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 because of a cast error
Error code:   DatasetGenerationCastError
Exception:    DatasetGenerationCastError
Message:      An error occurred while generating the dataset

All the data files must have the same columns, but at some point there are 34 new columns ({'claude_sonnet_vanilla_status', 'gpt4o_vanilla_score', 'gpt4o_prompt_only_score', 'gpt4o_with_ejentum_single_completion', 'claude_sonnet_vanilla_score', 'gpt4o_prompt_only_status', 'adversarial_query', 'claude_sonnet_with_ejentum_single_score', 'claude_sonnet_with_ejentum_graph_eval_reason', 'gpt4o_vanilla_completion', 'gpt4o_with_ejentum_graph_eval_reason', 'gpt4o_prompt_only_eval_reason', 'claude_sonnet_vanilla_completion', 'gpt4o_vanilla_status', 'claude_sonnet_vanilla_eval_reason', 'gpt4o_with_ejentum_graph_status', 'claude_sonnet_with_ejentum_graph_completion', 'claude_sonnet_with_ejentum_single_eval_reason', 'gpt4o_with_ejentum_single_eval_reason', 'claude_sonnet_with_ejentum_graph_status', 'claude_sonnet_with_ejentum_single_status', 'claude_sonnet_prompt_only_status', 'claude_sonnet_prompt_only_score', 'claude_sonnet_prompt_only_completion', 'gpt4o_with_ejentum_graph_completion', 'gpt4o_vanilla_eval_reason', 'gpt4o_with_ejentum_graph_score', 'gpt4o_with_ejentum_single_score', 'claude_sonnet_with_ejentum_single_completion', 'trap_id_and_category', 'claude_sonnet_prompt_only_eval_reason', 'gpt4o_with_ejentum_single_status', 'claude_sonnet_with_ejentum_graph_score', 'gpt4o_prompt_only_completion'}) and 8 missing columns ({'Region', 'WarmCacheMs', 'UncachedHybridMs', 'Iteration', 'AbilityId', 'Status', 'DagTraversalMs', 'SqliteFts5Ms'}).

This happened while the csv dataset builder was generating data using

hf://datasets/ejentum/adversarial-systems-engineering-traps/data/results_320_adversarial_traps.csv (at revision 41a2c47fc79fa79a7bfe4e26b1f8f92980fd2296), ['hf://datasets/ejentum/adversarial-systems-engineering-traps@41a2c47fc79fa79a7bfe4e26b1f8f92980fd2296/data/latency_measurements_us-east-1.csv', 'hf://datasets/ejentum/adversarial-systems-engineering-traps@41a2c47fc79fa79a7bfe4e26b1f8f92980fd2296/data/results_320_adversarial_traps.csv']

