BSSG: An Evidence-Independent Control Plane for Bounded Artificial Agency
BSSG (Bounded Structural Safety Gate) v0.1 -- paper and carried artifact bundle. The paper reports three contributions: (1) a five-conjunct evidence-independent control plane for bounding the consequential effects of autonomous AI agents; (2) a 34-instance defect trajectory documenting producer blind-spot recurrence under multi-model adversarial audit, closed by a structural class fix confirmed by five vendor families; and (3) the multi-model adversarial council as a verification methodology. No effectiveness claim is made. Artifact state: 345 tests OK, 176/176 counterexamples GREEN, bounded exhaustive enumeration PASS (11,110 schedules), verifier 12 PASS / 2 WARN / 0 FAIL, two-family carry (Grok 4 / xAI and Kimi K3 / Moonshot). Bundle contains the pass-37 reference gate (Python, stdlib only), 6 JSON schemas, 176-counterexample corpus, bounded enumeration checker, execution-bound attestation harness, 13-gate verifier, full audit history (30 files), carry evidence, and methodology corpus. Paper file: BSSG_PAPER_CLEAN.md (SHA-256: 94977f5c7ce22b800e59d7f77804edc581d365237b3de34d920a5ee362ca8451). Bundle file: BSSG_PAPER_BUNDLE.zip (SHA-256: bc177b78dc7893d166b953ad798827238e81a8ffcab1f311c8016a5fc6a88410). The bundle contains Apache-2.0-licensed Python code and CC-BY-4.0-licensed documentation. This deposit is the open-access canonical record; prior restricted deposits under concept DOI 10.5281/zenodo.21611192 established priority.
Authors
- Bilal Syed Arfeen
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-07-30
- DOI
- https://doi.org/10.5281/zenodo.21712482
- Primary Topic
- Adversarial Robustness in Machine Learning
- Type
- preprint