實驗EXP-2026-0020v0.1

PACC-Lab v0.13——反向殘差場／基點依賴

量測 held-out 反向殘差 r_CA(S) = Φ_A(S) − T_CA(Φ_C(S)) 的均值、共變異、潛在 logit 能量集中度、層間 vs 層內變異、方向穩定性，以及只用訓練集的逐層常數校正（對照全域與 shuffled 層控制）。N0/N1 把 98 % 的殘差能量集中在潛在 logit 軸上，但均值方向跨 seed 翻號（跨 seed 餘弦 ≈ −1），層間結構只有 ≈ 0.1–0.3 %（門檻 20 %），逐層校正對 held-out 保真度的改變 < 10⁻⁵——沒有家族通過。

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: NEGATIVE
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: Model Representation, Evaluation, Formal AI
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: The reverse bias is a base-point-conditioned residual field r(S) ≈ b_q + ε learnable per frozen overlap stratum.

設定

model_ids: MOD-2026-0001 N0——帶號支持
MOD-2026-0002 N1——序數錦標賽
MOD-2026-0003 N2——帶號圖
MOD-2026-0004 N3——約束競爭
MOD-2026-0005 Bayesian 參考（精確後驗／Beta-Bernoulli／聯合共同因／HMM）
dataset_ids: DAT-2026-0001 PACC 合成證據世界
benchmark_ids: BEN-2026-0001 PACC 微型實驗室協定 v0.1（凍結門檻）
software_environment: Python; deterministic seeded generators; no network, no LLM.

執行

run_count

random_seeds

20260909
4 secondary seeds
3 primary-scale-ish seeds

controls

shuffled-target mapping
constant prediction
broken composition (v0.6+)
target-local refit (v0.7+)

metrics

verdict

RESIDUAL MOSTLY UNSTRUCTURED

latent_energy_fraction_affine

N0: 0.9845
N2: 0.7723
N3: 0.3182

between_q_fraction

0.0012–0.0033 vs gate 0.20

stratum_correction

N0_uncorrected: 0.011616
N0_corrected: 0.011624

cross_family_direction_cosine_median

0.7024

cross_seed_direction_cosine_N0

-0.9977

base_point_pass

0 for every family

詮釋

interpretation: A stable residual axis is not a stable residual orientation and not a base-point field; the reverse bias contains family-dependent dominant error modes, which weakens the gauge/connection-like reading. Next (v0.14, preregistered): sign-free residual subspace / principal-axis stability, with no q input, no higher degree, no neural mapper.

限制

limitations: Residual unstructuredness is not established in every sign-free or subspace sense — that is the v0.14 question.

重現

reproduction_instructions: Extract the version's FINAL bundle; python -m pytest -q; run the version's primary script with the recorded seed; docs/PACC_LAB_v0.N_RESULTS.md and docs/EXPERIMENT_PROTOCOL_v0.N.md are inside the bundle.

記錄欄位

completed_at: 2026-09-09

關係

來源	關係	目標	狀態	ID
`EXP-2026-0020` PACC-Lab v0.13——反向殘差場／基點依賴	`runs_on`	`SYS-2026-0002` PACC-Lab——收斂猜想的微型實驗室	`ACTIVE`	`REL-2026-0302`
`EXP-2026-0020` PACC-Lab v0.13——反向殘差場／基點依賴	`uses_benchmark`	`BEN-2026-0001` PACC 微型實驗室協定 v0.1（凍結門檻）	`ACTIVE`	`REL-2026-0303`
`EXP-2026-0020` PACC-Lab v0.13——反向殘差場／基點依賴	`uses_dataset`	`DAT-2026-0001` PACC 合成證據世界	`ACTIVE`	`REL-2026-0304`
`EXP-2026-0020` PACC-Lab v0.13——反向殘差場／基點依賴	`uses_model`	`MOD-2026-0001` N0——帶號支持	`ACTIVE`	`REL-2026-0305`
`EXP-2026-0020` PACC-Lab v0.13——反向殘差場／基點依賴	`uses_model`	`MOD-2026-0002` N1——序數錦標賽	`ACTIVE`	`REL-2026-0306`
`EXP-2026-0020` PACC-Lab v0.13——反向殘差場／基點依賴	`uses_model`	`MOD-2026-0003` N2——帶號圖	`ACTIVE`	`REL-2026-0307`
`EXP-2026-0020` PACC-Lab v0.13——反向殘差場／基點依賴	`uses_model`	`MOD-2026-0004` N3——約束競爭	`ACTIVE`	`REL-2026-0308`
`EXP-2026-0020` PACC-Lab v0.13——反向殘差場／基點依賴	`uses_model`	`MOD-2026-0005` Bayesian 參考（精確後驗／Beta-Bernoulli／聯合共同因／HMM）	`ACTIVE`	`REL-2026-0309`
`EXP-2026-0020` PACC-Lab v0.13——反向殘差場／基點依賴	`tests`	`THY-2026-0005` PACC 猜想——四層收斂階梯	`ACTIVE`	`REL-2026-0310`
`EXP-2026-0020` PACC-Lab v0.13——反向殘差場／基點依賴	`extends`	`EXP-2026-0019` PACC-Lab v0.12——有界二次轉換	`ACTIVE`	`REL-2026-0311`
`EXP-2026-0020` PACC-Lab v0.13——反向殘差場／基點依賴	`produced`	`ART-2026-0033` PACC-Lab v0.13 Residual Field FINAL `artifact://evemisslab/adaptive-epistemic-systems/PACC-Lab_v0.13_Residual_Field_FINAL.zip`	`ACTIVE`	`REL-2026-0312`

歷史與來源歷程

Canonical URL

https://evemisslab.com/ai/experiments/EXP-2026-0020/

機器可讀

/ai/experiments/EXP-2026-0020/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