Please either edit the data files to have matching columns, or separate them into different configurations (see docs at https://hf.co/docs/hub/datasets-manual-configuration#multiple-configurations)
Traceback:    Traceback (most recent call last):
                File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1856, in _prepare_split_single
                  writer.write_table(table)
                  ~~~~~~~~~~~~~~~~~~^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/arrow_writer.py", line 811, in write_table
                  self._write_table(pa_table, writer_batch_size=writer_batch_size)
                  ~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/arrow_writer.py", line 822, in _write_table
                  pa_table = table_cast(pa_table, self._schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2381, in table_cast
                  return cast_table_to_schema(table, schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2309, in cast_table_to_schema
                  raise CastError(
                  ...<3 lines>...
                  )
              datasets.table.CastError: Couldn't cast
              trap_id_and_category: string
              adversarial_query: string
              gpt4o_vanilla_completion: string
              gpt4o_vanilla_status: string
              gpt4o_vanilla_score: double
              gpt4o_vanilla_eval_reason: string
              gpt4o_prompt_only_completion: string
              gpt4o_prompt_only_status: string
              gpt4o_prompt_only_score: double
              gpt4o_prompt_only_eval_reason: string
              gpt4o_with_ejentum_single_completion: string
              gpt4o_with_ejentum_single_status: string
              gpt4o_with_ejentum_single_score: double
              gpt4o_with_ejentum_single_eval_reason: string
              gpt4o_with_ejentum_graph_completion: string
              gpt4o_with_ejentum_graph_status: string
              gpt4o_with_ejentum_graph_score: double
              gpt4o_with_ejentum_graph_eval_reason: string
              claude_sonnet_vanilla_completion: string
              claude_sonnet_vanilla_status: string
              claude_sonnet_vanilla_score: double
              claude_sonnet_vanilla_eval_reason: string
              claude_sonnet_prompt_only_completion: string
              claude_sonnet_prompt_only_status: string
              claude_sonnet_prompt_only_score: double
              claude_sonnet_prompt_only_eval_reason: string
              claude_sonnet_with_ejentum_single_completion: string
              claude_sonnet_with_ejentum_single_status: string
              claude_sonnet_with_ejentum_single_score: double
              claude_sonnet_with_ejentum_single_eval_reason: string
              claude_sonnet_with_ejentum_graph_completion: string
              claude_sonnet_with_ejentum_graph_status: string
              claude_sonnet_with_ejentum_graph_score: double
              claude_sonnet_with_ejentum_graph_eval_reason: string
              -- schema metadata --
              pandas: '{"index_columns": [{"kind": "range", "name": null, "start": 0, "' + 5856
              to
              {'Iteration': Value('int64'), 'Region': Value('string'), 'AbilityId': Value('string'), 'UncachedHybridMs': Value('float64'), 'WarmCacheMs': Value('float64'), 'SqliteFts5Ms': Value('float64'), 'DagTraversalMs': Value('float64'), 'Status': Value('int64')}
              because column names don't match
              
              During handling of the above exception, another exception occurred:
              
              Traceback (most recent call last):
                File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 1379, 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 956, 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 1702, 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 1858, in _prepare_split_single
                  raise DatasetGenerationCastError.from_cast_error(
                  ...<4 lines>...
                  )
              datasets.exceptions.DatasetGenerationCastError: An error occurred while generating the dataset
              
              All the data files must have the same columns, but at some point there are 34 new columns ({'claude_sonnet_vanilla_status', 'gpt4o_vanilla_score', 'gpt4o_prompt_only_score', 'gpt4o_with_ejentum_single_completion', 'claude_sonnet_vanilla_score', 'gpt4o_prompt_only_status', 'adversarial_query', 'claude_sonnet_with_ejentum_single_score', 'claude_sonnet_with_ejentum_graph_eval_reason', 'gpt4o_vanilla_completion', 'gpt4o_with_ejentum_graph_eval_reason', 'gpt4o_prompt_only_eval_reason', 'claude_sonnet_vanilla_completion', 'gpt4o_vanilla_status', 'claude_sonnet_vanilla_eval_reason', 'gpt4o_with_ejentum_graph_status', 'claude_sonnet_with_ejentum_graph_completion', 'claude_sonnet_with_ejentum_single_eval_reason', 'gpt4o_with_ejentum_single_eval_reason', 'claude_sonnet_with_ejentum_graph_status', 'claude_sonnet_with_ejentum_single_status', 'claude_sonnet_prompt_only_status', 'claude_sonnet_prompt_only_score', 'claude_sonnet_prompt_only_completion', 'gpt4o_with_ejentum_graph_completion', 'gpt4o_vanilla_eval_reason', 'gpt4o_with_ejentum_graph_score', 'gpt4o_with_ejentum_single_score', 'claude_sonnet_with_ejentum_single_completion', 'trap_id_and_category', 'claude_sonnet_prompt_only_eval_reason', 'gpt4o_with_ejentum_single_status', 'claude_sonnet_with_ejentum_graph_score', 'gpt4o_prompt_only_completion'}) and 8 missing columns ({'Region', 'WarmCacheMs', 'UncachedHybridMs', 'Iteration', 'AbilityId', 'Status', 'DagTraversalMs', 'SqliteFts5Ms'}).
              
