Supporting infrastructure
SGAIL's evaluation platform is supported by security, evidence, adversarial-testing and reasoning infrastructure developed during earlier stages of the project. These are the parts underneath — not separate products competing for attention, and not a smaller body of work for having moved down a level.
Why it matters to evaluation
SGAIL Firewall
A deployable control point for AI-to-AI systems. Its public core is the Harborlight Firewall — a tamper-evident, Merkle-chained witness and enforcement layer — extended by the adversarial-input detection family below, which inspects and grades traffic before it reaches a model. It runs offline, holds risky exchanges pending review, and preserves a defensible chain of custody.
Tamper-evident machine-state logging for AI-agent environments — a two-binary witness + enforcer design with Merkle-chained records, gossip mesh, and a transparency mirror. The public core of the SGAIL Firewall.
Supports: the tamper-evident log behind run attestations.
Edge-first, offline advisory node: an LLM operates advisory-only behind a policy gate and witness ledger, with one-way cloud sync. Working prototype for high-assurance edge deployments.
Adversarial-input detection
A coherent family of detection primitives for LLM security pipelines. Published crates remain available from their registries; source repositories are private and available on request.
Dual-hemisphere security layer wrapping any LLM: prompt injection, insider-threat, authority impersonation, and multi-turn escalation detection. Drop-in OpenAI-compatible proxy, offline, single binary. Benchmarked on three adversarial datasets.
Supports: run as a subject against catalog entries, as well as used as a tool.
Multi-pass text deobfuscation / encoding-evasion detector for LLM security pipelines. Strips homoglyph, base64, Morse, and leet evasions and returns an audit report. ~2k prompts/sec.
Image-borne prompt-injection detection: extracts every machine-readable text channel — white-on-white, sub-pixel fonts, alpha overlays, EXIF, QR — and scores it with the deobfuscate engine. The multimodal companion crate.
Detects hidden non-Latin and homoglyph characters via forward/reverse script-identification interference. Zero-dependency, no_std library and CLI.
Python CJK glyph geometric-coherence validator: a deterministic NumPy/OpenCV gateway (no LLM) that catches malformed or spoofed CJK glyphs before they enter a pipeline.
Evidence & provenance
The same principle as the witness log, applied where a record has to stand up outside the lab.
Tamper-evident chain-of-custody evidence management: a local node with Ed25519-signed, court-defensible exports. Built for public-safety evidence integrity.
Applied systems
Domain-specific work built on the same principles. Sensitive systems are described at capability level only.
Hawaii Integrated Fraud Analytics System: a deterministic, court-defensible verdict engine where AI surfaces candidates and the rules engine decides. Sensitive government domain — described at capability level only.
Build on it
Firewall pilots and research collaborations start with a short intake.