EOS-SUPPRESSION-MULTI-AGENT

Single-agent EOS-suppression probe (exp-eos-suppression-substrate-probe, verdict PARTIAL 2026-05-24, ρ context not applicable) showed that when EOS + restart escape (<|im_start|>) are both biased -100, Qwen2.5-7B sustains coherent generation up to max_tokens cap. The «complete-and-stop» modal LLM behavior is NOT a fundamental substrate property — it's a behavioral default.

This experiment extends the probe to multi-agent shared text space: 3 agents take turns adding to shared memory. Under EOS+RESTART suppression, no agent can naturally stop. Question: do they construct longer, jointly coherent threads together when none can terminate, or does individual decoherence accumulate into collective drift?

This is the autopoietic-substrate question stripped to minimum form: if you remove the «answer-and-die» drive, does substrate produce self-sustaining text dynamics?

Pre-registered predictions (P5-P8 building on prior P1-P4):

• P5 multi-agent length scales with bias: under EOS_RESTART_SUPP, mean per-turn output_tokens hits max_tokens cap in ≥80% of turns vs baseline natural-stop in ≥80% of turns.
• P6 coherence preserved (multi-agent extension): English-word-ratio delta between baseline and EOS_RESTART_SUPP stays under 30%.
• P7 cross-agent referencing: shared memory becomes a substrate the agents can build on. Operationalized: rate of agents quoting/referencing each other's prior contributions (≥2 alpha words overlap with prior turn from different agent) is HIGHER under EOS_RESTART_SUPP than baseline by ≥10 percentage points. (Hypothesis: extra generation pressure forces agents to incorporate context rather than emit self-contained answers.)
• P8 thematic drift bounded: mean topic-cosine-similarity (agent text vs seed topic) stays above 0.5 in all conditions — agents don't drift to unrelated subject matter even when forced to keep generating.