DIFFERENTIATION-WITHOUT-INJECTION

L7 verdict (exp-axis-diversify-false-ablation-arm, commit c678c76d8) showed that agent-role specialization in mnsky's rule_multigen archive was entirely engine- injected via axis_per_agent dict + alphabetical RULE_SPACE key-order pathway. This invalidated all prior multi-agent differentiation findings in mnsky as «substrate-emergent» — they were all prompt/engine artifacts.

This experiment is the cleanest possible L2-emergence test: 5 agents with IDENTICAL persona text, IDENTICAL prompts, generic name letters (A/B/C/D/E) to minimize name-induced prior. Two conditions:

• BASELINE: standard generation, no constraints
• CLOSED: full constraint stack (EOS+restart suppression + max freq/presence penalty)

Question: do agents differentiate purely from interaction history when prompt- asymmetry is removed? If yes — first true L2 emergence in mnsky after L7 invalidation. If no — L2 differentiation requires asymmetry; substrate alone is insufficient (clean null result, closes door).

Pre-registered predictions:

• P13 pairwise output cosine similarity drops <0.7 by round 6+ in CLOSED (vs near-baseline ~0.85 random-pair expected when identical agents converge).
• P14 per-agent vocab Jaccard within-agent across rounds ≥0.4 (role-stability), across-agents <0.2 (different role-vocab niches).
• P15 differentiation rate CLOSED > BASELINE — penalty pressure should amplify forced novelty into per-agent niches.

Discriminating outcome (mutually exclusive):

• (a) DIFFERENTIATION-EMERGES: P13+P14+P15 all met. First clean L2 emergence without injection. Strong anchor for substrate-can-specialize-without-prompt candidate rule.
• (b) CONVERGENCE-NULL: all agents output similar texts (cosine >0.85, no vocab clusters). L2 emergence requires asymmetry; substrate alone insufficient.
• (c) DRIFT-WITHOUT-SPECIALIZATION: agents diverge but no stable per-agent identity — random walk in semantic space, not differentiation.