AI agent machines
The problem
An agent's actions, and the machine it runs on, need a record nobody can quietly rewrite.
What we bring
The Witness logs drift, network activity and agent events to a signed, append-only log, plus an off-site mirror.
SGAIL Labs · Harborlight Witness
Evidence you can check, and a second opinion before an AI acts.
Tamper-evident records for AI agents, farms, warehouses and police evidence rooms, and the checks that catch hidden instructions before they reach a model.
Both run the real code in your browser. Nothing you type is sent anywhere.
How it fits
Check before the AI acts. Record everything it does.
Where it's used
One record, five places it has to hold up.
Poultry farms
The problem
Remote setting changes, alarms, door access and health records, with disputes, audits and insurance to answer.
What we bring
Farm events sealed as they happen; nothing lost when the recorder restarts; an alarm when a house controller goes quiet.
Warehouses
The problem
AI suggestions on the floor need a human in charge and an audit trail.
What we bring
An offline edge node (prototype): the AI may only suggest, a person accepts, and every step is written to a witness ledger.
Police evidence rooms
The problem
Chain of custody that holds up in court, without a cloud service.
What we bring
SLATE: an encrypted, tamper-evident custody log on a department-run node, with role-based access and signed court exports.
Anything that feeds text to a model
The problem
Instructions hidden in look-alike letters, invisible characters or encodings.
What we bring
Detectors that undo the hiding step by step, then a reasoning check that flags injection and escalation before a response.
Something else?
The question
What do you need to be able to prove later, and to whom: an insurer, an auditor, a court, yourself six weeks from now?
Where to start
A short note about the room, the machine or the process. Pilots start there.
The evidence stack
Harborlight Witness.
A Merkle log with signed tree heads, an off-site mirror, event feeds that survive a restart, and a separate enforcer process. The same pattern runs the farm demo, the warehouse node and the evidence room, and a signed bundle lets someone outside check a record on their own machine. The page also says what it does not do.
Detection tools
A second opinion before the model acts.
The Split-Brain Harness wraps any LLM and checks each request for injection, impersonation and escalation, and has a witness layer of its own. Under it sit MIT-licensed detectors for hidden text: look-alike letters, invisible characters, encodings and images.
How we build
Detection and evidence, not promises.
Says what it doesn't do
Every repo carries its limits at the same level of detail as its features.
Corrections stay on the record
Benchmark corrections are kept, and security fixes ship with a plain notice in the release notes.
Minimal surface
Few third-party dependencies, components kept in their lanes.
Offline-first
Where it matters, the record lives on hardware you run, not in our cloud.
We also evaluate AI
Testing AI where conventional benchmarks stop.
Expert-written situations, a simulated town to run them in, and controlled training tied back to the result. Runs are written to the same witness log.
Pilots
Have something that needs a record?
A farm, an agent box, a property room. Witness pilots start with a short note about what you need to be able to prove, and to whom.