計畫PRG-2026-0101v0.1
智能的物理計量（IPM）
這個研究計畫問的不是「AI 有幾分聰明」，而是一個答案讓物理系統付出了什麼：花了幾次使用者回合、幾條生成軌跡、幾層隱藏 LOOP，做了多少有效語意工作，占用多少能量與計算時空，依賴多少外部鷹架，最後換回多少可驗證品質。十篇理論論文（2026-09-02）定義了測量物件——任務、品質、語意工作、物理計算、鷹架能力、測量詮釋資料——與五個可證偽命題；v0.2 實驗協定（Experiment A：single-pass vs scaffolded）落實為 XA-02…XA-06L 儀器套件，並在本地 9B 模型上執行過一次。
目標
goals- Replace token, FLOP, benchmark score and 'one turn' as units of intelligence with a typed event 𝔍_IPM = (task, quality, semantic work, physical computation, scaffolding, metadata).
- Measure the relation physical computation → effective semantic work → verifiable quality, and compare systems on a Pareto frontier instead of a single score.
- Put the five falsifiable propositions — token hypothesis, FLOPs sufficiency, binary burden, scaffolding separation, semantic intermediate utility — in front of experiments in the order A → D → B → C → E.
- Publish under the IPM Minimum Reporting Standard: boundary, energy type, hidden work, uncertainty and measurement grade on every number.
開放問題
open_questions- μI has no operational identification yet (Experiment C); the series itself says μI must pass a predictive/explanatory utility test or be revised or eliminated.
- The pilot's quality axis measured output-format compliance for two of three tasks; task semantic quality, output-contract compliance, system completion reliability and verifier reliability have to be separated before scaling.
- Same-model verifier serialization failed in 3 of 18 verifier trials; candidates abandoned on abort are not yet accounted; telemetry sampling costs about 24 % of trial wall time.
- CODE tasks have no measured quality until the pilot runs inside a disposable sandbox with code evaluation enabled.
- Experiments B, C, D and E are declared, not run.
里程碑
milestones- 2026-09-02: EML-IPM v0.1 theoretical series complete, 10/10 papers with a SHA-256 manifest; canonical index with unified notation and the v0.2 experimental entry point.
- 2026-09-02: Experiment A protocol (EML-IPM-XA-01 v0.1, READY FOR PILOT).
- 2026-09-03: instrument packages XA-02 (30-task pack), XA-03 (telemetry logger), XA-04 (A0→A5 runner), XA-05 (36-trial synthetic smoke gate), XA-06 (real-model pilot gate), XA-06L (local execution handoff).
- 2026-09-03: first real-model pilot on a local Qwythos-9B-v2 — 36 trials, sealed REAL_MODEL_PILOT_INCOMPLETE (33/36 complete).
- 2026-09-07: diagnostic analysis of the pilot with seven instrument revisions required before XA-07.
此計畫下的研究線與系統
| 來源 | 關係 | 目標 | 狀態 | ID |
|---|---|---|---|---|
RES-2026-0101 執行與物理線——一個答案在回合、語意工作、能量與計算時空上的代價 | belongs_to | PRG-2026-0101 智能的物理計量（IPM） | ACTIVE | REL-2026-0353 |
RES-2026-0102 品質線——不叫人打分數也能量成果品質 | belongs_to | PRG-2026-0101 智能的物理計量（IPM） | ACTIVE | REL-2026-0354 |
RES-2026-0103 能力線——拿掉 LOOP 還剩多少智能，以及統一的智能事件 | belongs_to | PRG-2026-0101 智能的物理計量（IPM） | ACTIVE | REL-2026-0355 |
關係
| 來源 | 關係 | 目標 | 狀態 | ID |
|---|---|---|---|---|
PRG-2026-0101 智能的物理計量（IPM） | released_as | ART-2026-0101 EML-IPM v0.1 canonical series package — 10 papers + canonical index + SHA-256 manifest (UTF-8 Markdown) artifact://evemisslab/intelligence-physical-metrology/IPM_v0.1_Canonical_Series_Package.zip | ACTIVE | REL-2026-0390 |
RES-2026-0101 執行與物理線——一個答案在回合、語意工作、能量與計算時空上的代價 | belongs_to | PRG-2026-0101 智能的物理計量（IPM） | ACTIVE | REL-2026-0353 |
RES-2026-0102 品質線——不叫人打分數也能量成果品質 | belongs_to | PRG-2026-0101 智能的物理計量（IPM） | ACTIVE | REL-2026-0354 |
RES-2026-0103 能力線——拿掉 LOOP 還剩多少智能，以及統一的智能事件 | belongs_to | PRG-2026-0101 智能的物理計量（IPM） | ACTIVE | REL-2026-0355 |
歷史與來源歷程
- Canonical URL
- https://evemisslab.com/ai/programs/PRG-2026-0101/
- 快照
AI-SNAPSHOT-v0.1-fe85b9694a45- 來源歷程
source- EveMissLab research collection: Intelligence Physical Metrology (真本體論13)
extracted_by- Splice (Claude Code), reading the canonical UTF-8 sources and each package's own reports
extracted_at- 2026-09-11
generator- tools/extract_all.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