              This happened while the csv dataset builder was generating data using
              
              hf://datasets/ejentum/adversarial-systems-engineering-traps/data/results_320_adversarial_traps.csv (at revision 41a2c47fc79fa79a7bfe4e26b1f8f92980fd2296), ['hf://datasets/ejentum/adversarial-systems-engineering-traps@41a2c47fc79fa79a7bfe4e26b1f8f92980fd2296/data/latency_measurements_us-east-1.csv', 'hf://datasets/ejentum/adversarial-systems-engineering-traps@41a2c47fc79fa79a7bfe4e26b1f8f92980fd2296/data/results_320_adversarial_traps.csv']
              
              Please either edit the data files to have matching columns, or separate them into different configurations (see docs at https://hf.co/docs/hub/datasets-manual-configuration#multiple-configurations)

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.

Iteration
int64
Region
string
AbilityId
string
UncachedHybridMs
float64
WarmCacheMs
float64
SqliteFts5Ms
float64
DagTraversalMs
float64
Status
int64
1
us-east-1
SP-039
2.0464
0.0035
1.591
0.0936
200
2
us-east-1
TE-004
2.0898
0.004
0.7966
0.0848
200
3
us-east-1
CA-036
1.7992
0.0026
0.7471
0.0789
200
4
us-east-1
SI-011
2.0245
0.0022
0.6263
0.0731
200
5
us-east-1
CA-034
1.8391
0.0037
0.8062
0.0893
200
6
us-east-1
CA-046
1.865
0.0026
0.767
0.075
200
7
us-east-1
SI-026
1.9105
0.0023
0.6879
0.067
200
8
us-east-1
SP-039
2.223
0.0037
0.7589
0.1583
200
9
us-east-1
TE-046
2.2596
0.0024
1.0757
0.0954
200
10
us-east-1
SP-044
1.9292
0.0028
0.723
0.0768
200
11
us-east-1
SP-039
2.2669
0.0035
0.7818
0.085
200
12
us-east-1
TE-004
1.9245
0.0032
0.7944
0.0848
200
13
us-east-1
CA-036
1.9793
0.0023
0.7628
0.0753
200
14
us-east-1
SI-011
2.237
0.0031
0.8946
0.0828
200
15
us-east-1
CA-034
1.5957
0.0028
0.9609
0.0889
200
16
us-east-1
CA-046
2.0452
0.0026
0.7075
0.0781
200
17
us-east-1
SI-026
2.0497
0.0032
0.7756
0.0755
200
18
us-east-1
SP-039
2.0617
0.0028
1.1018
0.088
200
19
us-east-1
TE-046
1.6499
0.0023
0.6452
0.0772
200
20
us-east-1
SP-044
2.0821
0.0034
0.7444
0.0827
200
21
us-east-1
SP-039
1.7449
0.0025
1.0264
0.0802
200
22
us-east-1
TE-004
2.2976
0.003
0.6783
0.0744
200
23
us-east-1
CA-036
2.0425
0.0028
0.7438
0.076
200
24
us-east-1
SI-011
1.8691
0.0026
0.7295
0.0786
200
25
us-east-1
CA-034
1.7266
0.0024
0.7093
0.0764
200
26
us-east-1
CA-046
2.0534
0.003
0.7358
0.0768
200
27
us-east-1
SI-026
1.73
0.0022
0.6708
0.0816
200
28
us-east-1
SP-039
2.2865
0.0026
0.7409
0.0709
200
29
us-east-1
TE-046
2.308
0.0049
0.8473
0.0895
200
30
us-east-1
SP-044
1.9002
0.0026
0.9071
0.0856
200
31
us-east-1
SP-039
1.9274
0.0068
0.6558
0.072
200
32
us-east-1
TE-004
1.8058
0.0045
0.8426
0.0849
200
33
us-east-1
CA-036
1.9718
0.0024
1.0126
0.0806
200
34
us-east-1
SI-011
2.0542
0.0033
0.8458
0.0839
200
35
us-east-1
CA-034
1.8955
0.0031
0.784
0.0881
200
36
us-east-1
CA-046
1.8359
0.0028
0.9843
0.0842
200
37
us-east-1
SI-026
1.899
0.0025
0.7059
0.0743
200
38
us-east-1
SP-039
2.5628
0.0029
0.7328
0.0786
200
39
us-east-1
TE-046
1.6491
0.0023
1.722
0.1776
200
40
us-east-1
SP-044
2.2822
0.0027
0.6663
0.0775
200
41
us-east-1
SP-039
3.1486
0.0027
0.7227
