FLOW: Functions as Live Overlays, Without Restart. An Architecture for Applications That People and Their Agents Change While They Run
The Live Layer Protocol (LLP) lets signed WebAssembly code replace parts of a running application without a restart and without losing state. It does not say who may change which part of an application, or what a change may do with a person's data. FLOW is an architecture on top of LLP that answers those questions. A company ships a core and declares two things about it: labelled data (public, personal, secret) and directions, the places where people's own functions may run and what they may read and use. A person, or an AI agent holding a time-limited grant from that person, signs a function into a direction. The device alone decides: it checks every signature against a root-signed keyring whose root includes a quantum-safe (SLH-DSA) key group, refuses before any code runs a function whose declared reads and capabilities could carry personal data off the device, runs functions as overlays that a company release can never be blocked by, and lets the person undo any function, restoring its old code and its old data across a change of data format. A fixed table of element meanings lets the same function appear as a web page, terminal text, speech, or a tool list for an agent. We report a JavaScript implementation wired to the unchanged LLP host (185 checks pass), a second verifier in Go that passes the same 40 frozen conformance vectors, and a seven-line "bar" run side by side against an ordinary web application: the web application fails six lines and partly passes one, FLOW passes all seven. On one 4-vCPU machine a signed release swaps in 4.2 ms (median of 15), a person's function installs in 3.0 ms and is undone in 1.1 ms. With 200 functions running, a company release still lands in 13-15 ms, but the reference host spends about 11 MB per function, which rules out cheap devices today. We close with a thesis, labelled as such: applications are turning into data plugins that bring their data and primitives while the system draws the interface on demand. We cite the platform moves that point that way, show the idea running on a calculator, and argue that the missing half of that future is the trust model FLOW provides.
Authors
- Nikhil Ajay Kadalge (ORCID: https://orcid.org/0009-0004-8859-5607)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-21
- DOI
- https://doi.org/10.5281/zenodo.24411134
- Primary Topic
- Security and Verification in Computing
- Type
- preprint