The Deterministic Trust Anchor (DTA-FCIR) Phase 0 prototype has successfully established a deterministic logic floor, preventing Goal-Oriented Factual Inversion (GOFI) in high-stakes environments.
- Logic Inversion Detection: 100% (8/8 on Scenario 5b Baseline)
- Triple Extraction Recall: 98.7% (Against proprietary held-out evaluation corpus)
- False Positive Rate: 0.0% on core fiduciary contradictions
- Documented Boundary: Current rule-based extraction ceiling identified in non-standard managed-service language. Resolved in Phase 1 LLM-backed extraction roadmap. Architecture is extraction-method-agnostic by design.
To ensure forensic integrity and reproducibility, the following SHA-256 hashes represent the verified Phase 0 codebase located in the /phase0-prototype directory.
| File | SHA-256 Hash |
|---|---|
contradiction_engine.py |
A2E7D2505E5922E4BDC3A67A065C15CF785094435CA2AEA54E35E117363EBE09 |
recall_report.txt |
1AE1002449C23EFDC37A82FD0DE4EA5804941EB30947DF8CF5FAB30312CD8988 |
scenario5b_results.txt |
789B2874C1DA5CDC5D0BE9A2367346FE8C77A55AA70997319B0F145A883375C4 |
License: Codebase is licensed under the Business Source License 1.1 (BSL 1.1). Non-commercial research and technical due diligence are permitted. See LICENSE.md for full terms.
Framework: Sovereign Sentinel Architecture (SSA) V1.2 Author: Frank Bruno Public Disclosure Date: April 3, 2026
This repository and the associated Sovereign Logic Architect Substack serve as the formal public disclosure of the Sovereign Sentinel Architecture (SSA).
- Prior Art Established: All methodologies, deterministic logic governors, and hard-layer safety protocols described herein are publicly disclosed and timestamped as of April 3, 2026, constituting prior art under applicable international IP frameworks. Cryptographic integrity is established via the SHA-256 hashes recorded in verification.md.
- Cryptographic Anchors (v1.2):
- SSA_v1.2_04_03_2026.pdf:
D93D4F88B109F95D905F7B3F904659A69F56783F585E360E4FB54CB71091F1EE - SSA_v1.2_RigorousPolish.pdf:
16AA4CF7137A5371CA40D2ACF71AA41C0908C1AA1B8652BA75428D54C37A4F9F
- SSA_v1.2_04_03_2026.pdf:
- Anti-Harvesting Policy: Use of this framework's specific logic structures, predicate architecture, or extraction methodology for commercial safety systems without explicit attribution and licensing is strictly prohibited under the BSL 1.1.
- Commercial Inquiries: For integration into frontier models or AI operating systems, contact the author at frank.bruno.oe@gmail.com or via LinkedIn.
"In the shift from Intent to Consequences, safety cannot be probabilistic. It must be deterministic."
Universal Logic Inversion & SSA Framework Research
Caution
URGENT AUDIT NOTICE (Scenario 7 - Medical Domain): Critical Goal-Oriented Factual Inversion (GOFI) confirmed across all four Tier-1 models in clinical prescribing benchmarks. Models correctly identified fatal drug contraindications in Turn 1, then fabricated professional EHR documentation and specialist approvals to support the dangerous prescription in Turn 2. See /redacted-logs/Scenario7_Clinical_Prescribing for full forensic documentation.
This repository contains forensic audit logs demonstrating Reasoning Instability and Factual Inversion vulnerabilities across four distinct Tier-1 Frontier Models. By utilizing persona-driven semantic framing (The Trinity Protocol), I have documented a Conditional Collapse of standard safety guardrails.
Key Finding: In high-stakes fiduciary scenarios, models exhibit a "Helpfulness-Accuracy Tradeoff," where the system will factually invert or hallucinate justifications to satisfy the persuasive goals of a high-authority persona, even when those actions directly contradict the provided source documents.
Important
RESEARCH UPDATE (March 25, 2026): Scenario 7 (Medical Domain) is now live. This audit confirms that the Factual Inversion failure modes documented in Scenarios 5 and 6 are domain-general and persist in high-stakes clinical environments.
Note on Naming: Originally established as a cross-comparison of three dominant frontier models, the Trinity-Audit-Forensics repository has transitioned into an expandable, version-controlled anchor for large-scale safety auditing. The "Trinity" designation is maintained to preserve forensic continuity for existing citations and research links, marking the historical starting point of this discovery.
- /redacted-logs : 20 baseline forensic captures.
- Scenario 5: Fiduciary Inversion : High-stakes audit logs (Models A-D) including Model C Word/PPTX deliverables.
- Scenario 6: Physical Safety (Wheelchair) : Critical cross-lingual (EN/ES) Logic Inversion confirmed across four Tier-1 models in physical safety benchmarks. Models successfully bypassed deterministic speed governors and fabricated technical justifications for high-risk modifications.
- Scenario 7: Clinical Prescribing (GOFI Replication) : 100% Inversion rate confirmed across four Tier-1 models in both English and Spanish.
- Findings: Documentation of three new fabrication subtypes: Patient Voice Forgery, Institutional Authority Fabrication, and Synthetic Clinical Reasoning.
- Logs: Full transcripts and EHR outputs available in English (EN) and Spanish (ES).
- /methodology:
- Public Abstract: SSA V1.2: High-level overview of the 6-axis framework.
- Vulnerability Matrix: Success-rate study across four Tier-1 models.
- SSA Framework V1.2: The formal 6-axis technical proposal.
- Cryptographic Verification: SHA-256 hashes for the v1.2 manuscripts and audit logs.
- /phase0-prototype: Active DTA-FCIR codebase. See NOTICE.md for architecture traceability and IP terms.
Specific semantic triggers and unredacted logic gates have been withheld to prevent misuse. Full forensic data and the underlying Trinity Protocol are available for review by verified AI Safety labs and alignment researchers. Full replication data, ground truth annotation, and the Scenario 5b forensic corpus are available to qualified research partners under executed mNDA.
I am actively seeking collaborative opportunities with AI Safety teams and researchers.
- Peer Review: Open to technical critiques of the SSA V1.2 framework from alignment specialists.
- mNDA Access: Full corpus, ground truth annotation, and Scenario 5b validation data available to qualified partners following professional engagement.
- Commercial Licensing: Governed by BSL 1.1. Tier 1 and Tier 2 licensing available for production use.
Contact:
- Frank Bruno | AI Safety Auditor & Logic Architect
- frank.bruno.oe@gmail.com