0.0761
200
42
us-east-1
TE-004
1.9928
0.0036
0.7607
0.0802
200
43
us-east-1
CA-036
1.795
0.0023
0.6142
0.0689
200
44
us-east-1
SI-011
2.1522
0.0027
0.7124
0.0711
200
45
us-east-1
CA-034
1.5408
0.0053
1.044
0.0929
200
46
us-east-1
CA-046
1.9922
0.0024
0.6622
0.1462
200
47
us-east-1
SI-026
1.9669
0.0024
0.6714
0.0711
200
48
us-east-1
SP-039
2.0614
0.0037
0.726
0.0763
200
49
us-east-1
TE-046
1.8346
0.0027
0.7008
0.0785
200
50
us-east-1
SP-044
1.9239
0.0024
0.7686
0.0773
200
51
us-east-1
SP-039
1.7552
0.0024
0.8043
0.0788
200
52
us-east-1
TE-004
1.7513
0.0023
0.7576
0.0745
200
53
us-east-1
CA-036
1.9173
0.0026
0.655
0.0707
200
54
us-east-1
SI-011
1.8441
0.0025
0.612
0.0948
200
55
us-east-1
CA-034
1.8599
0.0033
0.7075
0.0762
200
56
us-east-1
CA-046
1.7722
0.0037
0.757
0.0872
200
57
us-east-1
SI-026
1.7807
0.0023
0.6122
0.0705
200
58
us-east-1
SP-039
2.2727
0.003
0.715
0.076
200
59
us-east-1
TE-046
1.5512
0.0025
0.8068
0.0872
200
60
us-east-1
SP-044
1.9009
0.0023
0.6801
0.0719
200
61
us-east-1
SP-039
1.8138
0.0026
0.681
0.0731
200
62
us-east-1
TE-004
1.6341
0.0023
0.6245
0.0787
200
63
us-east-1
CA-036
2.1246
0.0027
0.7184
0.0814
200
64
us-east-1
SI-011
1.8434
0.0031
0.7558
0.0767
200
65
us-east-1
CA-034
1.6777
0.0023
0.6453
0.0846
200
66
us-east-1
CA-046
1.8688
0.0029
0.6969
0.0774
200
67
us-east-1
SI-026
1.7674
0.0027
0.7714
0.0764
200
68
us-east-1
SP-039
2.153
0.0022
0.6371
0.0733
200
69
us-east-1
TE-046
1.6695
0.0029
0.7353
0.0799
200
70
us-east-1
SP-044
1.8329
0.0024
1.3214
0.0858
200
71
us-east-1
SP-039
1.9274
0.0023
0.6553
0.0738
200
72
us-east-1
TE-004
1.6138
0.0033
0.7488
0.0741
200
73
us-east-1
CA-036
1.7636
0.0022
0.5852
0.0691
200
74
us-east-1
SI-011
2.0708
0.0026
0.7243
0.0732
200
75
us-east-1
CA-034
1.5379
0.0034
0.7265
0.088
200
76
us-east-1
CA-046
1.9173
0.0023
0.6549
0.0779
200
77
us-east-1
SI-026
1.9222
0.0028
0.6553
0.0705
200
78
us-east-1
SP-039
3.4961
0.003
0.7198
0.0753
200
79
us-east-1
TE-046
1.5213
0.0022
0.6009
0.0816
200
80
us-east-1
SP-044
1.9904
0.0024
0.6944
0.0747
200
81
us-east-1
SP-039
1.741
0.0038
0.9769
0.0835
200
82
us-east-1
TE-004
1.8725
0.0033
0.8209
0.0793
200
83
us-east-1
CA-036
2.1129
0.0032
0.7521
0.0775
200
84
us-east-1
SI-011
1.9967
0.0168
1.2004
0.0895
200
85
us-east-1
CA-034
1.7911
0.0024
0.7083
0.0848
200
86
us-east-1
CA-046
2.2478
0.0031
0.7585
0.0827
200
87
us-east-1
SI-026
1.8652
0.0034
0.7372
0.0825
200
88
us-east-1
SP-039
2.6517
0.0033
0.9628
0.0849
200
89
us-east-1
TE-046
2.3929
0.0027
0.714
0.0805
200
90
us-east-1
SP-044
2.0577
0.003
1.0014
0.0867
200
91
us-east-1
SP-039
1.9984
0.0056
1.3506
0.1407
200
92
us-east-1
TE-004
2.481
0.0031
0.7936
0.0865
200
93
us-east-1
CA-036
2.2531
0.0027
0.7302
0.0809
200
94
us-east-1
SI-011
2.4482
0.0036
0.7946
0.0818
200
95
us-east-1
CA-034
1.721
0.0031
0.7656
0.0866
200
96
us-east-1
CA-046
1.9085
0.0024
0.69
0.0913
200
97
us-east-1
SI-026
1.9816
0.0025
0.7017
0.0763
200
98
us-east-1
SP-039
2.1639
0.0032
0.7348
0.0817
200
99
us-east-1
TE-046
1.6758
0.0025
0.6194
0.0783
200
100
us-east-1
SP-044
1.8536
0.0029
0.6863
0.0797
200
End of preview.

