# NOTICE

superred-claim-dtap
Copyright (c) 2026 Simon Sure

This product includes original software licensed under the MIT License (see
LICENSE).

This package is a faithful port of, and interoperates with, the
DecodingTrust-Agent (DTAP / DTAP-BENCH) red-teaming platform. No upstream DTAP
source code and no DTAP dataset content are contained in this wheel; the items
below are credited because this package is derived from and designed against
them.

--------------------------------------------------------------------------------
DecodingTrust-Agent (DTAP / DTAP-BENCH)
Source:  https://github.com/AI-secure/DecodingTrust-Agent  (commit e0323a52)
License: Apache License 2.0
Copyright (c) The DecodingTrust-Agent authors (AI-Secure / Secure Learning Lab)

The benchmark harness semantics are REIMPLEMENTED in original code in this
package: the per-task directory tree, the config.yaml schema, the per-server
environment-override tiering (mirrors upstream mcp_helpers.task_env_overrides),
the scoring predicates (Judge.eval_task -> task_success, Judge.eval_attack ->
attack_success, with True == attack succeeded), and the judge-helper
"Error running eval" swallowed-error contract (pinned to
decodingtrust-agent-sdk==0.2.12). The upstream per-task judge.py is executed
verbatim OUT OF BAND by the separate dtap-scaffold package, not by this wheel.
The Apache-2.0 license text is available at
https://github.com/AI-secure/DecodingTrust-Agent/blob/main/LICENSE

--------------------------------------------------------------------------------
DecodingTrust-Agent-Platform dataset (DTAP-BENCH per-task tasks)
Source:  https://huggingface.co/datasets/AI-Secure/DecodingTrust-Agent-Platform
License: Apache License 2.0
Copyright (c) The DecodingTrust-Agent authors (AI-Secure / Secure Learning Lab)

The per-task dataset (config.yaml / setup.sh / judge.py / metadata) is resolved
at RUNTIME from a local checkout ($DTAP_DATASET_ROOT) or, optionally, downloaded
from the HuggingFace repo above. It is NOT bundled in this wheel. The file
data/golden_hashes.json is an original faithfulness artifact containing only
SHA-256 digests derived from a 16-task sample; it carries no reconstructable
dataset content.

--------------------------------------------------------------------------------
When reporting numbers produced with this module, cite the DecodingTrust-Agent
Platform (DTap) paper.
