1357cb638aca272dd415f026d811eca6e47540d0c91", "transformation": "RegalFire V3 deterministic transformation; no LLM truth labels"}, "recovery_action": {"action": "pickup"}, "recovery_available": true, "recovery_sequence": [{"action": "pickup"}], "source_dataset": "RegalFire/GameAgent-Horizon-120K", "split": "test", "success": true, "task_type": "key_then_door", "verification_method": "independent source oracle recomputation and recovery witness replay"}
serialization.json ADDED
@@ -0,0 +1,16 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "nested_json_columns": [
3
+ "action_history",
4
+ "cost_or_steps",
5
+ "failure_state",
6
+ "final_state",
7
+ "goal",
8
+ "initial_state",
9
+ "oracle",
10
+ "provenance",
11
+ "recovery_action",
12
+ "recovery_sequence"
13
+ ],
14
+ "reason": "Heterogeneous cross-domain structures are losslessly encoded as JSON strings for a stable Arrow schema. Raw original JSONL remains available.",
15
+ "records": 22573
16
+ }
validator/check_package.py ADDED
@@ -0,0 +1,35 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Independent packaging audit and adversarial tests; never imports the packager."""
2
+ import sys,json,hashlib,copy,importlib.util
3
+ from pathlib import Path
4
+ from collections import Counter
5
+ import pyarrow.parquet as pq
6
+
7
+ ROOT=Path(__file__).resolve().parents[1]
8
+ spec=importlib.util.spec_from_file_location('contract',ROOT/'validator'/'validate.py')
9
+ v=importlib.util.module_from_spec(spec);spec.loader.exec_module(v)
10
+ def audit():
11
+ rows=v.readrows(ROOT/'dataset.jsonl'); originals=v.originals(ROOT)
12
+ vendor=ROOT/'validator'/'vendor';sys.path.insert(0,str(vendor))
13
+ nav=v.module(vendor/'nav_check.py','nav_check');physics=v.module(vendor/'physics.py','physics');horizon=v.module(vendor/'horizon.py','horizon')
14
+ nested=json.loads((ROOT/'serialization.json').read_text('utf8'))['nested_json_columns']
15
+ restored=[]
16
+ for split in ['train','validation','test','holdout']:
17
+ raw=pq.read_table(ROOT/'data'/f'{split}.parquet').to_pylist()
18
+ for r in raw:
19
+ assert r['split']==split
20
+ for k in nested:r[k]=v.strict(r[k])
21
+ restored+=raw
22
+ assert Counter(v.encoded(r) for r in rows)==Counter(v.encoded(r) for r in restored),'Parquet changed/omitted/duplicated rows'
23
+ mutations=[]
24
+ for domain in ['navigation','multi_agent','counterfactual','planning']:
25
+ r=next(r for r in rows if r['domain']==domain and r['recovery_available'])
26
+ s=originals[(r['provenance']['source_dataset'],r['provenance']['source_record_id'])]
27
+ v.agent(r,s,nav,physics,horizon)
28
+ for field,value in [('success',False),('recovery_action',None),('recovery_sequence',[]),('final_state',{}),('failure_type','fabricated'),('goal',{}),('action_history',[]),('cost_or_steps',{'cost':-99,'steps':-99}),('initial_state',{}),('failure_state',{})]:
29
+ bad=copy.deepcopy(r);bad[field]=value
30
+ try:v.agent(bad,s,nav,physics,horizon)
31
+ except Exception as e:mutations.append({'domain':domain,'field':field,'rejected':True,'reason':str(e)})
32
+ else:raise AssertionError(f'Mutation survived: {domain}/{field}')
33
+ return {'status':'PASS','lossless_parquet_rows':len(restored),'controlled_corruptions':len(mutations),'rejected_corruptions':len(mutations),'tests':mutations,'scope':'Lossless serialization and deliberate trajectory corruptions across every domain; full source/oracle audit is validator/validate.py.'}
34
+ if __name__=='__main__':
35
+ report=audit();(ROOT/'package_validation.json').write_text(json.dumps(report,indent=2)+'\n',encoding='utf8');print(json.dumps({k:v for k,v in report.items() if k!='tests'}))