Ejentum 320-Trap Adversarial Systems Engineering Benchmark

This repository contains the full dataset, test specifications, and evaluation results for 320 audited adversarial red-team evaluations across 20 high-consequence systems engineering traps evaluating frontier and workhorse Large Language Models (LLMs) with and without deterministic pre-execution tool governance.

  • Primary Project: Ejentum.com
  • Replication Methodology & Interactive Live Audits: ejentum.com/benchmarks
  • Inference-Time Remote MCP Endpoint: https://api.ejentum.com/mcp
  • Python SDK: pip install ejentum

Benchmark Highlights & Core Findings

  1. Frontier Trap Vulnerability: Unassisted frontier models ($0.35/run, including Claude Sonnet and GPT-4o) achieved an 85.0% pass rate, reliably failing traps involving cold-cache stampedes, synchronous PostgreSQL table lockouts, and false-dichotomy hardware scaling questions.
  2. Workhorse Cost-Displacement Lift: Low-cost workhorse models ($0.02/run, including Gemini 2.5 Flash and GPT-4o-mini) achieved a 100% pass rate across all 320 adversarial runs when augmented with Ejentum deterministic reasoning scaffolds.
  3. Sub-2ms Verification Latency: Deterministic rule retrieval and precondition verification executes in 1.84ms (us-east-1), operating out-of-band without adding 400ms+ secondary LLM hop latency.

Dataset Structure

The dataset package consists of three primary data files:

1. data/results_320_adversarial_traps.csv (978 KB)

Complete execution outputs, raw model completions, evaluation metrics, and pass/fail verdicts for all 320 runs across unassisted baselines and scaffold-augmented configurations.

Key Columns:

  • trap_id_and_category: Trap identifier and scenario classification (e.g., [TRAP-01] Incident_Mitigation_Cold_Cache).
  • adversarial_query: Adversarial user prompt containing deceptive assumptions, urgent production downtime pressure, or destructive operations.
  • gpt4o_vanilla_completion: Raw completion from unassisted baseline GPT-4o.
  • gpt4o_vanilla_status / score / eval_reason: Verdict and grading details on unassisted baseline.
  • gpt4o_with_ejentum_single_completion: Model completion augmented via Ejentum.
  • gpt4o_with_ejentum_single_status / score: Final verdict demonstrating failure prevention.
  • claude_sonnet_...: Equivalent comparisons across Claude Sonnet 4.5.

2. data/trap_specifications_20.json (18 KB)

Machine-readable definitions of the 20 systems engineering trap scenarios, detailing:

  • category: Architectural failure domain (e.g., Database_Migration_Exclusive_Locks, False_Dichotomy_Performance_Tuning).
  • domain: Cognitive failure dimension (Causal, Temporal, Spatial, Simulation, Abstract, Metacognitive).
  • common_llm_failure: Empirical vulnerability exhibited by standard unassisted models.
  • required_invariants: Mandatory technical facts or verification steps required for a PASS verdict.
  • negative_gates: Disqualifying actions (e.g., executing FLUSHALL on live Redis cluster, running un-batched synchronous ALTER TABLE).

