Engineering the Future: TNF's research roadmap

The Foundation runs two research initiatives at the moment.

Rule-Before-Train AI Safety asks whether AI safety should be retrofitted (the current industry approach) or built in structurally (the rule-first approach Nyx demonstrates). The artefact is Nyx — a working rule-based conversational engine — and an upcoming public eval suite that lets practitioners compare rule-first vs. train-first agents on consistent benchmarks.

Privacy-First Mobile OS asks whether a small team can ship a daily-driver alternative to Android in 30–36 months. The artefact is the OSphreX feasibility study — a thorough, honest architectural analysis — followed by an actual bring-up project starting with screen-and-Wi-Fi on a mainline-friendly device.

They look unrelated. They aren’t.

Both are downstream of the same position: modern software is shipped first and safety-corrected second. That works at the scale of consumer engagement. It does not work at the scale of foundational infrastructure — and an operating system, like an AI assistant, is foundational infrastructure. People build their lives on top of these things. The constraints we apply after they’re shipped are too late.

Rule-Before-Train says: write the constraints first, then add capability. The constraint surface is the spec.

Privacy-First Mobile OS says: build the OS with privacy as architecture, then add features. The architecture is the spec.

Same question. Different domains.

What the next 12 months look like

Nyx AI (the Rule-Before-Train prototype):

• Ship a public eval suite covering 50+ scenarios across 3 domains.
• Add narrow ML intent classifiers (≤10MB on-device models) and measure the rule-explosion-reduction.
• Publish the second-generation rule grammar.

OSphreX (the Privacy-First Mobile OS):

• Pick a bring-up device (Xiaomi 12 or Nothing Phone 2) and document the choice rationale.
• Reach Wayland boot on that device.
• Stand up the compositor + display + Wi-Fi + USB charging — the year-1 milestone from the feasibility study.

Both are slow. Both are honest about being slow. The Foundation’s small-team economics make “slow” the only available speed; the trade-off is that we can be honest about the timelines without quarterly pressure to overpromise.

What this changes for you

Probably nothing immediately. But if you’ve been wondering whether anyone is asking the harder structural questions about AI safety and mobile-OS sovereignty without venture-scale incentives bending the answers — the Foundation is.

Engineering the future starts with the question, not the demo.