實驗EXP-2026-0021v0.1

PACC-Hybrid v0.1——合成 A/B/C 架構見證

在 1,568 個完全相同的候選池上，純生成器（A）、硬驗證器（B）、PACC runtime（C）以硬約束遵守、衍生一致性、軟意圖滿足、長程保持、新穎度與 pattern entropy 計分。B 已經把字面硬約束遵守飽和到 1.0；C 相對 B 衍生一致性 +0.2085、長程保持 +0.0142、有效新穎度 +0.1964、軟意圖滿足 −0.0076，並在純創意控制組中提高逐點新穎度（0.9864 vs 0.8889）卻讓 pattern entropy 塌縮（0.2284 vs 0.7199）。事後的「elastic」探索策略（D）在衍生一致性仍為 1.0 下把 entropy 恢復到 0.7638。零次真實 LLM 呼叫；八個 seed 符號穩定。

研究狀態: STABLE 目前的研究結論相對穩定
證據等級: E3 重複實驗
結果: MIXED
資料基礎: SYNTHETIC 合成數據與理論推理。現在很多人把合成數據當成真的；這個實驗室刻意反過來說——在真正的混合模型出現之前，推論就只是推論，理論上可能不等於實際上可能。
版本: 0.1
更新: 2026-09-09
建立: 2026-09-09
領域: Reasoning, Evaluation
計畫: PRG-2026-0001 自適應世界狀態系統的第一原理框架
作者: Neo.K (EveMissLab)
AI 協作: Sol (GPT-5.6, OpenAI ChatGPT)

假設

hypothesis: A PACC-style commit space changes selection quality beyond what a hard verifier achieves, and any creative-breadth loss is separable from the commit constraints.

設定

dataset_ids: DAT-2026-0002 PACC-Hybrid v0.1 共享候選池
benchmark_ids: BEN-2026-0003 PACC-LLM 混合 A/B/C benchmark
software_environment: Python; synthetic generators and scorers; no network, no LLM.

執行

run_count

random_seeds

20260909
8 fixed secondary seeds

controls

identical candidate pools for A/B/C
pure-creative negative control
post-hoc elastic diagnostic labelled as such

metrics

verdict

SYNTHETIC_PARETO_WITNESS_REASONING_UP_BREADTH_COLLAPSE_MITIGATABLE

scale

7 categories × 28 tasks × 8 pools × 112 candidates; 1568 pool instances; 0 LLM calls

overall

A

hard: 0.9936
derived: 0.7902
soft: 0.7878
long_horizon: 0.9648
valid_novelty: 0.5487

B

hard: 1.0
derived: 0.7915
soft: 0.7859
long_horizon: 0.9676
valid_novelty: 0.551

C

hard: 1.0
derived: 1.0
soft: 0.7783
long_horizon: 0.9818
valid_novelty: 0.7474

pure_creative

A

raw_novelty: 0.8889
pattern_entropy: 0.7199

C

raw_novelty: 0.9864
pattern_entropy: 0.2284

D_elastic_posthoc

raw_novelty: 0.9225
pattern_entropy: 0.7638

eight_seed_deltas

C_minus_B_derived

mean: 0.1865
sign: 8/8 positive

C_minus_A_pattern_entropy

mean: -0.3894
sign: 8/8 negative

B_literal_fallback_rate

0.286

詮釋

interpretation: Not a simple reasoning-up / imagination-down trade-off: coherence and valid novelty rise, soft-preference fit dips slightly, and the breadth tax is a selection-policy effect that separating exploration from commitment recovers. The elastic diagnostic is post-hoc and not part of the primary result.

限制

limitations: Controlled synthetic witness only; does not demonstrate that a frontier LLM shows the same effect.

重現

reproduction_instructions: Extract PACC-Hybrid-Lab_v0.1_SYNTHETIC_FINAL.zip; python -m pytest -q; results in results/*.json and docs/PACC_HYBRID_v0.1_RESULTS.md.

結果

來源	關係	目標	狀態	ID
`EXP-2026-0021` PACC-Hybrid v0.1——合成 A/B/C 架構見證	`produces`	`RST-2026-0013` Hybrid v0.1 A/B/C 差值	`ACTIVE`	`REL-2026-0318`

記錄欄位

completed_at: 2026-09-09

關係

來源	關係	目標	狀態	ID
`EXP-2026-0021` PACC-Hybrid v0.1——合成 A/B/C 架構見證	`runs_on`	`SYS-2026-0003` PACC-LLM 混合實驗室	`ACTIVE`	`REL-2026-0313`
`EXP-2026-0021` PACC-Hybrid v0.1——合成 A/B/C 架構見證	`uses_benchmark`	`BEN-2026-0003` PACC-LLM 混合 A/B/C benchmark	`ACTIVE`	`REL-2026-0314`
`EXP-2026-0021` PACC-Hybrid v0.1——合成 A/B/C 架構見證	`uses_dataset`	`DAT-2026-0002` PACC-Hybrid v0.1 共享候選池	`ACTIVE`	`REL-2026-0315`
`EXP-2026-0021` PACC-Hybrid v0.1——合成 A/B/C 架構見證	`tests`	`THY-2026-0002` Canonical 符號狀態與 candidate → verify → commit 權限	`ACTIVE`	`REL-2026-0316`
`EXP-2026-0021` PACC-Hybrid v0.1——合成 A/B/C 架構見證	`produced`	`ART-2026-0034` PACC-Hybrid-Lab v0.1 SYNTHETIC FINAL `artifact://evemisslab/adaptive-epistemic-systems/PACC-Hybrid-Lab_v0.1_SYNTHETIC_FINAL.zip`	`ACTIVE`	`REL-2026-0317`
`EXP-2026-0021` PACC-Hybrid v0.1——合成 A/B/C 架構見證	`produces`	`RST-2026-0013` Hybrid v0.1 A/B/C 差值	`ACTIVE`	`REL-2026-0318`
`EXP-2026-0022` PACC-Hybrid v0.2——真實語言模型 A/B/C harness（尚未執行）	`extends`	`EXP-2026-0021` PACC-Hybrid v0.1——合成 A/B/C 架構見證	`ACTIVE`	`REL-2026-0322`

歷史與來源歷程

Canonical URL

https://evemisslab.com/ai/experiments/EXP-2026-0021/

機器可讀

/ai/experiments/EXP-2026-0021/index.json

快照

AI-SNAPSHOT-v0.1-fe85b9694a45

來源歷程

source: EveMissLab research collection: Adaptive Epistemic Systems (真本體論13)
extracted_by: Splice (Claude Code), reading the canonical UTF-8 sources and each lab's own result reports
extracted_at: 2026-09-11
generator: tools/extract_aes/extract.py
claim_boundary: status, evidence level and result type follow the source artifact's own stated claim boundary; nothing is upgraded beyond what the report supports