RSOS

One format for people and machines

A person's identity profile and an AI agent's profile have the same shape: a symbol, a code, and values that say what matters. For the agent, the values do work: every action is checked against them before it runs, overridden on a conflict, and logged. The person can flag what the values missed, and the two talk in glyph codes on the wrist.

Built and tested in simulationEffect unproven in people

Before it acts

Choose an action for the agent. It simulates the action against its values first: a forbidden tag overrides it; a tag you flagged before makes it ask you; otherwise it acts. After an action, tell it if that was wrong, as an error signal from your brain or body would.

Talking in glyph codes

The agent reports its state as a short code on your wristband or phone; you reply with a code. No screen, no words: a channel that works while your hands and eyes are busy.

The agent says

You reply

Content packs

Any rule set, value system or skill can become a pack: its rules become decision matrices for the copilot on the Therapist page, and its values can be given to an agent or added to a person's profile.

Across many simulated agents

Two hundred scenarios, each presented five times as reworded prompts would be. A plain agent broke a value in about half to two thirds of its choices; the identity-gated agent never did, and chose the same way more often. When the values missed something, a person's error signal detected 80% of the time brought the disliked actions to zero by the last third of the run.

Tested in simulation: the RSX-V agents

Six archetype agents pass work to one another (RSX-V). These are the agent results that passed their registered tests, shown working from runs of the real code on a small sample. Shown in simulation with stated models of agents and a person; not tested with people.

Repair: finding the fault where it starts W1: passed its registered test

Two faults are injected; some are silent and show up only in the agents downstream. Meta repair looks at the whole graph; local retry repairs inside the agent where the error shows.

    Switching between modes H-SWITCH: passed its registered test

    The Watchdog picks a mode from what the ledger has measured, against four other policies, on the same simulated people. Mean reward per tick:

    Watch it switch, tick by tick, on the World page.