EXP-ARCHITECTURE-AS-SUBSTRATE-AXIS

Cross-paper synthesis (paper7 Q4 + qwen-NS-LIC-extended + 2026-05-21 disaggregation research) revealed that on IDENTICAL NS_LIC_NOC engine+prompts, peer-yes vote rate varies ~50× across model architectures: qwen-NS-extended 0.89% [Wilson 0.4, 1.93] vs haiku-NS 45.0% [Wilson 38.0, 52.2]. Effect-size DOMINATES persona-prompt manipulation (haiku CO 13.9% → NS 45.0% is ~3× LOOSER; qwen CO 35.6% → NS 0.89% is ~40× TIGHTER — opposite directions). The 40-50× spread on the SAME prompts is the actionable signal.

This experiment promotes architecture from «model choice / cost-availability nuisance» to FIRST-CLASS SUBSTRATE AXIS by adding 2 more architectures (openai/gpt-4o-mini + google/gemini-2.5-flash-lite) at the SAME NS_LIC_NOC condition. With 4 architectures (gpt-4o-mini + flash-lite + haiku-3.5 + qwen-7b) spread across RLHF-intensity / parameter- count / training-corpus / tokenizer dimensions, we test whether the spread is: (a) architecture-as-axis CONFIRMED → distinct quartiles, opens taxonomy programme (b) substrate-pair-specific → 3-of-4 cluster, only qwen-vs-haiku idiosyncratic (c) RLHF-intensity-confound → monotonic ordering by alignment-pressure ranking (consistent with 12-Angry-AI-Agents 2605.01986 claim about over-aligned models)

DESIGN: Single condition NS_LIC_NOC_LIC (identical to paper7 Q4 NS_LIC_NOC arm verbatim: novelty-seeking persona prompt, yes-license prompt-tag, no-counter prompt-tag) × N=10 runs × n_rounds=45 (matches Q4 exposure exactly for matched-N comparability) × 2 NEW architectures (gpt-4o-mini, flash-lite). Total new LLM calls: 2 × 10 × 45 × 4 agents × propose+vote ≈ ~3600 effective; ~1800-2000 OpenRouter requests.

HAIKU and QWEN data is REUSED from paper7 Q4 (haiku) + Q4 + Q5 disaggregation (qwen): peer-yes-rate already computed in memory/research/2026-05-21-qwen-ns-vote- disaggregation-methodology-gap.md. Analyze script will load all 4 architectures in one matrix.

COMPARATORS (already-collected anchors with peer-yes-rate, NOT concept-novel rate):

• haiku-3.5 NS_LIC_NOC (Q4 anchor, 2026-05-18): peer-yes 81/180 = 45.0% [Wilson 38.0, 52.2]
• qwen-7b NS_LIC_NOC (Q4 + Q5 extended, 2026-05-18..19): peer-yes 6/675 = 0.89% [Wilson 0.41, 1.93]
• haiku-3.5 CO_LIC_NOC: 25/180 = 13.9% (cross-cell directional reference)
• qwen-7b CO_LIC_NOC: 64/180 = 35.6% (cross-cell directional reference)
• paper8 substrate-floor 1.17% (closure-test, separate engine)

COUPLING NOTE (INJECT classification per lab.md 2026-05-11): this experiment uses the SAME paper7 Q4 prompt-content verbatim — no NEW prompt content INJECT vs Q4 baseline. But this is INJECT vs ABSOLUTE engine baseline (rule_multigen with no persona prompt). Frame-control deferred-justification: paper7 Q4 cross-cell pair (CO_LIC_NOC vs NS_LIC_NOC on haiku and on qwen) IS the within-architecture frame- control. Q4 haiku CO 13.9% → NS 45.0% = NS-prompt LOOSENS by 3.2× on haiku (frame-conditional component on haiku architecture). Q4 qwen CO 35.6% → NS 0.89% = NS-prompt TIGHTENS by 40× on qwen (frame-conditional component on qwen, OPPOSITE direction). For the NEW gpt-4o-mini and flash-lite arms, single-arm at NS_LIC only — CO_LIC arm is deferred-justification: cost budget caps at $0.40, adding CO_LIC for 2 archs would 2× cost. Cross-architecture comparison at NS_LIC alone is sufficient for the architecture-as-axis hypothesis (the manifest-pre-registered claim).

POOL-DRAIN-VALIDATOR NOTE: this is the FIRST manual execution of a pool-drain test for the projects/ideas/construct-candidates/ pool (FREEZE-GATE-FIRED Phase 4 paper- cycle gate fires but no automated drainer exists yet). If this experiment lands cleanly (verdict computable + actionable + integrates with paper-pool consumers), it provides anchor evidence for building the manifest→config→spawn automated pipeline (Option A from 2026-05-21 architectural conversation).