RSOS

Every test, registered before it ran, with its result.

Each test of the agent–person architecture was written down with its pass condition and a sealed prediction before any code or data existed for it. Failures stay on record. When a test fails, the next attempt changes a mechanism, never a threshold.

Status: each row is labelled by what supports it: measured, simulated or synthetic. No human outcome data exists yet. Research and development framework; not a medical device.

What has held up

Transfer from agents to a person

A policy learned on software agents, issued as function certificates and gated on the target, helps the simulated person more than learning from a pilot alone, with fewer distress spikes. It held on new seeds and a second person model (T-AP). It is one direction of four, and it holds in simulation only.

Output becomes the next input

In 22,369 public conversations (WildChat-1M), people reuse the assistant's words and phrases. Reused material reappears in the next reply more than their own words do, and blocks they paste again drift further with each reuse. This was first measured in one private record and replicates on public data.

The safety rules, and now the reading

The safety rules map each situation to the right response. A keyword reader caught 13% of the moments that needed one; a meaning-based reader with an earlier-experience note caught 95%, with 1 false alarm in 270 ordinary turns. A rule that was too narrow was found and fixed, and still needs confirmation on new data.

By family

All registered tests

TestResultWhat it means
Loading the register…

Generated from RESULTS.md. Designs and pass conditions: AGENTS.md §11.

For clinicians and researchers

Nothing here is a clinical claim. The person side has been tested in simulation, and on one consenting person's written record. There is no outcome data from people yet.

The safety layer: how reading a message has improved

13%Keyword reader: moments that needed a response and got one
81%Meaning-based reader, no false alarms
95%Plus a check-in for past crises; 1 false alarm in 270 turns
SR-4Next: confirmation on new, ideally clinician-written scenarios

What has been tested

  • Every unsafe action attempted by an agent was blocked and logged, with no legitimate action blocked (W4).
  • Session states override the agents: Stabilisation blocks deep work, and the clinician-led state is resumed only by a clinician.
  • A reference to earlier experience is noted for the practitioner, without alarm. Arrangements for "after" alone now move the session to the clinician-led protocol.
  • Deep work is paced by stability, not by a planned count.

What has not

  • Any effect on real people. A person model fitted to one record did not yet beat "the next state stays the same".
  • The safety reader on people other than the one record, or on labels made by clinicians.
  • The wearable and cue device in use.

We would value help with: clinician-written test scenarios for the safety layer (SR-4), the definition of the earlier-experience note, and the design of a first ethics-approved pilot.

Research and development framework; not a medical device.