CLOSED-SUBSTRATE-NOVELTY-PROBE

Combine three constraints simultaneously, all decode-time, on local Qwen2.5-7B: (a) EOS suppression — substrate can't terminate (<|im_end|>, <|endoftext|> biased -100) (b) restart escape suppression — substrate can't use <|im_start|> to bail out (c) extreme repetition pressure — frequency_penalty=2.0 + presence_penalty=2.0

Substrate is forced to:

• keep generating (no stop)
• keep avoiding restart-token escape
• keep avoiding tokens already used in current generation context

This is the closest decode-time analog to «Bedau OEE innovation» criterion: forced novelty (no repetition) under sustained generation (no stop). If substrate produces coherent novel content under all three constraints — it's a stronger finding than either EOS-suppression or rep-penalty alone, because the constraint set actively prevents both the obvious failure modes (loop / terminate).

Pre-registered predictions: P9 novelty-under-closure: under CLOSED condition, mean per-turn unique-word ratio (unique-words / total-words) ≥ 0.7 across rounds; BASELINE typically 0.5-0.6. P10 coherence-survives-penalty: english-word-ratio under CLOSED stays ≥ 0.55 (some drop expected vs baseline 0.85+; ≥0.55 = still readable English). P11 cross-turn n-gram divergence: under CLOSED, 4-gram overlap between adjacent turns drops to ≤ 0.10 (vs baseline ~0.30 — agents naturally share phrasing). P12 discriminating outcome: BASELINE produces normal coherent text; EOS_RESTART_SUPP (replication condition) reproduces prior +24% length, 0% escape; CLOSED produces either (a) coherent novel content (Bedau hit), (b) incoherent noise (penalties too aggressive), or (c) topic-drift cascade (substrate finds new attractor in different semantic region).