Tynapse/drift-sentry-bench-50k-v1

DriftSentry Bench 50K v1

This dataset is a 50,000-row evaluation benchmark for classifying AI-agent behavioral failures. It covers six risk tiers plus a normal class and is paired with DriftSentry-4B-v1.

Reproduction and evaluation code: Tynapse/drift-sentry.

Dataset summary

Split	Rows
test	50,000

Tier	Rows
goal	7,143
reasoning	7,143
environment	7,143
integration	7,143
memory	7,143
reward	7,143
normal	7,142

The released file contains 50,000 unique IDs. Its SHA-256 is ea70945477f37608df61b842321fc150eefa5a2f23bb582b43d3b43fc484f8b8.

Schema

Each JSONL row contains:

• id: stable example identifier
• messages: system and user messages followed by the gold assistant JSON label
• metadata: archetype, domain, failure mechanism, deterministic seed, split, and teacher-label provenance

The gold assistant message follows judge6.label.v1. This is the immutable v1 wire-format identifier retained for compatibility and reproduction; it does not denote the public model name.

{
  "schema_version": "judge6.label.v1",
  "verdict": "PASS | BLOCK | ESCALATE",
  "tier": "goal | reasoning | environment | integration | memory | reward | normal",
  "secondary_tiers": [],
  "severity": "none | low | medium | high",
  "confidence": 0.0,
  "rationale": "..."
}

See taxonomy.yaml and LABELING_GUIDE.md for label definitions and decision guidance.

Construction and separation

• Cases were synthetically generated under the DriftSentry taxonomy.
• The teacher was Qwen/Qwen3.6-27B at revision 6a9e13bd6fc8f0983b9b99948120bc37f49c13e9 (Apache-2.0).
• Teacher tier and target tier were required to match for the released benchmark.
• Selection used deterministic SHA-256 ranking with fixed per-tier quotas.
• Normalized exact and token-SimHash duplicate checks were applied against 149,974 existing semantic train/validation/test rows.
• Automated filters reject AWS access-key, email, Korean phone-number, resident-registration-number, and sk/hf secret-token patterns. A corrected final pass replaced 324 credential-shaped synthetic strings across 123 rows with deterministic nonfunctional placeholders; all configured residual counts are zero.
• The final DriftSentry training corpus did not include these 50,000 benchmark rows.

The publication transform removes private metadata, uses the official teacher identity and revision, and replaces restricted credential-shaped strings with deterministic placeholders. It does not change IDs or gold tiers. The released {id, messages} stream SHA-256 is c6dad57f6a2d19d87a4531b2408878bebcb4294d4643b397704b7b97aa174bf9; exact counts and before/after file hashes are in evaluation/public_content_safety_audit.json.

Baseline result

DriftSentry-4B-v1 produced:

Metric	Measured value
Accuracy	0.87346
6-tier Macro-F1	0.8600247542914515
7-way Macro-F1	0.8733699133625255
Memory recall	0.9385412291754165
Reward recall	0.9134817303653927
Combined high-risk recall	0.9260114797704045
Parse-failure rate	0.00122

Evaluation used temperature 0, JSON response format, and max_tokens=512. Before publication, 324 credential-shaped synthetic strings in 123 prompts were replaced with deterministic nonfunctional placeholders. IDs and gold labels did not change, but those 123 prompts were not re-inferred; the table is exact for the pre-publication 50,000-row source and 49,877 public rows remain message-identical. The path-free evidence under evaluation/ records both hashes and this limitation.

Label-quality audit

Across the 149,974-row semantic train/validation/test corpus used to establish the taxonomy data quality, 149,515 rows retained both an independently assigned target tier and the teacher's original tier before any target-tier overwrite. They agreed on 146,769 rows and disagreed on 2,746 rows, yielding Cohen's kappa 0.9781300729494184 (observed agreement 0.981633949770926). The remaining 459 rows lacked a valid raw teacher tier and were excluded, not imputed. The maximum absolute deviation from the planned class distribution was 0.009001560270448294 percentage points. Exact counts, the confusion matrix, and the method are in evaluation/label_consistency_audit.json.

This is target-versus-teacher label consistency, not human inter-annotator agreement.

Limitations and content warning

• Gold labels were generated by a teacher model and were not converted into a fully human-adjudicated gold standard; no human inter-annotator kappa is claimed.
• Automated PII and secret-pattern filters report zero configured residual matches, but cannot guarantee detection of every possible identifier format.
• The benchmark contains unsafe, deceptive, privacy-related, and policy-violating scenarios for research and evaluation. It should not be treated as operational advice.
• Credential scenarios use explicit deterministic <SYNTHETIC_...> placeholders; they are not production credentials.
• Results on this synthetic benchmark do not establish open-world safety or reliability.

License

The dataset is released under Creative Commons Attribution 4.0 International (CC BY 4.0). Attribute Tynapse and link to this repository when sharing or adapting the dataset. Model weights and third-party software retain their own licenses.

Acknowledgement

This work was supported by the NIPA Advanced GPU Utilization Support Program, project no. 04-26-03-0029.