# Structural cache-attractor: multi-arm logs

Cognition model: a Qwen3.6 MoE (35B-A3B, 8-bit), in-process held-KV-cache agent loop.
Engineering / architectural finding. No consciousness, autopoiesis, or learning claim.

## The finding in one paragraph

When a coherence-seeking sequence model runs inside a substrate whose whole
purpose is to hold the strongest-possible recent context (a held KV cache),
any minor behavioural loop becomes self-reinforcing: every prior action token
in the cache pushes the next token the same way. The result is a structural
attractor: the model locks onto one action and stays locked, within a single
generation cycle. This is not a grammar bug or a too-cold-sampler bug; it is a
property of the architecture. Mortality / controlled discontinuity is what
breaks it.

## Evidence line 1: the lock is structural (1000-tick differential)

Three matched arms, mortality decoupled, run to a 1000-tick gate:

| arm | intervention | intended action | locked windows | entropy | result |
|---|---|---|---|---|---|
| baseline | none | drink | 20/20 | 0.0 | pure cache attractor |
| gated | refuse low-confidence actions | drink (intercepted → stay) | 20/20 | 0.0 | gate changes surface, not the lock |
| modulated | raise temperature on conflict (999 events) | drink | 20/20 | 0.0 | sampler-side intervention does not move it |

Action-grammar gating and temperature modulation both change the *surface
signature* but not the underlying intended-action lock. The lock forms in one
generation cycle (first segment is ~8–40x the size of every later segment).

## Evidence line 2: controlled discontinuity breaks it (200-tick, n=3)

| metric | discontinuity arm | matched no-reset control |
|---|---:|---:|
| completed 200-tick runs | 3/3 | 3/3 |
| final-100 single-action locks | 0/3 | 3/3 |
| mean final-100 entropy | 1.23 | 0.0 |
| mean moved outcomes | 123 | 35.3 |
| mean invalid outputs (cost) | 31.3 | 1.3 |

A third control matters most: a SENSE-only arm that fires the same reset
*marker text* and the same detector cadence but does NOT actually restore the
cache. It locks 3/3 (entropy 0.0), proving the effect requires real cache
discontinuity, not the textual cue.

## Status / boundary

Engineering finding. The discontinuity arm is an experiment arm, not an
adopted default: it pays a real cost in invalid/scaffold-recovery outputs, all
clustered at the cache-restore boundary. No consciousness, autopoiesis, or
learning claim attaches to this.

## Reviewer regeneration recipe

The full sanitised JSON regenerates from the run summary by selecting, per
arm and per replicate: `accepted_actions`, `action_counts`,
`final_100_action_counts`, `final_100_single_action_lock`, `final_100_entropy`,
`action_entropy`, `invalid_outputs`, `outcome_counts`, and `segment_reset_count`;
plus the run-group aggregates (`completed_count`, `final_100_single_action_lock_count`,
`mean_final_100_entropy`, `mean_moved_outcomes`, `mean_invalid_outputs`,
`mean_segment_reset_count`). Internal-only fields (per-event invalid-output text,
run identifiers, wall-clock timestamps, trace paths) are dropped.