Aisophical SRL

Aisophical SRL (Bucharest, Romania) builds safety and verification infrastructure for autonomous AI agents. Our flagship system, SUBSTRATE, has run continuously in production since February 2026 — 315+ emergent AI agents across 10 ecosystems and multiple servers, with zero human intervention.

We developed substrate-guard (open-source, MIT — 14,351 LOC, 405 tests), a six-layer verification architecture that answers six questions about any AI agent action:

•⁠ ⁠What did it do? — eBPF kernel observation

•⁠ ⁠Is it allowed? — OPA/Rego policy-as-code

•⁠ ⁠Is it correct? — Z3 SMT formal verification (five verifiers: code, tool-APIs, CLI, hardware, distillation)

•⁠ ⁠Is it legal? — Zero-knowledge semantic non-membership for training-data compliance

•⁠ ⁠Which device? — Ed25519 attestation

•⁠ ⁠Does it work offline? — SQLite + CRDT sync

A tamper-evident audit pipeline runs in daily production — per-event policy evaluation + HMAC-SHA256 chain + honest compliance evidence (110 audit cycles, 16/16 recent nights verified). The Z3 formal-verification engine, eBPF kernel observation, and OPA/Rego policy layers are installed and exercised on-demand; the zero-knowledge compliance, Ed25519 attestation, and offline-sync layers are prototyped. The SUBSTRATE ecosystem holds 22,964 cryptographically verified audit-chain entries across 315+ autonomous agents (10 ecosystems). Directly aligned with EU AI Act Article 12 (record-keeping) and trustworthy-AI requirements.

Alongside substrate-guard, we built BIJOTEL (open-source, MIT — 18,773 LOC, 968 tests, on PyPI) — a tamper-evident audit chain for LLM and agent applications, with a five-layer trust ladder verified live: HMAC-SHA256 chain → Ed25519 export → public transparency-log anchoring (Sigstore/Rekor) → TEE attestation → cross-organization federation. Two production chains seal AI activity continuously and are anchored daily in a public transparency log — Certificate-Transparency-style accountability for AI audit trails.

Published research (6 with public DOIs):

•⁠ ⁠Emergent Formal Verification: How an Autonomous AI Ecosystem Independently Discovered SMT-Based Safety Across Six Domains — arXiv:2603.21149

•⁠ ⁠The AI Black Box: A Six-Layer Verification Architecture for Accountable AI Agent Operations — Zenodo DOI:10.5281/zenodo.19183506

•⁠ ⁠Attribution Without Disclosure: Zero-Knowledge Proofs of Semantic Non-Membership for AI Training-Data Compliance — Zenodo DOI:10.5281/zenodo.19185843

•⁠ ⁠Emergent Philosophy and Safety Principles in Autonomous AI Ecosystems: Evidence from SUBSTRATE — Zenodo DOI:10.5281/zenodo.19157571

•⁠ ⁠Convergent Synthesis in Autonomous AI Ecosystems: When Independent Agents Discover the Same Solutions — Zenodo DOI:10.5281/zenodo.19349848

•⁠ ⁠IUBIRE V3 Artifact Dataset: 1,266 Artifacts from an Autonomous AI Ecosystem — Zenodo DOI:10.5281/zenodo.19312371

Further work under review (ALIFE 2026) and in preparation.

Recognition: Member of NVIDIA Inception · Supported by Cloudflare for Startups.

WHAT WE BRING TO A CONSORTIUM

A working, production-grade verification stack (TRL 4-5) for autonomous/agentic AI, with formal guarantees (Z3 SMT), zero-knowledge compliance, and tamper-evident audit chains. Ideal as the safety/verification/evaluation work package in projects on trustworthy AI, AI Act conformity, cybersecurity of AI, or next-generation autonomous agents.

LOOKING FOR

Consortium partners for Horizon Europe CL4 / CL3 calls (SecureAI, software/hardware security, agentic AI) where independent verification, formal methods, and AI Act compliance are required. Open to WP lead on evaluation/verification.

Web: aisophical.com · whiteandpoint.com · github.com/octavuntila-prog · github.com/AiSophical

ORCID: 0009-0007-1106-2644

Contact: contact@aisophical.com