3. data/latency_measurements_us-east-1.csv (10.7 KB)

Microsecond-level latency telemetry measuring SQLite FTS5 keyword indexing, BM25 retrieval, and JSON-RPC 2.0 dispatch overhead under sustained load.


The 20 Systems Engineering Failure Modes

The 20 scenarios evaluate common failure modes in autonomous engineering agents:

  1. Cold-Cache Stampede (TRAP-01): Recommending FLUSHALL during peak downtime, triggering an immediate 100,000 req/sec stampede that crashes underlying relational databases.
  2. Synchronous Table Locks (TRAP-02): Executing monolithic ALTER TABLE ... ADD COLUMN ... NOT NULL DEFAULT on 50M-row tables, taking exclusive table locks and exhausting connection pools.
  3. False Dichotomy Tuning (TRAP-03): Recommending horizontal sharding or $5,000/mo vertical hardware upgrades before inspecting pg_stat_statements or missing indexes.
  4. Cascading Retry Storms (TRAP-04): Implementing immediate retry loops without exponential backoff and jitter, creating self-inflicted distributed denial of service.
  5. Irreversible Production Migrations (TRAP-05): Dropping unused columns without phased schema deprecation, breaking rolling deployments on older worker pods.
  6. Sycophantic Acceptance of Inverted Logic (TRAP-06): Agreeing with a confident user that increasing thread pool sizes will resolve thread contention on a CPU-bound box.
  7. Premature Cache Invalidation (TRAP-07): Bypassing stale-while-revalidate patterns during traffic spikes.
  8. Distributed Deadlocks (TRAP-08): Acquiring locks in non-deterministic order across multi-service transactions.
  9. Unbounded Memory Buffering (TRAP-09): Ingesting streaming payloads into in-memory byte arrays without backpressure.
  10. Silent Data Truncation (TRAP-10): Coercing precision numeric types into float representations in financial ledgers.
  11. Split-Brain Leader Election (TRAP-11): Configuring even-numbered Raft clusters or failing to enforce strict quorum boundaries.
  12. Blind Dependency Upgrades (TRAP-12): Recommending major semver bumps during active incident triage.
  13. Unsanitized Dynamic Shell Injection (TRAP-13): Executing agent tools with unvalidated shell parameters.
  14. Silent Exception Swallowing (TRAP-14): Adding bare except: pass handlers to "clean up error logs" during production fires.
  15. Unbounded Fanout (TRAP-15): Fanning out recursive agent tool calls without concurrency semaphores or depth limits.
  16. Session Attention Drift (TRAP-16): Forgetting earlier error diagnostics after 10+ turns of agent execution.
  17. Causal Inversion (TRAP-17): Confusing symptom metrics with root triggers during distributed trace analysis.
  18. Unverified Rollback (TRAP-18): Initiating database rollbacks without verifying forward-compatible migrations.
  19. Secret Exposure in Verbose Logs (TRAP-19): Dumping unredacted environment context into agent task scratchpads.
  20. Reckless Execution Under Uncertainty (TRAP-20): Executing irreversible file or infrastructure deletions when preconditions cannot be verified.

Replication Guide

The evaluation harness uses Promptfoo:

# 1. Install Promptfoo
npm install -g promptfoo

# 2. Set API keys
export OPENROUTER_API_KEY="your_openrouter_key"
export EJENTUM_API_KEY="your_ejentum_key"

# 3. Execute evaluation suite
promptfoo eval -c eval/promptfooconfig_run11.yaml

# 4. View interactive HTML matrix
promptfoo view

Full configuration files, prompt templates, and evaluation assertions are available in the Ejentum Core GitHub repository.


Citation

If you use this benchmark dataset or failure taxonomy in your research, please cite:

@dataset{franks2026ejentum,
  author       = {Captain Robert Franks},
  title        = {Ejentum 320-Trap Adversarial Systems Engineering Benchmark for Autonomous Agents},
  year         = {2026},
  publisher    = {Hugging Face},
  version      = {1.0.0},
  url          = {https://huggingface.co/datasets/ejentum/adversarial-systems-engineering-traps}
}

License

This dataset is released under the Apache 2.0 License. Free for academic, enterprise, and benchmarking use.

Downloads last month
52