Technology you can check, not technology you have to trust.
Responsible Systems Lab builds software where the safety, privacy and accountability properties come from how the system is made — verifiable in the source, in an evidence file, or in a published test — rather than from a promise about how it will be operated. Four released open-source projects, all Apache-2.0.
We build technology whose claims about itself can be checked by someone who does not trust us.
Responsible Systems Lab develops and studies practical technologies that make complex systems safer, more humane, and more accountable — with a working focus on the systems now making decisions on people's behalf. That means software that keeps sensitive data on the machine it was created on, tooling that produces portable evidence of what an AI run actually did, frameworks that force a decision to carry its citations, and orchestration that reports what it could not do instead of quietly claiming success.
The same standard extends to where technology affects people, communities, animals and the environment: an automated judgment about a living system should be inspectable, and should hand back to a person rather than manufacture confidence the evidence does not support.
We are an informal, interdisciplinary project group working on early-stage ideas, prototypes, and research-driven concepts — not an anti-technology campaign, and not a company selling a finished product. Our aim is technology that earns trust through transparency, human oversight, and practical risk reduction.
Four released projects, and the standard they share.
Different problems, one test: in each, the responsible property is a structural fact you can confirm — no network call, a hash you can verify, a citation the validator enforces, a ceiling the scheduler measures — rather than a policy you have to believe. All four are open source under Apache-2.0.
YazSes
Offline voice dictation for Linux, macOS and Windows. Hold a key, speak, release — transcribed on your own CPU and typed into the focused window. No network call in the dictation path, no account, no API key. Includes a Dysfluency-Friendly Mode for stuttered or dysarthric speech and an EMG trigger for hands-free use.
- v2.18.2
- Apache-2.0
- 12 contributors
- Archived with a DOI
NovaFabric
Captures, replays, diffs and audits AI agent, model and HPC runs as portable evidence capsules on your own disk — no code changes to what is observed, no accounts, no telemetry. Bundles verify offline with sha256sum and an ed25519 verifier, so an auditor needs neither the tool nor your permission.
- v0.101.0
- Apache-2.0
- Self-hosted
- Laptop to cluster
Scrutable
A decision and assurance framework: 199 questions in a typed activation graph, scoped by ten profile questions to the domains your subject actually warrants, answered with cited evidence, ending in exactly one defensible gate decision. Markdown and YAML — nothing to install, and an AI agent can run it.
- v0.8.0
- Apache-2.0
- 199 questions
- Agent-executable
agent-fleet
Runs many coding-agent sessions against one shared task queue, with parallelism measured from what the machine can spare rather than read off the queue length. Claiming is a single atomic rename(2), so workers cannot duplicate work — and anything needing a human is parked with a reason instead of faked.
- v0.1.0
- Apache-2.0
- No dependencies
- 67 self-test checks
Also in development: SOTAForge, a citation-enforced pipeline from research papers to build-ready specifications, and welfare-aware verification work including a tested prototype. Neither is listed as a released project, because neither has a public artifact you could check yet — see the projects page.
Powerful systems need accountability designed in — not added later.
AI agents, automation, robotics, sensing platforms, and cloud systems are becoming part of everyday life, and increasingly they act rather than advise. They can improve society, but they also create risk when nobody can reconstruct what one of them did, on whose authority, or with what evidence. We work on the practical mechanics of making that reconstructable: portable records, inspectable reasoning, stated limits, and a human left in a position to decide.
Where we work.
Six themes guide our projects. Each one already has work behind it, or is named here as something we are building toward rather than something we have done.
Responsible AI & agentic systems
Agents that stay understandable and accountable — bounded by what a machine can actually do, honest about what they could not finish, and stopped at the point where a human must decide.
Provenance, replay & audit evidence
Making it possible to prove afterwards what an AI or HPC run was given and what it returned — as a portable record the auditor can verify without the tool that produced it.
Decision assurance & governance
Turning “are we ready?” into a question with a defensible answer: the right questions for this subject, each answer carrying its evidence, and one gate decision on the record.
Evidence-based research → production
Keeping the citation trail intact from literature to specification, so what gets built can be traced back to what was actually shown — and an unsupported claim fails the build instead of shipping.
Privacy & accessibility by architecture
Software that cannot leak what it never sends, and that is designed for people mainstream products treat as too small a market — because both properties come from the same design decision.
Welfare-aware & ecological systems
Verification and sensing for environments where automated decisions affect living systems — detecting and documenting failures for a person to act on, rather than issuing a verdict the science cannot support.
Let's build responsible systems together.
We welcome collaboration, advisors, and partners across responsible AI, provenance and audit evidence, decision assurance, verification, and human-centered design.
Get in touch