Research foundations · Public evidence demo
LAILA NEXT
The work should survive a change of mind.
LAILA NEXT began with an integration question: what would happen if the strongest ideas from FRIDAY, LAILA, and Lexi became one cohesive, model-agnostic assistant?
The first step was an inventory of what those projects could contribute. FRIDAY brought experience with tools and agent execution. LAILA brought identity and permission patterns. Lexi brought lessons about interpersonal continuity. Muninn supplied an external memory boundary.
That led to an independent implementation. NEXT needed its own durable account of what was happening, what had been approved, and what had already taken effect. Experiments in the other projects needed room to continue without destabilizing it.
The central idea is that a model can be replaced while the work continues. A model proposes an action; the runtime checks it against current state and authority. If a person changes direction, an older model response must not overrule them. If execution stops halfway through, recovery needs to distinguish completed work from work that is still uncertain.
The research also asks when a more elaborate agent strategy earns its cost. Reported experiments improved some evaluation results without establishing a consistent advantage on harder challenges. The simpler production path remained in place.
LAILA NEXT is the strategic direction of this work. Its long-term aim is an assistant that maintains continuity across changing models and devices while remaining under its owner’s control. The current project develops and tests the foundations of that ambition.
State outlives a model turn
The runtime records durable activity, approvals, and effects separately from model cognition. Exact approvals bind an action to its current context; a new human instruction can invalidate older cognition through generation fences.
Recovery consults recorded execution state before proceeding. An uncertain external effect needs reconciliation rather than an assumption that retrying is harmless. This is not a universal exactly-once guarantee for external actions.
The Owner Node retains canonical authority. Discovery alone does not grant trust: enrollment requires owner confirmation, delegation is bounded, and revocation and quarantine remove authority. Portable bundles and delegation have fixture evidence; authenticated production transport and physical-device certification remain unfinished.
The frozen Full Shadow demonstration and the continuing Phase 8 certification are separate experiments. Full Shadow showed gains on its sealed tasks at higher token cost without establishing generalization. The Phase 8 report found no material cost-normalized advantage on its final challenge and kept the production DIRECT path. Neither result establishes universal model certification.
The public instrument presents a demonstration and sanitized telemetry. It does not expose the private execution environment.