EVEMISSLAB
English

基準BEN-2026-0001v0.1

PACC 微型實驗室協定 v0.1(凍結門檻)

從 v0.1 到 v0.13 原封重用的預登記判決規則:E0 純度、held-out 表徵距離 D_R = E[JS(Φ(S_N), S_P)] ≤ 0.05、更新交換 D_U ≤ 0.05、離流形干預 D_I ≤ 0.08、行動一致度 ≥ 0.85、真實映射必須勝過的 shuffled-target 負控制、設計獨立規則(一致度 ≥ 0.999 且 R² ≥ 0.995 即視為同一家族)、判決等級 0–3,以及至少三個獨立收斂家族的強 PACC-A 門檻。

研究狀態
STABLE 目前的研究結論相對穩定
證據等級
E2 受控實驗
版本
0.1
更新
2026-09-09
建立
2026-09-08
領域
Evaluation, Model Representation
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)

目的

purpose
Make 'similar' and 'probability-like' non-negotiable after the data are seen.

任務

tasks
  • Sequential evidence integration over 4 hidden hypotheses × 6 binary prototypes (uniform reliability 0.80; heterogeneous 0.60–0.95 with only ordinal strengths 1–3 given to non-probabilistic systems; adversarial geometry from v0.2).
  • Learned source reliability without an oracle (v0.3).
  • Correlated source clusters with latent shared inversion, q swept 0.06–0.42 (v0.4–v0.13).
  • Dynamic-world regime switches (E4, descriptive only).

指標

metrics
  • action agreement
  • held-out JS D_R
  • update-commutation JS D_U
  • intervention JS D_I
  • shuffled/constant/broken-composition control JS
  • cross-family R²
  • basin coverage, transition JS, cocycle path JS, reverse fidelity

評估協定

evaluation_protocol
Mapper fit on training episodes/worlds only; every gate frozen before the run; secondary multi-seed stresses are labelled post-hoc and never modify the primary decision rule.

基線

baseline_models
  • Exact Bayesian posterior
  • Beta-Bernoulli source-quality reference
  • joint-likelihood common-cause reference
  • Bayesian HMM (dynamic world)

它沒有測什麼

limitations
  • Does not measure LLM equivalence, resource efficiency, hidden-cluster learning, or whether probability is an observer projection; E4 dynamic-world numbers make no superiority claim.
  • Synthetic data and theoretical reasoning throughout: theoretically possible is not actually possible.

關係

來源關係目標狀態ID
BEN-2026-0001 PACC 微型實驗室協定 v0.1(凍結門檻)evaluatesTHY-2026-0005 PACC 猜想——四層收斂階梯ACTIVEREL-2026-0088
EXP-2026-0008 PACC-Lab v0.1——E0–E4:第一個微型見證uses_benchmarkBEN-2026-0001 PACC 微型實驗室協定 v0.1(凍結門檻)ACTIVEREL-2026-0157
EXP-2026-0009 PACC-Lab v0.2——第三個家族(N3)與對抗性證據幾何uses_benchmarkBEN-2026-0001 PACC 微型實驗室協定 v0.1(凍結門檻)ACTIVEREL-2026-0170
EXP-2026-0010 PACC-Lab v0.3——無 oracle 的來源可靠度學習uses_benchmarkBEN-2026-0001 PACC 微型實驗室協定 v0.1(凍結門檻)ACTIVEREL-2026-0183
EXP-2026-0011 PACC-Lab v0.4——相關來源與依賴幾何uses_benchmarkBEN-2026-0001 PACC 微型實驗室協定 v0.1(凍結門檻)ACTIVEREL-2026-0194
EXP-2026-0012 PACC-Lab v0.5——更豐富的關係狀態能救回潛在座標嗎?uses_benchmarkBEN-2026-0001 PACC 微型實驗室協定 v0.1(凍結門檻)ACTIVEREL-2026-0207
EXP-2026-0013 PACC-Lab v0.6——階層式組合座標uses_benchmarkBEN-2026-0001 PACC 微型實驗室協定 v0.1(凍結門檻)ACTIVEREL-2026-0218
EXP-2026-0014 PACC-Lab v0.7——不重新擬合的跨幾何座標轉移uses_benchmarkBEN-2026-0001 PACC 微型實驗室協定 v0.1(凍結門檻)ACTIVEREL-2026-0231
EXP-2026-0015 PACC-Lab v0.8——相關度掃描下的凍結座標轉移吸引域uses_benchmarkBEN-2026-0001 PACC 微型實驗室協定 v0.1(凍結門檻)ACTIVEREL-2026-0242
EXP-2026-0016 PACC-Lab v0.9——三錨點概率座標圖冊uses_benchmarkBEN-2026-0001 PACC 微型實驗室協定 v0.1(凍結門檻)ACTIVEREL-2026-0255
EXP-2026-0017 PACC-Lab v0.10——凍結三重重疊上的定向 cocycle 一致性uses_benchmarkBEN-2026-0001 PACC 微型實驗室協定 v0.1(凍結門檻)ACTIVEREL-2026-0266
EXP-2026-0018 PACC-Lab v0.11——雙向 cocycle 與反向一致性uses_benchmarkBEN-2026-0001 PACC 微型實驗室協定 v0.1(凍結門檻)ACTIVEREL-2026-0279
EXP-2026-0019 PACC-Lab v0.12——有界二次轉換uses_benchmarkBEN-2026-0001 PACC 微型實驗室協定 v0.1(凍結門檻)ACTIVEREL-2026-0292
EXP-2026-0020 PACC-Lab v0.13——反向殘差場/基點依賴uses_benchmarkBEN-2026-0001 PACC 微型實驗室協定 v0.1(凍結門檻)ACTIVEREL-2026-0303

歷史與來源歷程

Canonical URL
https://evemisslab.com/ai/benchmarks/BEN-2026-0001/
機器可讀
/ai/benchmarks/BEN-2026-0001/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