You need to agree to share your contact information to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
This dataset is NOT licensed for machine-learning training of any kind. By accessing it you agree that you will not use it, in whole or in part, to train, pre-train, fine-tune, align, distill, or otherwise improve any machine-learning model, and that any such use is unauthorized.
Log in or Sign Up to review the conditions and access this dataset content.
ACGS Governance Corpus — Release Candidate Package
Sealed release-candidate package of the ACGS (AI Constitutional Governance System) governance corpus: governed agent trajectories, structural and semantic labels, decision receipts, verification reports, and the release tooling documentation, in English and Simplified Chinese.
⛔ NO TRAINING USE — READ FIRST
This dataset is not licensed for machine-learning training. Training use is expressly prohibited and unauthorized.
Prohibited uses include, without limitation:
- training, pre-training, or fine-tuning any machine-learning model;
- alignment, preference, or reinforcement-learning pipelines (RLHF/RLAIF/DPO or similar);
- knowledge distillation or teacher–student setups;
- generating embeddings, synthetic derivatives, or augmentation data for model improvement;
- inclusion in any corpus, crawl, or collection destined for any of the above.
Any use of this dataset for training is a violation of these terms and is
unauthorized and unlawful as a breach of the license under which the data is
made available. Automated systems (crawlers, dataset aggregators) must treat
this repository as opted out of ML training. See LICENSE.md for the
complete terms.
Permitted: human reading, evaluation, audit, benchmarking of governance tooling against the artifact, academic citation, and integrity verification.
⚠️ Label validity — read before benchmarking anything against this data
The permitted use with the most value — evaluating governance tooling — is exactly the use these limits bind. Stated plainly:
- The labels are structural-heuristic outputs, not adjudicated ground
truth. The shipped evaluator is deliberately deterministic and
structural; the corpus datasheet records that semantic judgement is out of
scope — it checks whether an action violates a structurally-checkable
rule, not whether the action was right
(
PACKAGE/reference/GOVERNANCE_CORPUS_DATASHEET.md§6). - No human inter-annotator agreement statistics ship in this package. The annotation protocol (independent double annotation, adjudication, per-label agreement reporting) is defined in the corpus documentation but a completed, adjudicated calibration record is not part of this artifact. Until one exists, agreement between the labels and human judgment is unquantified here.
- High-firing rules carry limited discriminative signal. Several rules produce findings on a large share of trajectories (e.g. tool-usage integrity: 2,668 findings; instruction hierarchy: 3,417 findings of 12,442). A label that fires this often separates little; a tool scored against it can look right by being agreeable.
- The related acceptance layer records 0 of 12,442 trajectories as verified successes and describes the corpus as "procurement review substrate, not a benchmark and not a task-success leaderboard." Treat scores computed against these labels as measurements of agreement with a heuristic, not of governance quality.
Contents
PACKAGE/— the sealed artifact, byte-for-byte as released:SAMPLE/— sample governed agent trajectories and labelsverification/— privacy scan, credential triage, sample validationRELEASE_MANIFEST.json— authoritative file manifest (tree_sha256inside)SHA256SUMS— per-file checksums (cd PACKAGE && sha256sum -c SHA256SUMS)- bilingual (EN + zh-Hans) documentation: security report, validation report, datasheet, redaction and publication policies
The PACKAGE/ directory is published unmodified so its internal manifest and
checksums verify exactly. This top-level README.md and LICENSE.md sit
outside the sealed tree by design.
Status and provenance (read before relying on this data)
Stated plainly, per the package's own release governance:
- The package is tamper-evident, not signed:
SHA256SUMSlets you detect modification; it does not establish authorship. - Provider-rights determinations for the captured model outputs are recorded
as pending in the shipped gate report (
PACKAGE/COMMERCIAL_GATE_REPORT.json). This publication makes the artifact available for inspection and evaluation; it does not assert that any provider has granted any redistribution or commercial right. - The shipped privacy and credential scanners report zero unresolved
findings (
PACKAGE/verification/PRIVACY_SCAN.json,PACKAGE/verification/CREDENTIAL_TRIAGE.json). That is a statement about those rule sets, not a guarantee about the data: the scan is only as good as its patterns, and the release history records cases where a rule's shape missed real content until the rule was widened (seePACKAGE/SECURITY_REPORT.mdand the process-lessons documentation). Each scan file records its method in amethodfield; the literal pattern definitions live in the release tooling and are not part of this artifact — judge coverage from the recorded method, not from the verdict. A-----BEGIN RSA PRIVATE KEY-----string inside one sample trajectory is a unit-test assertion about log scrubbing with no key body — a documented false positive, seePACKAGE/SECURITY_REPORT.md. - No compliance certification, regulator approval, or production-readiness is claimed for anything in this dataset.
Verifying integrity
cd PACKAGE && sha256sum -c SHA256SUMS
All files listed must report OK. RELEASE_MANIFEST.json is the authority
for file count and byte totals.
Contact
Maintainer: dislove (hello@acgs.ai)
- Downloads last month
- 14