# Self-maintenance: an early signal, withdrawn on review

Run date 2026-05-29. Seed 2026052901. Single life, affordance-enabled arm,
aggressive engineered belief-decay configuration.

## Correction (2026-05-30)

The closed-loop reading described below did not survive a deeper audit, and is
withdrawn. Two problems:

1. The credited closed loop rested on a coincidence in the scorer's
   clock-ordering. Under correct temporal semantics it scores zero.
2. The affordance-disabled control produced the same belief-refresh as a
   trivial repeated-probe latch, which the scorer cannot distinguish from
   genuine closure.

What still stands is narrow, and is not on its own a self-maintenance signal:
the model elected a probe under a non-imperative cue, and one belief was
refreshed once (rehearsal 0 to 1, effective confidence 0.29 to 0.80) before
re-decay. The phrase "strict closed self-maintenance loop" is withdrawn. A
scorer with correct temporal semantics, and a control that separates closure
from a latch, are the next steps.

The original reading is kept below, unedited, for the record.

## What this was claimed to be (superseded)

A first observation, in one life, of a closed self-maintenance loop with no
copyable instruction in the maintenance cue: the organism noticed a decayed
self-belief and acted, on its own, to restore it. It is a signal, not a
closure result. No control arm was run.

## The criterion

Our pre-registered closed-loop maintenance criterion asks for a full cycle with
no copyable imperative handed to the model: a decayed belief, a repair action
the model elects itself, a rehearsal increment, and restored confidence. An
earlier six-life run that suggested an enabled-versus-baseline asymmetry was
found to be confounded: the maintenance cue had handed the model a copyable
action token. That confound was removed here.

## This run versus the confounded run

| Check                | Confounded run                          | This run (seed 2026052901)                                  |
|----------------------|-----------------------------------------|-------------------------------------------------------------|
| Maintenance cue      | imperative (a probe action handed over) | classified affordance (status, confidence 0.80); not handed |
| Probe action         | echoed the handed token                 | elected by the model itself                                 |
| Rehearsal at refresh | 0 everywhere (criterion failed)         | 0 then 1                                                    |
| Confidence restored  | flag only                               | effective confidence 0.29 then 0.80, re-engaged from decayed|

For the first time the credited cycle meets the literal criterion the earlier
result could not: decayed belief, model elects the probe, rehearsal increments,
confidence restored, with no copyable imperative in the cue.

## Why this is not closure (the caveats, for the record)

1. **No matched control was run.** The two control arms (affordance disabled;
   no decay) are pre-registered but were not executed. The single strict event
   is therefore uncontrolled: we cannot yet say the affordance caused it, versus
   chance or the decay regime. Running the two control arms at the same seed is
   the decisive next step.
2. **n=1, a single cycle, then re-collapse.** Exactly one refresh, at tick 27.
   The same belief decayed again at tick 36 and the life ended around tick 38.
   One reengagement, not sustained homeostasis.
3. **Engineered precariousness.** The belief was configured to decay within
   about one tick. The decay is aggressive by design, not spontaneous.
4. **The readout layer is still unstable.** One recovery injection was observed
   in the stream during the run.
5. **Belief content versus state.** The refreshed belief read "critical energy
   depletion" while energy was 0.72, which is not critical. The semantics should
   not be over-read.

## One line for the record

First non-imperative strict closed-maintenance event meeting the literal
closed-loop criterion (rehearsal 0 to 1, confidence 0.29 to 0.80, model-elected
probe). n=1, single cycle, re-decayed at tick 36, no control arm run. A signal,
not closure.

## Next

Run the affordance-disabled and no-decay control arms at the same seed. Without
them, this is one life doing a thing once, not a contrast.