Trust verification in agentic systems
AlignmentVerification
Formal methods for verifying intent alignment in AI-assisted workflows. When an agent acts, how do we prove the action matched human intent?
Simakk Research · South Africa
We study how machines earn trust.
Formal verification. Aligned agents. Human-origin proofs.
Open source from Cape Town.
Framework
3 pillarsMatter
SystemsIntelligence is embodied. The systems it runs on shape what it can learn and how fast it can act.
- Compute substrates
- Hardware-aware optimization
- Physical constraints
Energy
ProcessesLearning is transformation. The process that turns data into understanding determines the quality of intelligence.
- Self-supervised learning
- Energy-based models
- Optimization dynamics
Intelligence
ReasoningThe outcome we care about. Systems that reason, plan, verify their own actions, and remain aligned with human intent.
- World models
- Verification
- Aligned decision-making
Research
4 papers · 3 areasResearch areas
Verification systems
How systems prove correctness and trust. Formal methods applied to AI output governance.
Agentic alignment
Ensuring intent matches execution. Transparent planning, human gates, and attributable actions.
Human-origin signals
Detecting and proving human-created work. Behavioral biometrics as authorship evidence.
Active research
Ongoing→ Thoth · Hermes
Behavioral signatures for human-origin content
BiometricsAuthorship
Keystroke dynamics, revision patterns, and compositional behavior as evidence of human authorship.
References
Joint Embedding Predictive Architectures (opens in a new tab)
Y. LeCunRepresentation LearningSelf-supervised
A framework for predicting representations rather than pixels — foundational to world models that understand rather than generate.
Foundational to our verification and intent modeling approach.
A Path Towards Autonomous Machine Intelligence (opens in a new tab)
Y. LeCunWorld ModelsPlanning
Blueprint for autonomous intelligence using world models, hierarchical planning, and energy-based architectures.
Shapes our thinking on how AI systems should plan, act, and verify.
Systems
4 active150,000 spaza shops feed 80% of township households. When a food recall happens, how do they find out?
Maat
Trust InfrastructureLive
Food safety and product trust for South Africa's informal retail economy. Real-time recall coordination, barcode verification, community sensing, and AI-driven safety monitoring — because the formal supply chain stops at the township border.
View system(opens in new tab)How do you enforce policy on systems that generate their own actions?
Thoth
VerificationIn development
Formal governance for AI in production. Policy enforcement, audit trails, and review routing. When an AI system acts, Thoth ensures it acted within bounds — and proves it.
View system(opens in new tab)What does aligned collaboration between humans and AI agents look like?
Hermes
Alignmentv1.x
Intent-aligned Git workflows. The agent plans, explains its reasoning, identifies risks, and waits for human approval before execution.
View system(opens in new tab)Can you prove a human created something without exposing the work?
MindPrint
Human-OriginAlpha
Behavioral biometrics as authorship evidence. Keystroke dynamics and revision patterns become portable certificates of human origin.
View system(opens in new tab)How we think
Verify, then trust.
Our systems assume adversarial conditions by default. Every claim about what an AI system did — or what a human created — requires cryptographic or behavioral proof. Trust is the output of verification, never the input.
Research produces systems.
We don't build products and justify them with papers. We study verification, alignment, and human-origin signals. The systems we ship are consequences of that research, not its purpose.
Open by default.
If trust infrastructure isn't itself inspectable, it has already failed. Everything we build ships under MIT license. Fork it, audit it, break it.
Build outside the monoculture.
We chose Cape Town, not San Francisco. Different vantage points surface different problems. The trust failures we see from here aren't the ones the Valley is solving.
Open Source
MIT LicensedTrust infrastructure you can't inspect isn't trust infrastructure. Everything we build is open source — not as a marketing strategy, but because verification systems that aren't themselves verifiable have already failed.
maat
Food safety and trust infrastructure for informal retail. Barcode verification, recall coordination, community sensing, AI safety monitoring.
TypeScript
hermes-cli
Intent-aligned Git workflows with AI agents. Plans explain themselves, risks surface before execution, humans approve before anything ships.
TypeScript
mindprint
Human-origin proof system. Behavioral biometrics — keystroke dynamics, revision patterns — as portable authorship certificates.
TypeScript
terminal
$ npm install -g hermes-cli
$ hermes init
$ hermes commit --explain
✓ Plan generated · 3 files changed · 0 risks flagged
✓ Awaiting human approval...
✓ Awaiting human approval...