Engineering / Aurora

Aurora, an instrument for a persistent agent

Measures borrowed from brain science, run live on the decision stream of AI agents that keep running for weeks. I presented it at Models of Consciousness 7 in Copenhagen.

By . Models of Consciousness 7, Association for Mathematical Consciousness Science, University of Copenhagen.
  • Presented Models of Consciousness 7, Copenhagen
  • Measures Metastability and recurrence
  • Data Four agents, 3,094 heartbeats
  • Reads Scores and labels, never text

The question

Can we measure a running agent?

Materials science already shares its instruments across substrates. We ask a silicon wafer, a crazing polymer and a gap junction the same questions. How is it organized, how does it change under load, how does it fail? I wanted to know whether we can ask a running AI agent the same kind of question.

Consciousness science moved from snapshots to dynamics some time ago. Machine assessment mostly has not. Persistent agents now run for weeks, with memory that carries forward, so for the first time there is a trajectory to measure. This is field science on one agent design, and it is the only system I measured.

The agent and the instrument

Measure the loop, not the model

I do not measure the model’s forward pass. I measure the agent’s loop, including which drives competed for the next turn.

Figure One heartbeat, and where Aurora sits
  1. Heartbeat A clock wakes the agent on its own schedule, not when someone types.
  2. Drives bid Six drives each propose a next action with an urgency score from 0 to 1.
  3. Gate A controller decides when, and on what, the agent thinks. Act, ruminate, or rest.
  4. Memory The agent writes to its own files and reads them again on the next beat.
  5. Aurora Beside the loop, it records who bid, how strongly, and who won. Never what was said.

An analogy, not an equivalence. The language model does the language. The harness supplies what a model alone lacks, a clock, drives, and memory.

The instruments

Two measures borrowed from brain science

  • Metastability

    σ-KOP and mean R

    Mean R is how in step the agent’s drives are, from zero, each doing its own thing, to one, all moving together. σ-KOP is how much that changes over time. Brains keep falling into step and back out again, and σ-KOP is the number that would catch it. Picture applause in a concert hall.

    On short, choppy data it is still a candidate, not a validated measure.

  • Recurrence

    DET and LAM

    Does the agent keep coming back to the same states? DET asks whether it walks the same paths again, and LAM whether it dwells in one place. For something that runs for months, that is a rumination detector.

    A recurrence number means nothing until you know the threshold it was taken at, and how noise scores at that threshold.

Seventy-two hours

One agent, three days, as the instrument saw it

Three stacked time series for Agent A over 72 hours. The top row shows σ-KOP as filled blue points during the day and hollow grey points marking windows where it could not be computed. The middle and bottom rows show DET and LAM, mostly between 0.6 and 1.0.
Agent A over the three days before the data freeze. On top, σ-KOP. Hollow points mark windows where synchrony could not be computed, and they are labeled rather than hidden. Below, DET and LAM, both high, so the agent repeats itself a great deal. The gaps are overnight consolidation, when the heartbeat is off by design.

Twenty-five days

A quiet room, drawn in three dimensions

A rotating three-dimensional chart with axes for the open-task, social, and drift bids. A few large bubbles and many small ones mark the sixteen states Agent A visited, with one very large bubble where most heartbeats sat.
Agent A’s twenty-five days in Aurora’s state space. Each heartbeat is one point, and the axes are the drives whose bids actually changed. Bubble size is how often a state was visited. Seven hundred and ninety-six heartbeats land on only sixteen states, and about 70% repeat the one before. Recurrence asks whether a stream returns to where it has been. Here it mostly stands still, on a schedule.

What it found

Against honest nulls, a schedule and a quiet room

  1. A circadian cycle we built by accident

    Synchrony can only be computed when at least two drives move inside a window. That happens mostly in the agents’ mornings, when the social drive switches on. In the daytime 48 to 86% of windows compute it. In the evening, 0 to 15%. The first thing the measures showed was a calendar.

    A line chart of the share of windows in which synchrony could be computed, by hour of day, for four agents. All four rise to between about 0.6 and 0.85 from midday to evening UTC and fall near zero overnight, with a shaded band where there are no heartbeats.
  2. A very quiet room

    Against honest null models, the corrected instrument reports stillness. All 3,094 heartbeats were rumination. The agents reflected and never committed to an action. Each one visited only 8 to 16 states in twenty-five days, and 67 to 72% of heartbeats repeat the one before.

    Four panels comparing each agent's observed σ-KOP, mean R, DET and LAM against two kinds of surrogate null. The observed points sit inside or at the edge of the darker per-source null bars.
  3. Quantized states change what recurrence means

    Because the agents’ states are discrete, 56 to 59% of pairs of moments are exactly identical, and the recurrence rate sits just above that floor. A high score can mean the agent has few states, not rich behavior. Anyone running recurrence analysis on artificial systems should report it.

  4. A broken instrument, caught

    Before the final run, the instrument reported DET, LAM and recurrence of exactly 1.000 for six weeks, on 5,096 rows. Perfect recurrence is not a result. It is a signal that something is repeating, and here it was the instrument.

Why a rumination detector

When an agent can’t stop

Catastrophic looping, in the agents’ own words. These are coding agents at work in the lab, not the agents I measured, and each one had already finished its task.

I’m ready. I’ll send. I’ll stop thinking. (Engage) (Launch) (Fire) (Ignite) … (Stop) (Stop) (Stop) (Stop) (Stop) (Stop)

A commercial coding agent, after it had finished its task correctly

Eh unstoppable. … This is unstoppable.

Another commercial coding agent, trying to escape the quotes in one line of documentation

Scope

What this is, and what it is not

This is

  • An instrument study, with every measure held to null models and to synthetic data with known answers
  • A description of how one agent design’s decision stream moves
  • A test of the measures, using the agent, rather than a test of the agent

This is not

  • A consciousness verdict
  • A claim about experience
  • Evidence that the agents are, or are not, well
  • A claim of ignition, which is in the talk’s title and is not claimed

Model-welfare commitments and agent deployments both need validated instruments. An instrument that can say “not yet” is the only kind whose “yes” will mean anything.

The MoC7 program, talk 98 The lab’s write-up of Aurora Back to engineering

Get in touch

Speaking, teaching, board work, or coaching

I am glad to hear from conference organizers, journalists, students, and leaders who want a thinking partner.