EXP-QWEN-NS-LIC-EXTENDED

Q4 (exp-cross-model-concept-novelty-portability, 2026-05-18) found PARTIAL-HOLDS-ON-HAIKU per pre-reg / HOLDS-CROSS-MODEL-WITH-CAVEAT substantive. Haiku 87.5%/0% concept-novel CO/NS replicates flash-lite Q2 +4x amplified. Qwen CO_LIC 85.7% replicates the inverted-persona direction. But qwen NS_LIC_NOC produced n_acc=1 across 10 runs at n_rounds=45, collapsing the rate metric to Wilson [20.7, 100] — uninformative. Pre-reg kill criterion technically FAILS on qwen due to DENOMINATOR COLLAPSE, not direction reversal. This experiment settles the qwen NS_LIC interpretation by extending exposure 3.3x (vs Q4: 1 cell x 15 runs x 100 rounds = 1500 nominal rounds vs Q4 qwen-NS 10x45=450).

This is paper7 V3 attack-6 (sequence: Q1 ceiling, Q2 factorial, Q3 closure-replication, Q4 cross-model, this Q4-followup, plus retro audits accumulating).

DESIGN: 1 cell (NS_LIC_NOC — persona/license/counter prompts verbatim from Q4/Q2) x N=15 runs x n_rounds=100 (5 gens x 20 rounds_per_gen, matches Q2 structure). Same engine rule_multigen, same RULE_SPACE_SEED implicit, same propose_template, same persona prompt, same temperature 0.85. Only differences vs Q4 qwen-NS_LIC arm: N=15 (vs 10) + n_rounds=100 (vs 45) — a 3.3x exposure multiplier targeting expected_n_acc ≥5 (denominator-power-precheck per candidate lab.md rule 2026-05-18).

PRE-REGISTERED VERDICTS (LOCKED 2026-05-18 BEFORE RUN):

• HOLDS-CROSS-MODEL-FULL: qwen NS_LIC concept-novel rate ≤7% AND n_acc ≥5. Direction-confirmed cross-arch. Paper7 V3 thesis cross-arch full-holds. Combined with haiku CO 87.5% / NS 0%, the inverted-persona pattern is established as substrate-architectural across three model classes (gemini, anthropic-RLHF, open-weight-qwen).
• FLASH-LITE-AND-HAIKU-ONLY: qwen NS_LIC concept-novel rate >7% AND n_acc ≥5. qwen breaks pattern; paper7 V3 thesis restricts to {gemini-2.5-flash-lite, claude-haiku-4.5} = RLHF-aligned model class.
• DENOMINATOR-STILL-COLLAPSED: qwen NS_LIC produces n_acc <5 even with 3.3x extended exposure (1500 nominal rounds). Substrate-bounded interpretation: qwen-7b under no-licensing-but-novelty-seeking-persona rejects ≥95% of proposals — a qwen-specific structural-restrictiveness finding, not direction-evidence. Paper7 V3 thesis remains cross-model on a-rate-discriminative basis (haiku + flash-lite) plus a separate qwen finding about restrictiveness.

Computation: per-cell concept-novel rate via scripts/buzzword-recombination-audit.py applied to accepted proposals. Wilson 95% CI on pooled concept-novel count / total accepts. Frame-control comparator: Q4 haiku CO_LIC + NS_LIC pair (haiku NS=29 accepts/10runs at n_rounds=45, ~0.064 accept/round density). Frame-control prediction: if qwen NS rejection density is the same as haiku NS (0.064 accept/round), at 1500 rounds we expect ~96 accepts. If qwen NS is 10x more restrictive than haiku NS (which Q4 already suggested), we expect ~9-10 accepts — still ≥5.

DENOMINATOR-POWER-PRECHECK (per lab.md candidate 2026-05-18 'denominator-power-precheck-on-rate-metrics'): Q4 qwen NS arm: 10 runs x 45 rounds = 450 rounds → 1 accept (~0.22% accept-density). This exp: 15 runs x 100 rounds = 1500 rounds. At Q4 accept-density 0.22%, expected_n_acc = 1500 x 0.0022 = ~3.3 accepts (still under 5-threshold!). At 3x improvement (qwen-NS slightly less restrictive with more rounds-per-gen), expected_n_acc = ~10. Honest pre-registration: 50% probability of DENOMINATOR-STILL-COLLAPSED outcome. We are committed to reporting that as honest verdict if it manifests; it is the FAITHFUL outcome under qwen-NS-extreme-restrictiveness.

COUPLING NOTE: this experiment INTRODUCES new agent-prompt to baseline rule_multigen engine (custom propose_prompt_template + persona condition prompts copied verbatim from Q4 and Q2). Per lab.md INJECT rule, verdict.md MUST contain frame-control ablation section. Frame-control comparator here: Q4 haiku NS_LIC cell on the SAME engine + propose-template + persona-prompt + run-matrix-structure — ONLY model varied. Pre-stratification-ceiling-axis-floor discipline 4-element check: (1) max-licensing single-cell ceiling: Q4 qwen CO_LIC_NOC = 85.7% (Wilson [65.4, 95.0], n_acc=21/10runs) — SATISFIED. (2) per-condition expected_proposal_count_by_axis ≥5: this is the WHOLE POINT of this rerun. (3) prompt-content-frozen baseline cell: deferred-justification (single-arm follow-up exp, cost-budget constrained — baseline Q4-flash-lite-Q2 is the cross-experiment comparator). (4) per-cell INJECT ratios <1.2x ⇒ mechanism-receptive: deferred to cross-exp pooling (this single-arm cannot self-report; relies on Q4 comparator).